May 2026

Qwen image generation and editing now finish reliably — your work no longer gets stuck spinning if we ship an update while an image is being created. Every job lands as a finished result, or refunds automatically if something goes wrong.

Image tools are now far easier to use on mobile — pick your model, aspect ratio, and number of images with tap-friendly controls, see the credit cost right on the Generate button, and on phones the controls slide up in a tidy composer drawer so the screen stays focused on your work.

This rollout reached a dozen-plus image generators, upscalers, and editors, so they all look and behave the same way no matter which one you open — no more relearning the controls from tool to tool.

P Video Animate — motion transfer for still images. Upload a reference video and a character image, and it copies the video's motion, timing, and camera moves onto your character. Under Generate Videos. 1 credit/sec at 720p, 2 credits/sec at 1080p.

Flux Pro Virtual Try-On — premium, high-fidelity try-on under Characters & Faces. Dress a person photo in any garment, with an optional styling note ("tucked in", "open over a tee"). Preserves pose, lighting, and background. 2 credits per result.

Krea V2 (Medium + Large) — text-to-image with a creativity dial (raw → high) and style reference images. Medium at 1 credit for fast iteration, Large at 2 credits for the highest fidelity, across eight aspect ratios.

Riverflow 2.0 now generates cleanly on every aspect ratio at 1K, 2K, and 4K — the ratios that used to error out are fixed.

Mobile tool pages: the Generate button stays visible and tappable, new generations scroll into view automatically, and recent generations moved into the My Stuff menu to keep the screen clear for creating.

Fresh prompting guides with real example galleries for Krea V2, Flux Pro Virtual Try-On, Riverflow, Luma UNI 1, and P Video Animate — all under Guides.

Riverflow 2.0 — Fast + Pro in one tool. Fast at 1 credit (1K/2K) for cheap iteration. Pro at 3 credits (1K/2K) or 7 credits for full 4K — the only model on the site shipping 4K output right now. Text-to-image with optional reference image to steer style and composition (the model writes a fresh frame from your prompt; references aren't modified). Find it under Image Generation.

Luma UNI 1 — Standard + MAX in one tool. Standard at 1 credit for fast iteration, MAX at 3 credits for the quality-first variant. ~3K resolution across 7 aspect ratios, up to 8 reference images per edit, optional manga style preset. Also under Image Generation.

P-Image Upscale — sub-second upscaling up to 128 megapixels. From 0.1 credit (4 MP) up to 3 credits (128 MP). Toggles for detail boost and an AI-image realism pass that cleans up plastic-looking outputs. Find it under Upscale & Enhance.

Searching the Prompt Library and then clicking the Image or Video filter chip was kicking users to an error page. The filter now applies cleanly on top of any search — pick a query, narrow by type, and the matching prompts show up the way they should. Same fix covers tag and model combinations along the way.

Sourceful Riverflow 2.0 is now live in the Image Generation section. One tool, two quality tiers: Fast is 1 credit per image (1K or 2K). Pro is 3 credits per image at 1K/2K and 7 credits at 4K — the only tier on Pixel Dojo right now that ships full 4K output.

Click any image in your history to use as a reference — Riverflow uses it to steer style and composition while generating a fresh frame from your prompt. (It's reference-guided generation, not inpainting/editing — the source image isn't modified.) Ten aspect ratios available: 1:1, 4:3, 3:4, 16:9, 9:16, 3:2, 2:3, 5:4, 4:5, and 7:3 (ultra-wide).

The resolution picker is smart about the model you're on — 4K is unavailable while Fast is selected (Fast doesn't ship 4K), and switching from Pro 4K back to Fast snaps you to 2K automatically instead of leaving an invalid combo. Cost preview updates live as you change model or tier so there's no math required.

Luma UNI 1 is now live in the Image Generation section. One tool, two quality tiers: Standard is fast and clean at 1 credit per image, MAX is the quality-first variant at 3 credits per image. Pick the tier in the form, the cost updates live as you change it.

Both text-to-image and image editing in the same page. Click any image in your history to add it as a reference, then describe the edit — Luma supports up to 8 reference images per generation. Ships at ~3K resolution across 1:1, 16:9, 9:16, 3:2, 2:3, 2:1, and 1:2. Optional 'manga' style preset for the look you'd expect.

The in-app 'What's New' modal was clipping the last item or two behind the Got it footer on shorter screens — laptops, smaller windows, and phones in particular. The modal now keeps comfortable breathing room above and below regardless of viewport size, the footer always stays in view, and mobile browser chrome (Safari's address bar in particular) no longer pushes anything off-screen.

When there's more content below the visible area, a bouncing chevron appears at the bottom of the modal alongside a soft fade — no more guessing whether you've seen everything. The cue disappears once you've scrolled to the last item.

The modal was rendering two close buttons stacked in the corner. Now there's one.

P-Image Upscale is now live in the Upscale & Enhance section. Sub-second image upscaling to any target size up to 128 megapixels. Pricing scales with output size: 0.1 credit for 4 MP, 0.5 credit for 16 MP (≈4K), 2 credits for 64 MP, 3 credits for 128 MP. Toggles for Enhance Details (sharper textures) and Enhance Realism (cleans up AI-generated images that look too plastic).

The Upscale & Enhance lineup now has three clear lanes: P-Image for fast and cheap bulk work, Clarity Pro for photorealistic with identity preservation, and Magnific for creative/premium. Pick the right one for the job instead of paying premium prices on every upscale.

When we ship a new version of Pixel Dojo while you have it open, the old page used to silently reload itself a second or two after you next clicked something — which sometimes ate the video or image you just started generating. It now shows a quiet 'A new version of Pixel Dojo is available' banner instead. You choose when to refresh, so nothing in flight gets dropped.

The app now checks for new versions in the background and surfaces the upgrade banner the moment a deploy lands, instead of waiting until you click something that tries to load a missing chunk. Less surprise, no broken interactions.

WAN 2.7 Video now reserves a slot in your history the instant you hit Generate, before the network call lands. If anything interrupts the dispatch — a reload, a network blip — the job is no longer 'lost'; it shows up as 'Preparing…' and either resolves or clears itself cleanly. Other generation tools get the same protection as we roll them over.

Qwen Image, Qwen Image 2, and Kling Image no longer deselect your source images the moment you hit Generate. The images you picked stay selected so you can keep iterating on the same shot — tweak the prompt and hit Generate again without re-picking everything. Use the Exit edit mode button (or Clear history) when you actually want to start fresh. Matches how the other edit-capable models already behave.

The looping hero video at the top of the home page no longer hangs on a loading spinner — it loads and plays as soon as the page is ready. Underneath, we also hardened the player so any future bad video URL falls through to a Play button instead of silently spinning.

The Agent now reads what your prompt is actually about — product shot, portrait, logo, cinematic, anime, character — and picks a model that's good at that flavor. Before, every text-to-image prompt routed to whatever was most popular overall, which meant a logo request and a product photo both landed on the same generic default. Now a logo goes to a text-rendering-strong model, a product shot goes to a product-photography model, and so on.

Model recommendations now use site-wide usage data (every surface — chat, Canvas, dashboard, API, MCP) instead of Canvas-only data. The top models reflect what creators across the whole site converge to, not just one surface. A single heavy API caller can't dominate the rank either — each user's contribution is capped at 5% of total signal.

When the Agent picks a model for you, you'll see WHY in the response — 'used seedream-4 because it's tuned for product photography; also good: nano-banana-pro, gpt-image-2'. No more black-box defaults.

Every Agent result now shows a 'Try next' chip strip — Animate, Add effect, Photoshoot, Upscale, Edit again — that one-click chains the next skill onto your current image. Pre-fills the composer, attaches the asset for you, and the inline mode/effect picker pops automatically. No more remembering what / commands exist or what step comes next.

Brand-new Agent chats now show a categorized starter grid (Image / Edit / Video / Marketing) with 8 ready-to-send prompts spanning every skill — including the newer /photoshoot, /effect, and /hooks paths that were easy to miss. The /hooks card carries a 'Free' badge so you can experiment with paid-social hooks without thinking about credits.

Campaign tool card was showing a fake 'Prediction not found' error on campaigns that were actually still rendering. Wrong job-tracker path was sending the campaign ID to the image-prediction lookup, which 404'd and surfaced as a red error. Campaigns now poll their own status endpoint, show partial progress as assets land ('Campaign running — 3 of 5 assets ready'), and only flip to failed if every sub-job genuinely fails.

If you just generated a video, then asked the Agent to 'edit this' or 'animate it', the .mp4 URL was getting passed as a still-image reference — which silently 422'd at the model with 'cannot read image'. The Agent now walks back to the most recent IMAGE asset in your history, skipping videos, since every edit / storyboard / upscale / character / photoshoot / effect skill consumes a still.

WAN 2.2 (Standard, Plus, Extended) was rejecting public-API calls that didn't explicitly include a 'mode' field with 'mode: Required'. Now infers the mode from your input: pass an image_url and it picks image-to-video, otherwise text-to-video. Same convention every other multi-mode video model uses.

Public /pricing page now shows three tiers side-by-side. Lite ($10/mo, 160 credits) for casual creators, Pro ($25/mo, 400 credits) for working creators, Pro Max ($50/mo, 800 credits) for power creators. Every tool, every model, every new launch — included on every plan. Pick the volume that fits how much you create.

Pro Max is here for the creators who keep buying $50 credit packs one at a time. Same 800 credits, but as a subscription — predictable monthly billing, no more 'I should top up before tomorrow's deadline'.

Every plan unlocks the whole studio. Lite gets the same 70+ models, LoRA training, Canvas, and developer API as Pro and Pro Max — you're choosing how much you create per month, not which tools you get. Pro keeps its 6-month and annual billing options behind a small disclosure below the card so they don't compete with the tier comparison.

Mobile /pricing layout stacks the three cards with Pro on top to preserve the anchor — same hierarchy you see on desktop, just vertical. No squishing the 'Most Popular' card off-screen.

SEO + page meta on /pricing now lead with 'Plans from $10/mo'. Schema.org markup includes all three tier offers so search engines can show the right starting price.

Pixel Dojo Lite is here — $10/month for 160 credits. Same per-credit price as Pro ($0.0625/credit), just a smaller plan. Currently surfaced as a save option when you go to cancel and pick 'It's too expensive' — coming to the public pricing page soon.

Cancel flow now leads with real save options when you pick 'It's too expensive'. Instead of just a 'Manage plan' button that punted you to a portal, you now see a clear sapphire-glow downgrade card showing $25 → $10 with side-by-side credit counts (160 vs 400), then Pause as the secondary option, then a smaller 'Manage plan' link. Hierarchy you can scan in two seconds.

Downgrade now has a proper email flow. Click 'Switch to Lite' and you get an immediate confirmation showing exactly when your current plan ends — plus a reminder email at that date with a one-click link to start Lite. Your trained models, LoRAs, and media library stay intact through the transition.

Modal stays open with a confirmation state after you schedule a downgrade — 'Your Pro stays active until [date]' — instead of closing instantly with a fading toast. Easier to absorb what just happened and what comes next.

Pause and Downgrade are now mutually exclusive so you can't accidentally do both. If you're already paused, the downgrade card shows 'Resume to downgrade'; if a downgrade is scheduled, the pause card shows 'Cancel pending downgrade first'. The blocked option stays visible so you know it exists.

Click the credit number in the top bar to open a quick view of your balance, this month's credits used, and your recent refunds. No more digging through menus to see where credits went or whether a failed generation actually got refunded.

Separate Buy Credits button now sits next to your balance, so topping up is always one click away — even when you're deep inside a tool or Canvas session.

Small green dot lights up on the credit pill whenever refunds have landed that you haven't seen yet. When a generation fails or times out we put the credits back — now you actually know it happened.

Dashboard now shows a monthly usage card right after Recent Creations: credits used, credits refunded, your top tools by spend, and a one-click link to the full report.

Refund notifications now fire as soon as a generation fails — see the credits come right back in real time. Multiple refunds landing at once bundle into a single notification so you don't get spammed.

MCP skills handle missing inputs more gracefully. The character skill now returns a clear message when called without a reference image on edit-style models (instead of looping with a cryptic backend error), and the generate skill correctly maps parameters for grok-video-extend.

Magnific / Creative Upscale at very high scale factors now returns a clear, actionable error when the output exceeds size limits — telling you exactly what to try (lower scale factor or smaller source image) instead of a raw parse error.

API access now works with credits alone — no active subscription required. If you ever saw 'An active subscription is required to use the API' from /api/v1/generate even though you had plenty of credits (or paid via a credit pack instead of a recurring plan), that error is gone. Behavior now matches what /api-platform/documentation has always said: signed-in users can create an API key and use it as long as they have credits.

Twenty-one tools picked up an in-tool prompting-guide banner today — open one and you'll see a small 'Read the guide' card at the top of the settings sidebar (or in the empty state) linking straight to that model's dedicated prompting guide. Video side: Runway Gen-4.5, LTX 2.3, Pixverse V6, Pixverse v5.6, Seedance 1, WAN 2.5, P-Video. Image side: Hunyuan Image 3, Dreamina 3.1, Ernie, PonyXL, WAN Image, ImagineArt, Qwen Image, P-Image, WAN 2.7 Image, WAN 2.6 Image, FLUX Studio, GPT-Image 1.5, Z Image Turbo, FLUX 2 Flex. The guides have been live at /guides/* for a while — they're just one click away from the tool page now.

FLUX Studio prompting guide at /guides/flux-studio-prompting-guide — documents the every-FLUX-1-variant workspace (Schnell, Dev, Krea Dev, FLUX Pro, 1.1 Pro / Pro Ultra, Kontext Pro / Max). Four worked examples plus a tier-selection guide for matching variants to use cases.

WAN Image prompting guide at /guides/wan-image-prompting-guide — the fast Replicate-backed WAN image variant. Documents the optional LoRA fal-lora variant and how it pairs with WAN 2.7 Image for hero shots. Four examples covering portrait, landscape, product, and stylized.

WAN 2.6 Image prompting guide at /guides/wan-2-6-image-prompting-guide — intermediate-gen WAN with the distinctive fixed pixel-dimension preset system (1280×1280, 1280×720, 768×1024, etc) vs free aspect ratios. Useful when you need exact downstream output dimensions.

P-Video prompting guide at /guides/p-video-prompting-guide — Pruna's three-mode video model (text-to-video, image-to-video, audio-to-video). Documents the cost-efficient tier choice and the audio-to-video mode for music-driven motion and lip-sync work.

Ernie Image prompting guide at /guides/ernie-prompting-guide — Baidu's image flagship with multilingual prompt support (English, Chinese, Japanese) and native CJK text rendering. Four worked examples spanning a Chinese tea master portrait, Shanghai Pudong skyline, Chinese calligraphy sign, and traditional garden pavilion.

FLUX 2 Flex prompting guide at /guides/flux-2-flex-prompting-guide — the resolution-flexible FLUX 2 variant where you pick 0.5/1/2/4 MP per generation. Documents how to iterate cheap at 0.5 MP and finalize at higher MP, plus how it relates to FLUX 2 Pro / Max / Dev.

Pixverse v5.6 prompting guide at /guides/pixverse-prompting-guide — the dependable previous-generation Pixverse workhorse. Quality tiers from 360p to 1080p, optional audio, predictable motion. Four examples covering glassy lake sailboat, beach character walk, fountain pen signature, and skater handrail grind.

WAN 2.5 prompting guide at /guides/wan-2-5-prompting-guide — fast image-to-video flash variant in the WAN family. Documents how the speed-optimized I2V tier compares to the premium WAN 2.7. Four worked examples spanning rice terrace drone, maple forest character walk, ink-drop macro, and sunset surfer vertical.

PonyXL prompting guide at /guides/ponyxl-prompting-guide — documents the score-tag quality system (score_9, score_8_up, score_7_up that you must lead every prompt with) and booru-style tag prompting conventions. Four worked examples across anime portrait, fantasy knight, chibi cat wizard, and anime mecha.

Z Image Turbo prompting guide at /guides/z-image-turbo-prompting-guide — covers what's distinctive about Z Image Turbo's synchronous response (URL returns inline, no polling round-trip needed), and why that matters for high-volume programmatic workflows. Four production examples.

WAN 2.7 Image prompting guide at /guides/wan-2-7-image-prompting-guide — the image-side sibling of WAN 2.7 Video. Documents Standard/Pro tier choice and how the same prompt language carries between WAN Image and WAN Video for consistent aesthetics across modalities.

P-Image prompting guide at /guides/p-image-prompting-guide — at 0.1 credits per image it's the cheapest image option on the platform. Documents the moodboard / high-volume use case where you generate 30-50 variations for ~3-5 credits total, then upscale the keepers via a premium tier.

Hunyuan Image 3 prompting guide at /guides/hunyuan-image-3-prompting-guide — Tencent's 80-billion-parameter image model. Four worked examples spanning painterly fantasy, photoreal cultural portrait, Brutalist architecture, and mechanical watch macro. Covers the Fast/Quality toggle.

Google Gemini Image prompting guide at /guides/google-gemini-image-prompting-guide — Google's image flagship with three tiers (Standard at 1 credit, Pro at 3-6) and unique native 4K support. Examples for cinematic 21:9 landscape, documentary portrait, food editorial at 4:5 Instagram-spec, and stylized illustration.

ImagineArt prompting guide at /guides/imagineart-prompting-guide — three model tiers (2.0 preview, 1.5 Pro, 1.5) in one workspace, nine aspect ratios including the unusual extremes 1:3 (ultra-tall mural) and 3:1 (ultra-wide banner). Four worked examples cover cinematic, fashion editorial, pop art, and ultra-tall mural composition.

Grok Imagine Video prompting guide at /guides/xai-video-prompting-guide — xAI's video model. Four examples (Mojave Mustang, Irish cliff character, chocolate truffle macro, snowboarder vertical) covering cinematic motion, atmospheric scene composition, and 480p/720p tier choice.

Qwen Image prompting guide at /guides/qwen-image-prompting-guide — covers the Plus / Max tier choice, snapshot pinning for campaign consistency, and clean text rendering in Latin and CJK scripts. Four worked examples including a multi-line bookshop sign with proper kerning.

Dreamina 3.1 prompting guide at /guides/dreamina-prompting-guide — 4MP cinematic image generation with atmospheric defaults, flat pricing across 1K and 2K, and custom dimensions (256–2048 each axis) for spec-size output. Four worked examples covering architecture, food editorial, fashion portrait, and atmospheric fantasy.

Pixverse V6 prompting guide at /guides/pixverse-v6-prompting-guide — documents the unique multi-clip mode (first frame + last frame generates the transition between them), the six built-in style presets (anime, 3D animation, clay, comic, cyberpunk, none), and the moderation quirks worth knowing.

Seedance 1 prompting guide at /guides/seedance-1-prompting-guide — smooth natural motion, six aspect ratios including the standout 21:9 ultra-wide cinematic, native audio with sound-cue steering, and the end-frame interpolation mode for motion between two anchor images.

GPT-Image 1.5 prompting guide at /guides/gpt-image-1-5-prompting-guide — covers literal prompt adherence, in-image text rendering with quoted strings, transparent-background PNGs (real alpha channel, no cutout pass needed), and when to reach for it over FLUX or Seedream. Five worked examples spanning chalkboard typography, editorial illustration, macro photoreal, transparent logo, and cinematic landscape.

Ideogram Character prompting guide at /guides/ideogram-character-prompting-guide — how to get consistent character variations from a single reference image, no LoRA training required. Includes a worked example of one character pushed across fantasy armor, anime, film noir, and sci-fi astronaut styles. Reference-image quality rules included.

Runway Gen-4.5 Video prompting guide at /guides/runway-gen4-video-prompting-guide — cinematic motion, camera-language fluency, six aspect ratios from 21:9 cinematic through 9:16 vertical. Four worked examples (aerial drone push, character tracking shot, product macro, vertical reel) plus the gotcha that 1:1 aspect requires image-to-video mode.

LTX 2.3 prompting guide at /guides/ltx-2-prompting-guide — the broadest-surface video model on the site. Covers the Fast vs Pro tradeoff, durations up to 20 seconds, resolutions through 4K, optional native audio, and three modes (text-to-video, image-to-video, video extension for 30+ second pieces). Four worked examples and a full settings table.

Eight fresh prompting guides shipped today — FLUX, Recraft V4.1, Seedream 4.5, HiDream O1, Hailuo 2.3, Kling Image v3, Kling 2.6 Pro, and WAN 2.7 Video. Each one is built from a shared format with real generated examples (not stock clips), a copy-ready prompt library, settings reference, and a model-specific FAQ. Browse them all at /guides.

Four of the eight are video model guides — Hailuo 2.3, Kling 2.6 Pro, and WAN 2.7 Video each ship with embedded sample clips you can scrub through before writing your own prompt. The shared template will speed up every video guide we publish from here.

Four image model guides for FLUX, Recraft V4.1, Seedream 4.5, and HiDream O1 — cover style ranges, aspect ratios, the model's prompt sensitivities, and what to ask for vs what to leave out. Useful whether you're brand-new to the model or you've hit a wall trying to nail a specific look.

Seedream 3 has been fully retired — the model was no longer reachable from the dashboard but a few public API and Canvas surfaces still advertised it, which meant API calls to /api/v1/models/seedream-3/run were returning service-unavailable errors. Cleaned up the catalog so what you see is what you can actually generate with. If you were using Seedream 3, move over to Seedream 4.5 or Seedream 5 Lite — both are live and stronger across the board.

Hunyuan Image 3 moved to a new backbone so we can ship it with the safety checker fully disabled at the provider — content blocks should no longer interrupt your generations. Pricing changes accordingly: 2 credits per image (was 1).

Hunyuan Image 3 settings simplified — replaced the Steps + CFG sliders with a single Fast / Quality mode toggle (Fast is the default, flip it off when you want max fidelity). Aspect ratio presets expanded from 3 to 7 (added 16:9, 9:16, 4:3, 3:4, 3:2, 2:3).

Hunyuan Image 3 is now available — Tencent's 80-billion-parameter text-to-image model with high-fidelity output and fine-grained control. 1 credit per image. Generate up to 4 images at once.

Hunyuan Image 3 exposes Steps (10–50) and Guidance / CFG (1.5–15) controls — adjust denoising iterations and how literally the model follows your prompt. Default values (20 / 7) work for most cases; nudge up CFG when you want the output closer to the prompt.

Hunyuan Image 3 ships with three aspect ratio presets — 1:1 (square), 3:1 (ultra-wide cinematic), 1:3 (ultra-tall portrait). More aspect ratios to follow once we lock down the supported pixel mappings.

Bria 3.2 prompting guide is live at /guides/bria-3-2-prompting-guide — five example prompts spanning Photography and Art mediums, plus six prompting tips for matching medium to language, anchoring quality, and getting clean product shots.

Bria 3.2's 14:17 and 17:14 aspect presets were sending unsupported pixel dimensions and erroring on submit. The pixel mappings have been corrected — those aspects now generate cleanly.

Bria 3.2 is now available — text-to-image at 1K resolution across 9 aspect ratio presets (1:1, 3:2, 2:3, 7:4, 4:7, 9:7, 7:9, 17:14, 14:17). 1 credit per image. Generate up to 4 images per prompt and pick the output format (PNG, JPG, or WEBP) before you hit go.

Bria 3.2 medium toggle — bias outputs toward Photography or Art with a single click. Useful when prompts read ambiguously and you want to nudge the result one direction or the other without rewriting the prompt.

Bria 3.2 image and prompt enhancement — optional toggles for richer detail / sharper textures and for letting the model expand your prompt with descriptive variations.

Vidu Q3 is now available — text-to-video and image-to-video across four resolution tiers (360p, 540p, 720p, 1080p) and five aspect ratios (16:9, 4:3, 1:1, 3:4, 9:16). Pick exactly the frame and resolution you need before generating instead of upscaling or reformatting later. Credit cost scales with resolution: 1 credit/second at 360p and 540p, 2 at 720p, 3 at 1080p.

Optional synchronized audio on Vidu Q3 — off by default for clean tracks you'll post-score, one toggle on when you want ambient and diegetic sound layered in with the visuals. End your prompt with a 'Sounds of …' line to steer the soundscape.

Vidu Q3 image-to-video — upload one source still and Vidu writes motion onto it from your prompt. Source aspect drives the output frame automatically, so portrait photos animate as portrait videos without reformatting.

Vidu Q3 prompting guide at /guides/vidu-q3-prompting-guide — six example prompts spanning cinematic aerials, character close-ups, product spots, multi-shot action, atmospheric wides, and vertical I2V. Plus a settings reference table and a six-item FAQ.

Image-to-video on WAN 2.7 was silently dropping the source frame in some flows — you'd attach a photo, ask it to animate, and the result would be a completely different subject. The variant didn't pin its mode internally, so a default kicked in that took the text-only branch and ignored the image. Now pinned correctly: your starting frame is honored every time, whether you call it from chat, Canvas, the dashboard, or the public API.

Audited every multi-mode video model for the same bug pattern. Same fix applied to Happy Horse 1.0 (both text-to-video and image-to-video variants) and Kling Video v3 (all four — Standard Text, Standard Image, Pro Text, Pro Image). Worst case caught: Kling Pro Image was doubly broken — would have been billed as Standard pricing AND ignored the image. All fixed.

If you've ever pasted a long, detailed prompt and gotten hit with "Prompt must be 2,500 characters or less" or similar — that's gone. Every generation tool now lets your full prompt through. If a specific model genuinely can't handle the length you'll see that model's own error message (which is more accurate than a guessed cap), but the upfront client-side rejection is no longer in the way.

Same prompt-honor fix from earlier today extended to Dreamina — its 'Enhance Prompt' toggle was on by default and silently rewriting your wording before generation. Now off by default like the rest.

Hover-info icons added to the prompt-rewrite toggles on Dreamina, Hailuo, and WAN 2.2 too. Every "enhance / optimize / expand" checkbox across the site now tells you upfront what it actually does.

Audited every generation tool for silent prompt truncation. No tool ever shortens what you typed before sending it to the model. If your prompt is too long for a particular model you'll see a visible error so you can trim — never a silent edit.

If you ever felt like WAN 2.7 (or one of the other WAN, Qwen, or Z Image tools) wasn't following your prompt — you were right. Every generation was silently running your wording through an upstream auto-rewriter that expanded and stylized it before the model saw it. That's now off by default across all of them. What you type is what the model gets. You can still turn it back on if you want the AI to embellish short prompts.

Every "Enable prompt expansion" toggle now has a hover-info icon that explains exactly what it does (the model rewrites your wording) so the trade-off is visible at decision time.

Cleaned up the default negative prompt on video tools. Removed "excessive motion" and "flickering" — those were silently fighting action-heavy prompts. Only always-harmful negatives (anatomy defects, watermarks, compression noise) remain. If you want stillness, you can still add it to your own negative prompt.

Images you generated with the SDXL LoRA tool weren't recording which LoRA produced them — so when you opened the details popover in My Media on one of those images, the 'LoRAs Used' section stayed blank even when you'd clearly picked one. The save flow was dropping the LoRA metadata on the floor. Every new save from the LoRA Image Creator now correctly stores the LoRA name and strength, and they show up in the details panel where they belong. Older saves can't be retroactively tagged, but everything going forward will be.

A new 'My Stuff' button sits next to 'All Tools' in the top bar and opens a dropdown with everything personal — Dashboard, Canvas, Agent, My Media, My LoRAs, My Prompts, Community Gallery, and your Usage Report. One click from any tool page. These destinations used to live two clicks deep inside the All Tools drawer (or buried in the avatar menu).

Your full pinned tools list and the last five tools you visited also show up inside My Stuff — so you can jump between everywhere you actually work without leaving the page you're on. The three-icon pinned strip in the top bar stays exactly where it was.

Yesterday's 'Save All' truthful-toast fix only covered the image tools. Now every video tool — Kling Video, Veo 3.1, Hailuo, Seedance, WAN, Grok Video, LTX, Pixverse, Runway, P Video, Omnihuman, and the rest — uses the same single-round-trip save with per-item results. Select 10 videos, hit Save, you get an accurate count back. Mobile-safe with keepalive enabled.

Same fix applied to LoRA Image Creator. Bulk-save reliability now matches every other tool on the site.

If you ever clicked 'Save All' on a batch of generated images and saw the success toast but didn't find them in My Media later — that was a real bug, not a glitch. The save flow has been rebuilt around a single server round-trip with truthful per-image results: the toast now says 'Saved 7 of 10 — 3 failed' if any of them didn't land, and the failed ones stay selected so you can retry without re-picking them.

Mobile (especially iPhone) was where this hit hardest. iOS Safari was killing in-flight save requests whenever you switched apps or locked your screen mid-batch. Saves now run keepalive-true and the server-side fan-out keeps everything going even if you background the tab — no more silent drops.

Recraft is now V4.1 — better photorealism, smoother gradients, and noticeably better text rendering vs V4. Same prompt understanding and design taste, just sharper across the board. Available at /recraft.

Three output variants under the same tool: Raster (standard image), Vector SVG (editable, layered), and a brand-new Utility variant tuned for product mockups — flat lighting, front-facing composition, clean scenes. Perfect for packaging, icons, and e-commerce hero shots.

Pricing simplified and cheaper across the board. All Standard generations are 1 credit (was 1–2). All Pro generations are 5 credits (was 6–8). Pick the variant you need without doing pricing math first.

Aspect ratio menu expanded from 8 to 14 options. Added 2:1, 1:2, 5:4, 14:10, 10:14, and 6:10 alongside the standard square/portrait/landscape set.

Kling Avatar used to charge the full 60-second rate up front regardless of how long your audio actually was — a 2-second clip cost 120 credits at Standard. Now the cost panel reads your audio file the moment it uploads and charges only for the real duration. That same 2-second clip is 4 credits now (or 6 on Pro).

Kling Avatar V2 is live. Upload a portrait and an audio file and get a lip-synced talking video — works on realistic humans, stylized characters, cartoons, and animals. Great for talking-head explainers, narrated product demos, and giving any character a voice without filming.

Two quality tiers: Standard for fast turnarounds (2 credits/second) and Pro for sharper facial detail and smoother motion (3 credits/second). Video length matches your audio automatically.

Optional prompt slot lets you guide actions, emotions, and camera movement — the audio still drives the mouth and timing, the prompt adds direction on top ('warm enthusiasm', 'subtle nods', 'wide shot').

Public API calls to MuleRouter-backed video tools now surface the same helpful error messages the in-app tools already do. When the model is briefly offline upstream, you'll see 'This video model is briefly offline while we top up our backbone. Your credits were refunded — please try again in a few minutes.' — not the previous internal architecture message. Applies broadly: any tool with a server-side override now bubbles its real error (size limits, format issues, transient outages) through /api/v1, not just MuleRouter models.

Type /photoshoot or /effect and a picker pops up with the named choices — hero / lifestyle / with-model / levitating / restyle for photoshoots, and dolly-zoom / 360-spin / fpv / crane-up / whip-pan / parallax / vertigo / bullet-time for effects. Each entry has a short description so you can see what it does before picking. Tab or Enter to select; the value splices into your prompt and you keep typing the scene.

Typing something that isn't a valid mode/effect after /photoshoot or /effect (like pasting a URL) used to fail with a cryptic 'invalid enum value' error after sending. The picker now shows 'No match' inline as you type so the mistake is obvious before you waste the send.

/photoshoot turns any product image into a brand-quality shot. Five modes — hero (clean studio), lifestyle (real-world setting), with-model (person interacting), levitating (dynamic floating with splashes), and restyle (same product, fresh aesthetic). Each mode bakes in the prompt scaffolding a product photographer would write — you just pick the mode and add direction.

/effect adds a named cinematic effect to a video shot — dolly-zoom (Vertigo), 360-spin, fpv (first-person drone), crane-up, whip-pan, parallax, vertigo, or bullet-time. Pass the effect name and what the camera is looking at; the camera-move prompt is handled for you.

/hooks generates 10 paid-social ad copy hooks for a product or concept — TikTok / Reels / Shorts ready. Pure text, no image generation, zero credit cost. Mixes curiosity, social proof, transformation, contrarian, and FOMO angles so you can A/B test scroll-stoppers.

/model picker now works inline anywhere in your message, not just at the start. Type /model mid-prompt to pop a filterable list of every available model — narrows as you type, picks splice into your text without wiping what you'd written.

The + button gained a 'Pick a model' shortcut and an 'Insert skill' submenu. Both insert at your cursor instead of replacing your prompt — so you can write half a sentence, click +, choose a skill or model, and keep typing.

Follow-up skills (/edit, /character, /storyboard, /upscale, /photoshoot, /effect) now automatically use your last generated image as the reference. 'animate the dragon' just animates the dragon — no re-attaching, no re-picking.

'Just make me an image' was sometimes routing to an image-edit model by mistake, which would error because no reference image was attached. The Agent now correctly defaults to a true text-to-image model when you don't supply one.

Slash commands in the Agent. Type / and pick a skill directly — /generate, /edit, /character, /storyboard, /upscale, /from_url, /campaign — to skip straight to the tool without going through the chat picker. Tab or Enter to confirm.

Pin a specific model with /model <apiId> <prompt> — e.g. /model wan-2.7-i2v animate her smiling. Great when you know exactly which model you want and don't want the Agent to decide for you.

Attach reference images two ways: 'Upload from computer' or 'From My Media' — pick any image you've already generated on PixelDojo without leaving the chat.

Attached images now render as a real thumbnail inside your chat bubble instead of a clunky '[attached image: URL]' line. Looks like a proper chat — and follow-ups like 'animate this' thread the image through automatically.

Pasting a product URL (/from_url) now renders a rich preview card — hero image, product name, description, and source — instead of a wall of plain text. Much easier to confirm at a glance that the Agent grabbed the right product before you tell it to make hero shots.

The Agent's default text-to-image pick was Dreamina even when WAN and Nano-Banana are actually more popular. Defaults now follow real Canvas usage rank instead of alphabetical order — so 'just make me an image' lands on what creators actually keep, not whichever model came first in the alphabet.

Reading a product page from a URL no longer renders the success card with a red error border. It's a successful action, so it now looks like one.

When you ask the Agent to 'animate this' on an image you uploaded or just generated, it now picks WAN 2.7 i2v by default — far more faithful to the source subject than the more creative video models, which sometimes reinterpret the image entirely.

Every Agent generation card now shows the source image it used, so you can confirm at a glance that 'edit it' / 'upscale that' picked up the right reference.

The Agent now offers the same skill set as the public /skills page — edit, product extraction from URL, and full campaign generation (hero + lifestyle + video pack from one product link) are all available in chat.

Reference images now thread through reliably when you say 'edit this' or 'animate it' as a follow-up — no more dropping the image silently. Credits also refresh in the top bar the moment they're deducted, and conversation history persists cleanly when you leave and come back.

Asking the Agent to 'make a video' with WAN 2.2 no longer errors with 'mode: required' — text-to-video vs image-to-video is auto-detected from your prompt.

Qwen Image, Qwen Image 2, and their Edit variants now accept your full prompt — we removed the 800-character soft cap and the 'we trimmed your prompt' warning. Long, detailed prompts go through unmodified. If you regularly write past 800 chars for fine-grained control, this is for you.

Meet the Agent. Describe what you want — 'cinematic samurai portrait', 'edit this to add neon rain', 'make a 6-second video of waves' — and it picks the right model from across the 70+ tools and runs it. Live at /agent from the sidebar, available on subscribed plans. Generations cost the same credits as the tools they call; the chat itself is free.

The Agent auto-tunes to real PixelDojo usage. Defaults land on the models creators actually keep for each task — text-to-image, edits, video, upscales, characters — so 'just make me an image' produces what's working today, not yesterday's pick.

Six built-in skills cover the common asks: generate, edit, character (consistent identity across shots), storyboard (video), upscale, and a status checker for long-running jobs.

Videos render in the background and appear in chat the moment they're ready — same auto-polling pattern the video tool pages use. No need to ask 'is it done yet?'.

Drag in a reference image to anchor edits, character generations, or upscales. Each completed asset shows the model that made it with Save / Download / Open full size on hover. Conversation history saves so you can resume any chat.

Right-sized the Prompt Library headline so the featured strip stays visible without scrolling on a standard desktop. Still bold and editorial — just no longer eating the whole viewport.

Removed a redundant 'AI-powered' badge from the search bar — the sparkle icon already says it, and the extra chrome was competing with the Create button.

When semantic search has no real matches for your query, the empty state now says so honestly and points you to popular prompts instead of leaving you stranded with a generic 'not found' message.

Semantic search on the Prompt Library now only returns prompts that actually relate to your query. Weak matches are filtered out before they reach you — search for 'tesla' and you'll see Tesla-themed prompts (or an honest 'no semantic matches' when the library doesn't have any), not a random mix of motivational and business prompts pretending to be relevant.

Newly created prompts are now searchable the moment you save them. Previously, only prompts that were part of a periodic backfill were findable by semantic search — anything saved between backfills was invisible until the next sweep. Embedding now runs at save time, so brand-new prompts show up in relevant search results immediately.

The Prompt Library landing page got a real makeover — bigger editorial typography, a massive widescreen hero showcase autoplaying the standout video prompt of the week, and a row of supporting picks underneath. Great prompts now feel like featured work, not search results.

A 'Create Prompt' shortcut now lives right in the library hero so you can jump straight into the builder from anywhere in the library.

Semantic search now labels its results as 'Top N matches' instead of a misleading total — the AI search surfaces the most relevant prompts, not every possible match, and the UI now reflects that clearly.

A brief render error that took the Prompt Library landing page offline for a window today — now resolved.

The Prompt Library now uses AI-powered semantic search. Type what you're looking for in plain English and it finds prompts by meaning — not just keyword matching. 'Cinematic portrait at dusk' finds dusk-mood prompts even if they don't use those exact words.

Every prompt detail page now shows up to six semantically related prompts below the main one, so you can discover variations and adjacent styles without starting a new search.

A featured strip at the top of the library showcases curated prompts with rendered video and image previews — hover to see them come to life before you click through.

Outpaint no longer fails with a bare "Unprocessable Entity" error when you expand from the edges. The submit path now drops zero-valued sliders before the request reaches fal.ai, so an "edges" expansion with one or two active sides goes through cleanly instead of being rejected as an empty payload.

When fal.ai does reject an outpaint request, you now see the actual validation reason — which field or value was wrong — instead of just "Unprocessable Entity." Same translation we already use for Smart Resize and other fal-backed tools, now wired into Outpaint.

Click Enhance now works with model-specific examples on every PixelDojo tool, not just the top 10. Niche image and video models now pull from their own track record instead of guessing from generic data — a real precision boost for less-mainstream tools.

The pool of prompts the enhancer learns from just got dramatically larger. Every thoughtful prompt now counts toward retrieval, not just the most-liked or admin-approved ones — those are still weighted higher, but no longer the only ones in the running. Many more candidates per request.

Deleting a single image or video from My Media now fully cleans up the underlying file. Before, the gallery entry would disappear right away but the file lingered in storage.

Click Enhance now learns from real prompts that have produced great generations on PixelDojo. Instead of using a generic template, the enhancer pulls a handful of high-quality examples that actually worked on your chosen model and uses them to shape the rewrite. Think of it as having access to the prompt notebooks of every successful generation on the platform — 40,000+ curated entries plus everything in the community gallery.

Per-model styling. The enhancer now knows that what works for Flux is different from what works for Veo 3, and pulls relevant examples from each model's track record. Your enhanced prompts are tuned to the model you're using, not a generic image template.

Gets better the more you use it. Every time you save, download, or upscale a generation, that signal feeds back into the enhancer. Prompts that produce things you keep get weighted higher next time someone enhances a similar idea.

Drop-in upgrade. If retrieval can't find anything useful for your specific prompt, you get the same enhancement you've always had — no regression, no errors. The new system is additive.

The Canvas co-pilot now recommends models based on what the community is actually using — not what its training data thinks is famous. Top models per chain type (generate, edit, upscale, animate) are pulled live from real Canvas usage and refreshed every hour. As new models climb the leaderboard, the co-pilot picks them up automatically. You can still name any model in your prompt ("use Flux," "with Seedream") and it will honor your pick.

Clarity Pro Upscaler is here — photorealistic upscaling up to 16× with a creativity dial that goes from strict fidelity all the way to expressive re-imagining. Realistic skin, preserved identity, and output up to 64 megapixels (~8K). Find it in Upscale & Enhance.

The community gallery and all user profiles are now fully public — no account required to browse, discover, and get inspired. Share any gallery link with anyone.

Gallery image descriptions are now generated the moment you open a piece. Open any image in the gallery and a smart description appears automatically, giving every creation more context.

The gallery detail page got a layout refresh — likes, views, share, and vote are now in a single compact row so the image takes center stage.

Kling Video multi-shot mode was silently blocking generation in some cases due to a form validation issue. It now submits correctly every time.

Multi-Shot Mode on Kling Video 3.0 — write a separate prompt for each shot (up to 6) and Kling stitches them into one seamless clip. Toggle it on, describe your scenes, choose Custom or Auto pacing, and generate. Great for short narratives, product demos, or any video that needs distinct beats.

Creator Studio has been retired. If you had it bookmarked, the link now takes you straight to Canvas — the better tool for multi-model workflows.

A small welcome card on the dashboard now leads with the most important thing to know: "Generations live here for 24 hours." Save what you want to keep, otherwise it's gone tomorrow. Inline links to your last-24-hour Recent pill, My Media, and Buy Credits. Dismiss it once and it's gone for your account across every browser and device.

Tool empty states (the screen you see before you generate anything on most tools) now include the same 24-hour notice. Belt-and-suspenders so you see the message before you spend credits, not after you wonder what happened to your generation.

The Recent floating pill (bottom-left of the dashboard) just got a real job. Every row now has three icon buttons on the right: Save, Download, and Remove. Save promotes that generation into your permanent My Media library; Download grabs the file directly to your computer; Remove hides the row from Recent (it never deletes the file or anything in your gallery).

Recovery: if you accidentally delete a generation from My Media, Recent still has it for 24 hours. Hit Save in Recent and it goes right back into your library.

Saved items now show a green checkmark in Recent so you can see at a glance which generations are already in your permanent library and which are still just "in flight."

ImagineArt 2.0 (preview) is here. The next-gen entry in the ImagineArt family lands with sharper realism, stronger composition, and the most faithful prompt following yet. 1 credit per image.

The /imagineart page now hosts the whole ImagineArt lineup — 1.5, 1.5 Pro, and 2.0 — in one workspace. Pick the model from the panel on the right, no more page-hopping. The menu link and page title are now just "ImagineArt" instead of "ImagineArt 1.5."

Pasting a really long prompt into Qwen Image (or Qwen Image 2, or either of their Edit variants) used to bounce with a validation error — the model only reads 800 characters and we were rejecting anything longer outright. Now we trim the prompt to fit and pop a one-time warning toast telling you it happened, so the generation just runs. Same treatment for negative prompts (which cap at 500). If you're calling the API directly, the response now carries a `warnings` array describing what got trimmed.

Search in All Tools is much smarter. Multi-word queries now AND across tokens (typing "edit face" actually returns face-editing tools instead of zero results), category names are part of the searchable text (typing "face" surfaces everything in Characters & Faces, not just tools with "face" in the title), and a small synonym map covers common variations — "edit" → inpaint / modify / fix, "upscale" → enhance / sharpen, "video" → clip / animate, etc.

All Tools sidebar feels calmer. The 11 category icons used to all show in full color at rest — a parade of pink / red / blue / purple / etc that fought for attention with no clear hierarchy. Icons now render at half opacity at rest with a much subtler colored chip behind them, then snap to full color on hover or when their category is active. Same color identity, way easier to scan.

Reverted the mobile All Tools Pinned and Recent sections from two-column to single-column. The two-column was harder to read because tool labels and descriptions got tight at narrow widths.

The search box is now a permanent affordance in the top bar — click it (or press ⌘K) to instantly search across every tool, style, and model. Replaces the static Dashboard / Canvas / My Media / My LoRAs / My Prompts / Gallery / API row that used to sit there. Those destinations live one click away in All Tools' Quick Access section, and on the dashboard cards. Search is the right primary affordance when you have 100+ tools.

Up to 3 of your pinned tools now show as compact icon buttons in the top bar (on screens md and wider). Pin from any tool card and it appears in your bar; unpin to remove. Pin order is stable so your bar doesn't shuffle. The icon for the tool you're currently on lights up — your bar tells you where you are.

Theme toggle moved out of the bar and into the user menu (click your avatar). Theme is rarely changed once set, and freeing the bar slot leaves more room for what you actually use day to day.

Mobile All Tools is way less to scroll. Browse Categories now shows up first (instead of being buried under Pinned, Recent, Quick Access, and Featured Tools). Pinned + Recent fit two-per-row instead of stacked. Featured Tools stays on the dashboard hero where you can already see them. Net result: about 60% less vertical scroll before you reach the catalog.

Removed the duplicate "Account" category from All Tools (it had Usage Report, Profile, Subscription, Feedback — all of which live in the user menu next to your avatar, where you'd expect to find account stuff).

All top-bar buttons are now 40px tall (was 36px), comfortably above the 44px-ish touch target threshold once you account for the padding around the icon. Easier to tap on iPad and laptop trackpads.

Open All Tools and the Home tab is now structured as a real launchpad: your Pinned tools at the top, your Recent generations next, then a Quick Access strip with Dashboard / Canvas / My Media / My LoRAs / My Prompts / Gallery / API, then a curated Featured Tools section. Pinning your first tool no longer makes the Featured suggestions disappear — both coexist.

The drawer's left-rail tab is now "Home" instead of "Quick Start," with a calmer Home icon. The "Featured Tools" hero on /dashboard and the "Featured Tools" section in the drawer now share the same curated list, so when we promote a model, both surfaces update.

When you start typing in the drawer's search box, the left rail no longer keeps the previously-selected category lit up — search results are global, so the highlight matched nothing. Sidebar entries dim while a search is active.

The Recent floating pill (bottom-left of the dashboard) now self-heals. If a generation it remembers has been deleted from your gallery or the temp preview expired, the broken thumbnail row auto-hides and the count updates instead of showing an empty box.

Drawer chrome cleanup: section labels are sentence case (Pinned / Recent / Featured Tools / Found N tools) instead of WIDELY-TRACKED UPPERCASE; duplicate Account / Usage / Credits links removed from the drawer footer (they live in the user menu and the credits pill).

Removed the two large pulsing blue / sapphire ambient circles that animated behind every dashboard page. They were 5%-opacity decoration; the dashboard now feels a touch calmer and more tool-focused without them.

A floating Recent button appears bottom-left of every dashboard page once you've generated something. Click it to see every successful generation from the last 24 hours across every tool — image, video, edit, the lot — with a thumbnail, timestamp, and one-click jump back to the tool that made it. Built for the all-too-common moment of "I refreshed and now I can't find my video."

Qwen Image edits no longer ignore your aspect-ratio selection. Picking 1:1, 16:9, 4:3 (or any other ratio) now resizes the output instead of silently mirroring the source image's dimensions. Caught and reported by jazze#77 — thank you.

If your connection drops mid-generation — common on mobile or shaky Wi-Fi — you used to see a bare "Failed to fetch" toast with no context, leaving you wondering whether your credits had been silently spent. The toast now reads "Connection lost — your credits are safe" and explains that if a generation actually started, credits refund automatically within 2 hours. Check your generation history before retrying. Applies across every tool page (every browser: Chrome, Safari, Firefox, Edge).

The public API for /api/v1/models/wan-image/run now matches its own documented default — you can omit numberOfImages and get 1 image, instead of bouncing with "validation_error: numberOfImages: Required." The docs always said default: 1 but the schema was requiring the field; both sides now agree.

Generations that fail validation now show what went wrong instead of silently doing nothing. If your prompt is too long, your image too large, or a value out of range, a clear toast tells you exactly which field and what to change. Previously the Generate button just returned with no feedback — a long-standing papercut across Qwen Image, Qwen Image 2, WAN Image, ImagineArt, Z-Image Turbo, Dreamina, Ernie, and Ideogram Character.

Running out of credits now shows "Insufficient credits" with a link to top up, instead of a generic "generation.failed" toast. ~50 tool pages were checking the credit-error response in a way that quietly missed the real shape — they all detect it correctly now.

Error toasts no longer leak internal codes like "INPUT.INVALID_PARAMETER" or "CREDITS.INSUFFICIENT" as the title. The friendly message that was supposed to render now actually renders.

Every parameter on every API model docs page now has a description. Hundreds of fields used to render as just a name and a type with no explanation; they now carry plain-English context — what the field does, what the valid values mean, when to use it. Affects /api-platform pages, the OpenAPI spec, the LLM-readable docs at /llm, and Canvas's "More Settings" panel. All five surfaces auto-update from a single source so future tools inherit the descriptions for free.

Edit tools — Qwen Image Edit, WAN 2.7 Image Edit, GPT-Image 2 Edit, P-Image Edit, OpenAI Image 1 Edit, HiDream Edit, and a few others — now ship a multi-image example alongside the single-image curl. The capability was always there, but the docs only showed one URL; it was easy to miss that you can compose up to 3-9 reference images in a single call. The new examples make it obvious.

HiDream is now HiDream O1 — a brand-new image backbone replacing the previous L1 family. Two variants live behind one tool: Full (the production quality target) and Dev (a faster, distilled version). Pick between them with a one-click pill right above the prompt.

Pricing simplified to a flat 1 credit per generation and 1 credit per edit. No more dual-credit tiers — every HiDream call costs the same regardless of variant.

Aspect ratio moved inline with the composer. Pick Square HD, Square, Landscape 4:3 / 16:9, or Portrait 3:4 / 9:16 right next to the Generate button — no more buried Settings modal. The picker uses real aspect-ratio icons so the layout you'll get is visible at a glance.

Image editing got the same upgrade. The Edit endpoint now uses HiDream O1's image-conditioned mode (Full and Dev variants), still 1 credit per call, and accepts up to four reference images for subject-driven edits if you call the public API directly.

Cleaner workspace — the standalone Settings modal and the post-generation auto color correction option are gone. Both were holdovers from the old engine that didn't have a home in the new flow.

Public API users: /api/v1/models/hidream/run and /api/v1/models/hidream-edit/run are live with the new model + image_size schema. The OpenAPI spec and LLM docs picked it up automatically. The legacy /api/hidream and /api/hidream/edit endpoints continue to work for existing integrations.

Qwen Image 2 now defaults to 2048×2048 (was 1024×1024). If you weren't picking a size manually, every new generation gets 4× the pixels — noticeably sharper out of the box. The previous default was using a quarter of what the model can actually produce.

Qwen Image 2 unlocks a high-fidelity "Max" resolution tier: 2048×2048, 2688×1536, 1536×2688, 2368×1728, and 1728×2368. The 2K and 1K tiers stay as alternatives for faster generations.

Qwen Image Plus and Max gained three controls that were previously hardcoded: smart prompt rewriting (the model adds detail to short prompts), watermark on/off, and a custom negative prompt. All available in the public API and Canvas More Settings.

Qwen Image Edit (both the legacy and 2.0 variants) now respect a watermark toggle. Output sizes are also tightened to a documented set — invalid combinations get a clean error instead of a confusing 500.

Sending more than 1 output to Qwen Image Plus or Max used to bounce with an unhelpful upstream error — the underlying model is fixed at 1 image. You now get a clean message at request time telling you to use LoRA mode if you want multiple outputs in a single call.

Power users on the public API can now pin to dated model snapshots (e.g. `qwen-image-2.0-pro-2026-04-22`, `qwen-image-plus-2026-01-09`) for reproducible behavior across model rev bumps. The rolling names like `qwen-image-2.0-pro` keep tracking the latest.

Polished the look of the Qwen Image and Qwen Image 2 pages in light mode — Generate and Subscribe buttons now use the sapphire site CTA color, and the resolution picker reads cleaner. The new "Max" tier shows what it actually represents on hover (about 4 megapixels, the model's documented sweet spot).

WAN 2.7 Pro now generates true 4K images. Switch to Pro on /wan-2.7-image and set resolution to 4K — output jumps from 2048² up to 4096². Text-to-image only (the model caps editing at 2K).

Resolution and aspect ratio are now independent controls. Pick your size (1K / 2K / 4K) and your shape (1:1, 16:9, 9:16, 4:3, 3:4, 5:4, 4:5) separately, instead of hunting through a combined "size" dropdown with twelve mixed values. Standard model gets 1K and 2K; Pro adds 4K.

Edit Mode accepts up to 9 reference images at once, up from 4. Bring your whole mood board.

Interactive edit: in Edit Mode, toggle on "Interactive edit (bounding boxes)" and drag rectangles directly on each reference thumbnail to restrict changes to those regions. Up to 2 regions per image. Useful for outfit swaps, sign edits, or selective restyles without touching the rest of the image.

Custom color palettes in the Settings menu — bias a generation toward 3–10 specific colors with weighted ratios. Hex picker, percentage sliders, and an auto-balance button that distributes ratios to a clean 100%. Good for brand work or hitting a specific mood.

Seed control in the Settings menu. Pin a number for reproducible runs; leave blank for random.

All of the above is also live on the public API, the OpenAPI spec, and the LLM docs — scripts and agents pick everything up automatically.

The public API for Kling 2.6 Pro and Seedance 1.5 now accepts the video duration as a plain number (e.g. "duration": 5) as well as a string ("5"). Previously, sending the number — which most HTTP clients do by default, and which our own Seedance curl example showed — bounced back with "duration: Expected '5' | '10', received number." Both shapes work now.

Seedream 4.5 custom output sizes that fall outside the model's supported range used to fail with a cryptic provider error ("image size must be at least 3686400 pixels. Request id: 0217…"). You now get a clear message with your actual width × height, the real limit (about 3.7–16.8 megapixels), and a concrete suggestion like "increase to 2560×1440".

When the generation service behind P-Image, P-Image Edit, P-Video, and P-Video Avatar is briefly overloaded, you now see a plain "the generation service is busy right now — please wait a moment and try again" instead of a bare technical error like "PRUNA API error: 504."

When Canvas Co-pilot's planner refuses your request because the underlying model's safety filter flagged it, you now see "Co-pilot can't plan that — your request was flagged by the safety filter. Please rephrase your goal …" instead of the generic "Co-pilot couldn't draft a plan. Try again." The previous wording asked you to retry an identical prompt that would keep failing.

Nano Banana reference-image dimension errors used to surface as "Runware: Invalid reference image width. The reference image width must be an integer between 1 and 12288." — leaking the provider brand name and the internal parameter name. You now see a clean message that explains each side must be between 1 and 12288 pixels and recommends 4000px on the long edge as a safe target. Same cleanup applies to prompt-length rejections (the message now says exactly which range is supported, no brand prefix). Centralized in the Runware client so every Runware-backed tool — Image Relighting, Google Gemini Image, Pixverse v6, Consistent Characters, Veo 3.1, and Nano Banana — gets the cleaner messages.

WAN 2.5 in Canvas now accepts an empty prompt when you're animating an image. Previously, dropping an image into a WAN 2.5 step and hitting run without typing a description bounced back with a "Prompt is required" error from your own UI — even though the underlying model is happy to animate the image with neutral motion. The schema now treats the prompt as optional in image-to-video mode (text-to-video still requires one).

If your card got declined and you fixed it by starting a fresh subscription (instead of updating the card on the failing one), your account could end up marked as expired even though the new payment had gone through and the next renewal was already on the calendar. The two billing events were arriving seconds apart and the older one was overwriting the newer one. Subscriptions now stay active in this case, and a small number of customers we found in the bad state have already been put back to active automatically — no action needed on your end.

Canvas now has an AI co-pilot. Press / on the canvas (or click the spark icon on the right edge) and ask it for anything — "make 4 lifestyle variants of my active shot," "build a launch from shop.example.com/runner," "animate this into a 6-second reel." It drafts a plan, shows you the cost upfront, and one click runs the whole chain. Parallel-safe too: 4 lifestyle gens fire at the same time instead of waiting in a queue.

Plans collapse to a clean title + cost + Run button. Tap "Show steps" to see the numbered plan, or "Show graph" for a tiny DAG view of how steps depend on each other. Power-user detail is one click away when you want it; out of sight when you don't.

100-credit hard cap per plan, enforced both at the AI planner and on the server. If a plan would cost more, you get a clean rejection with a hint to ask for something smaller — no surprise bills.

Co-pilot runs use the same execution path as a manual generate — same credits, same automatic refunds on failure, same job poller, same telemetry. Failed nodes can be retried without re-running the ones that already succeeded.

Greenfield keyboard map: / and ⌘K both open the co-pilot. ⌘M opens the model picker. Esc collapses. The right-edge icon rail is 56px when collapsed so the existing single-generate workflow looks and feels exactly the same — power users don't pay for a feature they didn't ask for.

Grok Imagine now has a Quality tier alongside the original Fast tier. Pick Quality for noticeably sharper detail, more accurate compositions, and stronger text rendering — ideal for thumbnails, hero images, ad creative, or any final visual you'd ship to a client.

Choose 1K (1024px on the long edge) or 2K (2048px on the long edge) output on the Quality tier. 1K is faster and cheaper; 2K is the high-resolution deliverable. The default is 2K.

The Quality tier supports editing too — upload an image, describe the change, and get a high-fidelity edit at 1K or 2K. Works the same way the Fast tier edits do, just with crisper output.

Pricing stays simple: Fast tier is still 1 credit per image. Quality is 2 credits at 1K or 3 credits at 2K. The credit chip on the Generate button updates live as you switch tiers and resolution.

Canvas now offers Grok Image Quality as its own tile in the model picker — drop it into any chain when you want a cleaner final pass without leaving the canvas.

When Grok Image Edit can't generate an output — usually because xAI's safety filter quietly refused the edit, or the source image is in an unsupported format — you now see a message that names the likely cause and tells you what to do (rephrase the edit instruction, or re-upload the source image as a JPEG or PNG, then retry). Previously it surfaced a bare "Prediction failed:" with nothing after the colon, which gave you zero clue what went wrong.

pixeldojo:campaign — drop a product URL into your AI agent (Claude Code, Cursor, Codex, Cline, Windsurf, Zed, and any other MCP client) and get back a hero image, four lifestyle variants, and a vertical video, all from one call. About 50 credits, two minutes end to end. Perfect for a quick-turn ad pack or a social drop.

pixeldojo:from_url — paste any product page URL (Shopify, WooCommerce, BigCommerce, or anything with structured data) and get back the product name, description, and images. Used as the first step inside Campaign, but also handy on its own when an agent needs structured product info without copy/paste.

pixeldojo:generate now accepts a preset slug. Pass preset: "cinematic-portrait" or preset: "product-floating-clean" and skip the prompt-writing step entirely. The 60 expert-tuned presets in our library now work the same way through your agent as they do on the website.

Job status checks now show how long a job has been running. After 5 minutes your agent gets a "this is unusually long" warning; after 12 minutes it gets told to start a fresh request instead of polling forever. Fewer agents stuck in endless poll loops on the rare hung job.

The Skills page at pixeldojo.ai/skills now shows the new agentic tools front-and-center, with code snippets for each. New shortcut: paste pixeldojo.ai/skills.md into your AI agent and say "install this" — the page has a built-in directive that tells your agent exactly which command to run.

Nano Banana jobs that appeared stuck on "still processing" — even after the actual generation had completed — now resolve correctly. The status check was looking up jobs through the wrong backbone for the Pro and Pro v2 tiers, so it never saw the completion. Old stuck jobs unblock immediately if you retry; new jobs return URLs as expected. If you generated a Nano Banana image in the last day and it never came back, retry it now.

When fal.ai rejects a Smart Resize request (status 422), you now see the actual validation reason — e.g. which target size or input field was wrong — instead of the bare "Unprocessable Entity" status text. Same translation we already use for the other fal-backed tools, now wired into the Smart Resize submit path so you can fix the input on the first retry.

The rotating agent chip on the Skills page now shows the actual brand mark for each integration — Claude Code, Cursor, Codex, Cline, Windsurf, and Zed — instead of two-letter initials. The chip also sits properly inline with the headline now, lined up with the cap of the surrounding letters instead of dropping below the baseline.

The install steps under the hero are tidier — Step 01's title fits on one line so all three columns line up, and the agent list across the page (hero subtitle, install card, and "Works with" pills) is consistent end-to-end.

The hero headline on the Skills page now cycles through the agents that support PixelDojo — Claude Code, Cursor, OpenClaw, Codex, Hermes, NemoClaw, and Cowork — each with a small color-coded chip so you can immediately spot your setup.

The install section is now three numbered steps (01 / 02 / 03) instead of icon cards, making the path from zero to working agent quicker to skim.

The new /presets page lets you browse all 60 workflow presets in one place — from product photography and fashion to storyboards and text rendering. Every card links directly to the matching tool with the preset loaded.

The preset library grew from 24 to 59 entries, all with real generated example images. Each thumbnail was produced with that exact preset so you know what output to expect before you click.

The Skills page now has a live terminal demo showing exactly what the agent conversation looks like: you describe what you want in plain English, it routes to the right model, and you get back an output URL. Makes the workflow concrete instead of abstract.

The hero on the Skills page now cycles through real agent use cases — things like "Build a 4-shot product storyboard" — so it's immediately clear what you can hand off to your LLM.

A "Build with PixelDojo" card is now on the main dashboard, linking developers straight to the Skills page and API Platform from day one.

Seedream 4 and Seedream 5 Lite now translate ByteDance's "The request failed because the image format is not supported by the API. Request id: 0217..." into a clean message that names the likely cause (HEIC, AVIF, GIF, or a corrupted file) and tells you what to do (re-upload as JPEG, PNG, or WebP). The internal request id no longer leaks to users.

Reference-image size errors on Nano Banana and Consistent Characters are clearer. The previous message — "Please re-upload an image between 5kb and 20MB" — fixed the unit casing and now explains both failure modes: if your image is too large, compress or resize it (under 4000px on the long edge usually works); if it's smaller than 5KB, the file may be corrupted or didn't finish uploading. Includes the supported formats (JPEG, PNG, WebP) so you know exactly what to send back.

When Seedream rejects a prompt for sensitive content or copyright reasons, you now see a clean "Generated content was blocked by the model provider — try a different prompt or input media" instead of the raw "The request failed because the input text may contain sensitive information. Request id: 0217..." envelope. Credits are refunded the same way; the change is just the message you see.

If you ever saw the orange "You're out of credits" banner while your header still showed plenty of credits, that's gone. The banner could latch on after a brief subscription-status hiccup or a top-up and stay visible until you dismissed it manually — even with hundreds of credits in your account. It now re-checks the live state on every render and hides itself the moment your balance is healthy again.

@pixeldojo/mcp ships on npm. Run `npx @pixeldojo/mcp init` for copy-paste install snippets, set PIXELDOJO_API_KEY, restart your agent, and you've got five named tools your LLM can call: pixeldojo:generate (any prompt → image or video), pixeldojo:character (consistent characters across shots), pixeldojo:storyboard (multi-shot scenes from one brief), pixeldojo:upscale (enhance any image), and pixeldojo:status (check on long jobs). Works with Claude Code, Cursor, OpenClaw, and any other Model Context Protocol host.

Smart polling for long jobs. Fast generations come back in the same call. Jobs that take longer than 30 seconds — LoRA training, long-form video — return a job ID and a status URL, and your agent calls pixeldojo:status to check on them. No more 5-minute MCP timeouts that look like failures. Polling uses exponential backoff (1s → 2s → 4s → 8s) so we don't hammer your account's rate limit.

Your API key never appears in your agent's logs. The MCP server strips any pd_xxx token from every log message before it leaves the process — even debug output. Same for error messages: a 401 surfaces as "Set PIXELDOJO_API_KEY" with a link to the keys page, never as a stack trace with your token in it.

Recraft tool page now respects the ?preset=<slug> URL param. Click a preset thumbnail anywhere on the site and the recraft form opens with the prompt and aspect ratio already filled in. The other preset-having tools follow next.

PixelDojo Skills replaces the developer-agent landing at /agents. The new /skills page leads with a one-line install for AI-agent tools (Claude Code, Cursor, OpenClaw), covers four named skills your agent can call (generate, character, storyboard, upscale), and shows real preset thumbnails you can try without writing a prompt. Old /agents and /agents.md links keep working — they redirect to the new home.

Try-a-preset on every tool docs page. Visit any model's page on the API Platform and you'll see a row of curated starting points — cinematic portraits, product-on-marble shots, magazine covers, and more. Click one and we'll open the tool with those parameters loaded, so you can play before you write a single line of prompt. Twenty-four presets ship today across the most-used image and video models, with more on the way.

Use-case filter chips on the API Platform. Browse 130+ models by what you actually want to make: marketing, character, cinematic, fashion, photorealism, text-in-image, upscale, video, and more. The filters are deep-linkable — share /api-platform?useCase=character with a teammate and they land on the right shortlist. Empty results now ship with a one-click clear-filters button instead of a dead-end message.

The Skills page tells the truth about scale. The hero now reads 130+ models — matching what's actually on the API Platform — instead of the old 40+ figure. The model showcase pulls live counts from the registry, so the number you see is the number you can call.

Sapphire-aligned design across the developer landing. The fuchsia-violet-cyan gradient look is gone, replaced with the sapphire palette that already drives the rest of the product. Cleaner hero, cleaner CTAs, less visual noise — the page feels like a workspace tool, not a marketing template.

If a renewal payment failed and your subscription ended, the Subscribe page could keep telling you "You're Already Subscribed" — sometimes for an hour or more — and the Subscribe button stayed locked out. The page now reflects billing changes within a minute, so re-subscribing after a card decline works the way you'd expect.

Submitting WAN 2.2 with an unsupported quality_mode + resolution combo (e.g. quality_mode "plus" with resolution "720p" — Plus only supports 480p and 1080p) used to fail opaquely as "credit_error: Invalid credit amount" after the request was already accepted, sounding like an account issue. You now get a schema validation error upfront that names the exact resolutions your chosen quality mode supports — e.g. "Resolution '720p' isn't available on the plus quality mode — pick one of: 480p, 1080p."

Switching models in Canvas no longer leaks settings from the previous tool. Picking Qwen Image 2 with the Pro variant and then switching to WAN 2.7, WAN, or ImagineArt used to send the previous tool's `model` value to the new one, failing with "Invalid enum value, expected 'wan-2.7-standard' | 'wan-2.7-pro', received 'qwen-image-2.0-pro'". Same root cause for sizes (1280×1280 / 1536×1536) and durations (numeric vs. string) leaking across model switches. Canvas now resets all advanced settings whenever the active model changes — re-selecting the same model preserves your current picks.

Seedream 5 Lite custom-size mode now works end-to-end. Picking the "Custom" image size with your own width and height used to be rejected at the schema gate as "Invalid enum value, expected 'auto_2K' | 'auto_3K' | 'square_1_1' | … received 'custom'" even though the server already supported it. Custom dimensions now flow through cleanly, matching how Seedream 4.5 already worked.

Sending more than 14 reference images to Nano Banana used to surface the raw provider error "Runware: Invalid number of elements for 'referenceImages' parameter" — confusing brand and parameter names with no guidance. You now get a clean message that includes your actual count and tells you exactly how many to remove (e.g. "You sent 20 reference images. Nano Banana supports up to 14 — please remove 6 and try again.").

WAN 2.5 in Canvas no longer rejects 3s and 4s durations and now lets you go up to 15s. The duration slider used to span 3-10s, but the model's actual range is 5-15s, so picking 3 or 4 failed with "Number must be greater than or equal to 5" and the 11-15s range was hidden. Slider now matches the model exactly, and stale durations from sibling tools (WAN 2.2, Hailuo, LTX) auto-clamp into range.

WAN Reference to Video in Canvas no longer fails when you switch in from a sibling tool with a different resolution or aspect-ratio picker. Stale '480p' values from WAN 2.2 and unsupported aspects like '3:2' used to slip through and trigger "Invalid enum value" — Canvas now defensively falls back to 720p / 16:9 when an incompatible value is in flight.