P Video Avatar is live. Upload a portrait, write a short script (or paste an audio URL for lip-sync), and get a talking-head video. 30 named voices, 10 languages — English (US/UK), Spanish, French, German, Italian, Portuguese, Japanese, Korean, Hindi. 1 credit per second at 720p, 2 credits per second at 1080p. Find it under Characters & Faces in the All Tools menu.
Switching to Hailuo 02 in Canvas after a model with a different resolution picker (WAN 2.5, PixVerse, etc.) used to fail with "resolution: Invalid enum value, expected '768p' | '1080p', received '720p'" because Canvas carried over the previous model's value. Canvas now exposes Hailuo's exact resolutions and falls back to 768p when an incompatible value is in flight.
LTX 2 in Canvas no longer rejects odd-numbered durations. The duration slider used to step by 1 (so picking 7s or 11s failed with "Duration must be one of: 6, 8, 10, 12, 14, 16, 18, 20 seconds"); it's now an explicit list of valid lengths. Picking longer than 10s on the Fast tier also auto-locks 25fps + 1080p so the cross-tier requirement is always satisfied.
Runway Gen-4.5 in Canvas now constrains duration to 5s or 10s the same way the standalone Runway Video tool does, so stale 6s/8s values from sibling models no longer leak through and fail with "duration: Invalid input".
Happy Horse Reference to Video is here. Upload up to 9 reference images (characters, props, costumes), describe your scene with character1 / character2 tokens, and the model generates a video with each subject locked to its image. 720p or 1080p, 3-15 second clips, 4 credits per second at 720p and 6 at 1080p.
Happy Horse Video Edit lets you transform existing footage with a prompt — style transfer ("make this look like a Studio Ghibli film") or local replacement ("swap the sweater for the one in the reference image"). Pair a source clip with up to 5 reference images, choose to keep your original audio or have new audio generated. Same pricing as the rest of the Happy Horse family — 4 credits per output second at 720p, 6 at 1080p. Find both in Generate Videos under WAN Family.
Happy Horse 1.0 is now live. Generate fast, expressive video from text or a starting image at 720p or 1080p, with clip lengths from 2 to 15 seconds. 4 credits per second at 720p, 6 credits per second at 1080p. Find it under Generate Videos in the All Tools menu.
Public API calls to qwen-image-2.0, qwen-image-2.0-pro, qwen-image (Plus/Max), the Qwen Image Edit model in image-relighting, and the Qwen Image 2 Pro model in consistent-characters were hanging indefinitely. The dashboard worked fine because it polls a different endpoint, but the public API's polling didn't recognize the internal job ids these tools were returning. They now return image URLs directly when called via the API — first /api/v1/jobs/{id} GET serves the result. Dashboard / Canvas / Creator Studio behavior unchanged.
The Qwen Image Edit, Qwen Image 2 Edit, and HiDream Edit endpoints are now reachable via the public API at /api/v1/models/{qwen-image-edit, qwen-image-2-edit, hidream-edit}/run. Sister set to Generation now matches every dashboard tool 1:1 — every customer-facing capability has a public apiId.
Each /api-platform/{apiId} doc page now shows a "See also" section linking related apiIds (gen ↔ edit pairs across 16 tools: Flux, GPT Image, Nano Banana, Kling, P-Image, Qwen, Seedance, WAN, XAI, HiDream). Customers building integrations can discover companion endpoints without leaving the docs.
Canvas's "More Settings" panel now auto-populates for any new tool — derived from the tool's schema enums + inputParameters. Add a tool.config.ts and the tool gets a working settings panel for free, no manual registration in the Canvas params table.
Two new CI gates so /api-platform docs and Canvas can't drift from the schemas: canvas-param-schemas-parity (every Canvas option must be schema-acceptable) and no-passthrough-schemas (schemas reject unknown fields, customers see clean 400s instead of silent drops).
CLAUDE.md now formally documents the single-source-of-truth principle: tool.config.ts is canonical. Adding a tool means writing one file — every doc surface, the public API, Canvas, Creator Studio, and OpenAPI all derive from it automatically.
Closed the rest of the schema-vs-docs drift across the API platform: 11 more tools (Hailuo 02, Image Relighting, Pony XL, Qwen Image 2, Recraft, Seedream 4.5, Seedream 5 Lite, WAN 2.6 Image, WAN 2.6 Image Edit, WAN 2.7 Image, WAN 2.7 Image Edit) had documented enum lists for parameters like resolution, image_size, scheduler, output_format, and aspect_ratio while the schema accepted any string. The schema now enforces the same set the docs advertise — bad values fail fast with a clear 400 instead of bouncing off the provider after credits are deducted.
Added a reverse-direction parity check to CI (tests/contracts/tighten-loose-string-audit.test.ts): every documented enum value must be enforced by the schema. Combined with the existing forward check (every advertised value must be acceptable), the docs and schemas can't drift apart again.
Aspect-ratio and resolution fields on the public API now reject invalid values up-front instead of being passed through to the provider for a confusing 422. ~16 tools (Flux, Flux 2, Flux Edit, Google Gemini Image, Nano Banana, Dreamina, Seedream 3, P-Image, P-Image Edit, Qwen Image, WAN 2.2 / 2.5 / 2.7 video, WAN Image, Ideogram Character, HiDream, Image Editor, Video Reframe, Video Transform, WAN 2.2 Animate) had loose `z.string()` validation that let unsupported values through silently — they now use the same enum the dashboard offers.
WAN 2.5 / 2.7 video API now accepts the documented `720p`/`1080p` form (the dashboard's value) — previously the schema only accepted uppercase `720P`/`1080P` even though every doc said lowercase. The API now matches the docs and continues to send the upstream-correct uppercase form to DashScope on your behalf.
HiDream resolution is now a proper enum on the API (`1024x1024`, `1360x768`, etc.). The dashboard's human-friendly form (`1024 × 1024 (Square)`) still works, so you can use either.
Runway Gen-4.5 was 500'ing on every Canvas chain because the duration picker emitted "5" or "10" as text but the schema expected a number. Both formats now work, so animating a Canvas frame with Runway goes through on the first try.
LTX-2 was occasionally rejecting Canvas runs with "FPS must be one of: 24, 25, 48, 50" when the FPS dropdown was set — same number-as-text mismatch. Coerced at the schema, so any FPS option you pick now works. Also added a guard that drops stale resolution values (like 480p left over from a previous WAN step) and falls back to 1080p instead of failing the run.
Consistent Characters and Google Nano Banana now show "Your reference image is outside the allowed size range. Please re-upload an image between 5KB and 20MB and try again." or "We couldn't fetch your reference image — the link may be unreachable or the format unsupported. Please re-upload (JPEG or PNG) and try again." instead of a raw Runware JSON envelope with internal taskUUIDs and documentation URLs.
Image-edit endpoints for GPT Image 2, GPT-Image 1.5, and OpenAI Image 1 are now available via the public API: POST /api/v1/models/gpt-image-2-edit/run, /gpt-image-1-5-edit/run, and /openai-image-1-edit/run. Send a prompt plus image_urls (or input_images) and the same edit pipeline the dashboard uses runs end-to-end — image URL sanitization included. Existing /api/v1/models/{wan-2.7-image-edit, wan-2.6-image-edit, kling-image-edit, p-image-edit, xai-image-edit}/run already worked the same way; this fills in the orphans.
Each new edit apiId is documented on /api-platform with example curls and shows up in /llm/{apiId}.txt for AI-assistant-driven integrations.
API documentation now matches the schemas the public API actually validates against. Previously several tools' /api-platform pages and /llm/{model}.txt docs advertised parameters the schema would silently drop — for example `num_outputs` listed on Kling, Qwen, WAN, Seedream, Imagineart, P-Image, and others when the underlying schemas use `num_images`, `numberOfImages`, or `n` (or only generate one image per call). Renames + removals applied across ~15 tools so the names you read are the names the API accepts.
Google Nano Banana 2 and Nano Banana Pro (text-to-image variants under /api/v1/models/google-gemini-image/run) advertised `reference_images` for guided generation but the schema didn't accept the field. Reference images now work end-to-end on those variants — same image URL sanitization and Runware backbone the dashboard uses.
/llm/{model}.txt now derives parameters from each tool's Zod schema when no hand-written list exists — so reference-image fields, custom dimensions, and other inputs always show up in the LLM-friendly docs even for tools we haven't manually documented yet. Matches the same fallback /api/v1/models/{model} already does.
Heygen Avatar API parameters renamed to match the schema: `expression` → `expressionHappy`, `background.type` → `backgroundType`, `background.value` → `backgroundValue`. The dashboard already used these names; the API docs now agree.
Deleted ~600 lines of stale hand-rolled API model entries that had drifted from each tool's actual schema (e.g. a static `nano-banana` apiId that competed with `google-nano-banana`, a `seedream-4` entry advertising `reference_images` while the live schema accepted `image_urls`). Every public apiId now comes from a single source of truth: the tool's own tool.config.ts.
If you upload an unsupported image format (HEIC, AVIF) or a corrupted file to Seedream or Magnific, you now see "Could not read your reference image — the file format may be unsupported (HEIC, AVIF) or the image may be corrupted. Please re-upload as JPEG or PNG." instead of a raw provider JSON dump or Python traceback.
When a referenced image URL returns 404 / 403 / 401 (link expired or access denied), Grok Imagine Edit and other image-editing tools now say "We couldn't load your image — the link returned 404 (not found). The image may have expired or moved. Please re-upload and try again." rather than "Failed to fetch image: 404".
Canvas was offering a 480p option for WAN 2.5 that the backend doesn't support (silently failed every run). Removed it from the Resolution dropdown. WAN 2.7 Video and Seedance 2 also now defensively normalize stale resolution values from previous Canvas steps so chains like "Kling Image → Seedance 2 video" no longer carry over the wrong setting.
WAN family + Qwen models running on Alibaba DashScope now show "Alibaba DashScope is temporarily unreachable. Please try again in a moment." instead of a raw "fetch failed" error when the upstream service is briefly down. Same treatment for ByteDance / Seedream when its servers are overloaded — credits are still refunded automatically.
WAN 2.2 image-to-video was occasionally rejecting valid runs with "sample_shift: Must be greater than or equal to 1" from Replicate. The schema now bounds the field correctly and clamps any out-of-range value, so the run goes through cleanly.
Reference-image fields on public API endpoints (Nano Banana's image_input, Seedream's image_urls, Veo 3.1's reference_images, and every other tool that accepts source or reference URLs) were being silently dropped because the public API bypassed the per-tool payload shaping that tool pages, Canvas, and Creator Studio already used. Public API calls now run through the same submission pipeline — including image URL sanitization, multi-tier routing (e.g. Veo 3.1 Lite correctly uses its own backbone), and field-name translation — so a curl to /api/v1/models/{id}/run behaves identically to the same model on the site.
Synchronous image models (Seedream 4.5, Grok Imagine, Z-Image Turbo, Recraft sync, Nano Banana Pro) now return completed output on the first /api/v1/jobs/{id} poll — no more spinning while waiting for a provider job that never existed.
API calls to LTX-2 sub-models like /api/v1/models/ltx-2-pro-i2v/run were silently being validated as text-to-video instead of the mode the endpoint name implies, so callers got confusing errors about "text-to-video" rules while posting to an image-to-video endpoint. Each variant's mode is now applied automatically from the endpoint you call — a minimal {prompt, image_url} call to ltx-2-pro-i2v just works.
LTX-2 fps validation used to return "fps: Invalid input" with no list of valid values. It now says "FPS must be one of: 24, 25, 48, 50." Aspect ratio errors similarly spell out the supported values and when each applies.
The Community Gallery and user profile galleries are now behind sign-in. Browsing generations, exploring model pages, clicking into individual images, and viewing other creators' profiles all require a PixelDojo account going forward — the same account you use for the dashboard. Signed-out visitors get bounced to sign-in and routed back to the page they wanted.
Video Upscaler and Video Transform were throwing "Prediction ID is missing in the response" on every submit after a recent behind-the-scenes change. Both tools are working again — upload your clip and go.
If a P-API model (P-Image, P-Video) can't reach its upstream service, you now see "Pruna AI service is temporarily unreachable. Please try again in a moment." instead of the raw "fetch failed" error. Credits are still refunded automatically.
Canvas was failing silently on Flux Image-to-Image and Face Enhance when you tweaked More Settings like Guidance Scale, Prompt Strength, or Scale Factor — the inputs emit text and the schema expected numbers. Both now accept either, so your Canvas runs go through on the first try.
Canvas was offering a 21:9 aspect ratio for Grok Imagine that the backend doesn't support, so that pick always 500'd. Canvas now only shows the 7 aspect ratios Grok actually accepts.
If you pick Flux Pro or Flux 1.1 Pro with 21:9 or 9:21, you now see "Flux 1.1 Pro does not support 21:9 — try 16:9, or switch to Flux Schnell or Flux Dev for ultra-wide ratios." instead of a raw 422 dump from Replicate. No credits spent when the pre-check rejects.
Creative Upscale was occasionally 500'ing from Canvas when Creativity or Resemblance sliders landed outside the provider's range. Values are now clamped to what Clarity Upscaler accepts (Creativity 0–1, Resemblance 0–3) so the run goes through cleanly.
Prompt safety checks now run server-side on every tool, Canvas, Creator Studio, the public API, and LoRA training fields — not just the dashboard pages that had the client-side check. Three layers: a public-figure name list, a deterministic phrase scanner, and an AI moderation backstop. Anything flagged returns a clear, actionable message about what to change.
The image safety check that previously ran only on a couple of tools now applies to every tool that can produce explicit imagery, and uses a more reliable classifier than before. Photos that depict children are rejected with a specific message. Other tools (upscalers, background removers, analyzers) are unchanged.
Every image stored on PixelDojo is now eligible for CSAM hash scanning at the storage layer. All upload paths — generation outputs, reference images, and LoRA training datasets — route through the same scanning coverage.
When an uploaded image is over our 12 MB cap, we now show your file's actual size and tell you to resize or compress — no more guessing why the upload failed. Applies to every tool that takes an image.
On Seedream 4, if your reference image is bigger than the 36-megapixel limit or the host is slow to respond, you now get a clear message with the actual dimensions and next steps — instead of a raw technical error.
You can now paste Civitai model URLs from the new civitai.red domain anywhere we accept Civitai links — LoRA training, model imports, and HuggingFace uploads all recognize it automatically.
Z Image Turbo — the sub-second photorealistic model — now has the sticky mobile composer. Describe, pick aspect ratio, add an optional LoRA, and generate up to four at a time. Eighteen tools on the new shell.
ImagineArt 1.5 now has the sticky mobile composer. Describe your image, pick aspect ratio and model tier, and generate up to four at a time from a single pull-up panel. Seventeen tools on the new shell.
Portrait Series Generator (Consistent Characters) now has the sticky mobile composer. Pick a source portrait, choose a model, set a background, and generate up to 12 portraits — all from a single pull-up panel. Sixteen tools on the new shell.
Lip Sync, Video Autocaption, and Video to Sound all got the sticky mobile composer. Fifteen tools on the new shell now. The interior UI on these three is still pre-design-system (custom headers and buttons) — a polish pass will line them up with the rest.
Text To Music now has the sticky mobile composer. Describe the music, drag the length slider, tap an example prompt, generate — all from a single pull-up panel. Eleven tools on the new shell.
Hunyuan 3D now has the sticky mobile composer. Pick text or image input, tweak advanced settings, and generate — all from a single pull-up panel. Ten tools on the new shell.
Subject Control, PuLID Characters, and Image to 3D have all been removed. Subject Control overlapped with Virtual Try-On and Style Transfer. PuLID and Image to 3D were both on disk but never made it into the sidebar menu — nobody could navigate to them. Consistent Characters covers identity preservation, Hunyuan 3D covers 3D generation.
Dreamina 3.1 in Canvas now only offers 1K and 2K for resolution — the two values the model actually accepts. If your Canvas session had 4K or HD carried over from another model, Dreamina runs were failing at submit; stale state is now politely remapped to 2K instead of erroring.
Magic Lighting now has the sticky mobile composer. Upload, pick a model, tap a lighting preset, and generate — all from a single pull-up panel. Eight image tools on the new shell.
Style Transfer now has the sticky mobile composer. Source, style, prompt, and advanced settings all live in one pull-up panel. Seven image tools on the new shell now.
Change Camera Angle now has the sticky mobile composer. The orbit visualizer, sliders, presets, and generation options all live in one panel you can slide up from any scroll position. Six image tools on the new shell now.
Virtual Try-On now has the sticky mobile composer. Upload your garment, pick your model, and generate from any scroll position — five image tools on the new shell now.
Portrait Upscaler now has the sticky mobile composer. Four image tools on the new shell so far — upload, set your scale factor, and generate from any scroll position on mobile.
OmniHuman's 'Generate from Text' button works again. A plumbing change a week ago had disconnected the text-to-speech response from the page — you'd click Generate Audio, nothing would appear, no error. Now it polls to completion and shows the audio ready to use. Caught while sweeping other tools for the same class of bug as Background Remover.
Background Remover now actually shows your finished image again. A recent plumbing change had disconnected the final result from the page — you'd click Remove Background, wait, and get sent back to the empty state with no image. Fixed, verified end-to-end.
Extract Frame and Image Analyzer now have the same sticky mobile composer as Background Remover. Tap the bar at the bottom to open controls, tap again to see your result. Three image tools on the new shell now; more following on the same cadence.
Canvas signing now rotates on a 15-minute window, matching Creator Studio. If a signed link ever leaked from the page, its useful life is minutes, not hours. Zero visible change to how Canvas works for you.
On mobile, Background Remover now has a sticky composer at the bottom of the screen — tap to open the upload controls from any scroll position, tap again to collapse and see your result. Same pattern video tools have used for a while. First image tool to get it; the rest will follow.
Under the hood, image and video tools now share one shell. Three duplicate components (ToolWorkspaceShell, ToolPageHeader, ToolEmptyState) replace six parallel ones (VideoToolShell, VideoPageHeader, VideoEmptyState + image siblings). Every page you touch lives on the same plumbing now — future polish lands everywhere at once.
Every Creator Studio frame and video generation now flows through the same reliability backbone as tool pages and Canvas. If a model has an issue, we spot it and recover the same way whether you hit it from a tool page, Canvas, or Creator Studio. Zero visible change to your workflow — just one shared plumbing instead of two.
Flux LoRA trainers — all six of them — now use the proper Flux purple gradient on their Generate buttons and chrome. Orange-red and yellow-orange drifts are gone. Every Flux-family tool, training or generating, speaks the same color.
Softer corners across marketing and landing pages — hero cards, pricing tiles, community gallery, model showcases, and tutorial blocks now all share the same corner radius as the rest of the product. 42 files touched. The dashboard-to-landing transition is visibly one product now.
Image tools and video tools now share one empty-state component. Video tools keep their icon pills and info tips (like "Lite supports 720p or 1080p in 16:9 and 9:16"), image tools keep their colored-dot pills, and the visual style is consistent across both. Under the hood this eliminated a whole duplicate component tree.
The shared component library had a longstanding bug where every 'purple' accent was silently rendering as sapphire (the primary brand color) instead of actual purple. Across the Flux family (Flux Studio, Flux 2 Flex, HiDream, PonyXL, LoRA Image Creator, Feedback) that meant empty states, prompt boxes, aspect ratio pickers, credit cost displays, and mode toggles were all speaking the wrong color. Fixed at the source — 16 shared components across tool-components, video-components, history-grid, pricing-plans, admin, and my-images. Every purple-accented surface now renders actual purple.
Kling Image and Text to Music Generate buttons moved to amber — the proper Kling-family and audio-category color. Matches the rest of the Kling and audio surfaces across the product.
xAI Image's Generate button cleaned up from a sapphire-to-rose mix to pure rose — consistent with xAI Video and the rest of the xAI family.
Magnific Upscaler, Creative Upscale, and Google Gemini Image are now the category-correct blue on Generate and chrome, instead of sapphire, yellow-amber, or blue-purple mixes.
Ernie and P-Image are now pure cyan — matching Qwen, Recraft, ImagineArt, and the other cyan-family tools. No more stray rose or sapphire on their chrome.
LoRA Image Creator's cost display and Generate button are now the Flux-family purple gradient throughout instead of drifting back to sapphire.
The Subscribe CTA on the Qwen Image 2 page moved from a mixed emerald-teal gradient to solid sapphire — matching the product-wide rule that site-level actions like Subscribe are solid, while per-tool Generate buttons carry the gradient.
Video utility tools (Video Reframe, Video Transform, Video Upscaler, Lip Sync, Merge Videos) now share a consistent green across their Generate buttons and chrome — matching the category identity.
Subject Control, a character tool, now uses rose — the character-tool family color — instead of sapphire.
The Subscribe page CTA moved from a sapphire-to-blue gradient to solid sapphire — site-level actions like Subscribe are solid sapphire across the product now, gradients are reserved for the single Generate button inside each tool.
OpenAI Image 1's Generate button moved from a blue/sapphire gradient to emerald — the proper OpenAI-family color, consistent with GPT Image 2 and GPT Image 1.5.
Video to Sound's Generate button now uses amber — the audio-tool family color — instead of blue. Matches Text to Music and other audio surfaces.
Softer corners on cards, sheets, and empty-state tiles across 17 tool pages — from Flux Studio to Seedance 2 to Purchase Credits. The whole product sits at the same corner radius now instead of a handful of pages looking bubble-rounded.
Flux Studio's Generate button now uses the proper Flux-family purple gradient instead of blue — matches PonyXL and restores the family visual identity. The Subscribe-to-create CTA was cleaned up to solid sapphire (site-level actions are sapphire across the product now).
Removed a shimmer sweep animation from Generate buttons on Flux Studio and Seedream 4, plus the focus glow behind their prompt boxes. Tool surfaces are meant to be quiet; your prompt and your outputs hold focus now.
Toned down all-caps-shouting button labels on Billing, Settings, My Prompts, Merge Videos, and Usage Report. Labels are readable at a glance now instead of broadcasting at you.
Creator Studio's chrome now matches the rest of PixelDojo. Buttons and labels across the project header — Save, Characters, Scene, Shot Settings — use the same quiet treatment you see on every tool page. Your project title reads in title case instead of all caps.
Softer corners on the pre-project picker cards and the shot settings sheet. Same radius as the rest of the product, so Creator Studio sits next to Canvas and your tool pages without looking like a different app.
The in-project workspace is calmer. Removed a decorative background glow so your storyboard holds focus. Ambient motion still lives on the pre-project picker where it belongs.
Seedance 2, Seedance 2 Reference, and Seedance 2 Video Edit have moved to a new backbone. Same models, faster and more reliable queuing — no more jobs getting stuck on "queued."
Most Seedance 2 generations now cost fewer credits. Fast 480p is down to 2 credits/sec and Fast 720p to 4 credits/sec. High 720p is 4 credits/sec (5 with a reference video), and High 1080p is 10 credits/sec (12 with a reference video). The quality popovers and Canvas cost displays reflect the new rates automatically.
Seedance 2 Video Edit keeps Standard tier and full 480p / 720p / 1080p output; reference images and reference videos work the same way you're used to.
Your dashboard now recognizes whether you've generated before. If you're a returning creator, the top of the page becomes a compact welcome strip with Create Image and Create Video ready to click — and your Recent Creations move up to become the visual hero. First-time visitors still see the full onboarding intro so they know where to start.
The home page buttons are now solid sapphire across the board — no more yellow Buy Credits pill, no yellow hover flash on Create Image, no gradient Subscribe button. Cleaner, more consistent, and it lets your art stay the hero.
Featured Tools on the dashboard now show honest category labels — Image, Video, Editing — instead of marketing claims like "Best Quality" or "Most Flexible." You can see at a glance what each tool is for.
Video tutorials got their own home at /tutorials. The Master AI Creation video library moved off the main dashboard to a dedicated page, so your workspace stays focused on getting work done. You can jump to them anytime via the new Videos link in the Prompting Guides section.
The returning-visitor welcome strip keeps a Buy Credits (or Subscribe for free accounts) button right alongside Create Image and Create Video — styled as a subtle sapphire-accented pill so it's always one click away without competing with your creative actions.
GPT Image 2 Prompting Guide is live at /guides/gpt-image-2-prompting-guide — a hands-on playbook with seven generated examples covering near-perfect text rendering, dense infographic layouts, multilingual signage (Latin + CJK), world-aware photorealism, product mockups, character consistency, and UI mockups. Every prompt is copyable.
Every prompting guide now has a clean breadcrumb trail — Home › Prompting Guides › <Title> — so you always know where you are and can jump back to the index with one click.
The site footer has a new "Prompting Guides" column led by an "All Prompting Guides" link to the index, plus direct links to every major guide including GPT Image 2, Nano Banana Pro, Flux 2, Qwen Image 2, Seedream 5, Veo 3.1, Kling 3, and Seedance 2.
The /guides index now lists every guide we've published.
The AI Image landing page now leads with GPT Image 2 across its hero, FAQ, and structured data, and features a new Prompting Guides section so you can browse model-specific playbooks while choosing a generator.
Uploading an unusually large reference image to Nano Banana Edit (Pro tier) used to return a raw provider JSON blob. You'll now get a short, actionable message telling you the image dimensions are outside the supported range and to resize to within 1–12288 pixels per side before retrying.
GPT Image 2 is live at /gpt-image-2 — OpenAI's next-generation image model, with native 4K UHD rendering and noticeably sharper typography. Great for hero art, editorial covers, and print-ready marketing.
Six explicit render sizes in the size picker: 1024×768, 1024², 1024×1536, Full HD 1920×1080, QHD 2560×1440, and 4K UHD 3840×2160. Each 4K tile wears a small amber "2×" chip so you know when you're on the premium tier.
Tiered credit pricing that tracks compute honestly. Low quality is a flat 0.5 credits per image at any size. Medium is 1.5 credits (3 at 4K UHD). High is 5 credits (10 at 4K UHD). The quality popover shows the live per-image cost for the size you picked, so there are no surprises.
GPT Image 2 also supports edit workflows — upload or pick reference images and describe your change. Multi-image refs, optional edit masks, and the same 4K rendering carry through.
Available in Canvas as a full generate step with size, quality, and format controls in More Settings, and exposed in the public API as the gpt-image-2 model ID.
The Seedance 2 duration control in Canvas now shows the three durations the model actually supports — 5s, 10s, 15s — as discrete chips instead of a free-range slider. The slider let you pick 6 or 7 and then the model rejected the job; those nudges are no longer possible.
PixelDojo is now discoverable by AI agents and LLM-powered coding tools out of the box. The homepage and API reference advertise their machine-readable documentation (llms.txt, LLM-optimized reference, OpenAPI spec, ai-plugin manifest) through standard HTTP Link headers — so agents like Claude Code, Cursor, and ChatGPT can find the API contract on first visit without any extra setup on your side.
The API reference at /api-docs now speaks both HTML and Markdown from the same URL. Visit it in a browser and you get the rich HTML page; have an AI client fetch it with Accept: text/markdown and you get the LLM-optimized reference directly — now with the correct text/markdown content type so strict agents accept it on first try.
Added an API catalog at /.well-known/api-catalog (RFC 9727) — a single machine-readable index pointing to the OpenAPI spec, HTML and Markdown documentation, AI plugin manifest, and health endpoint. Agents that follow the well-known convention can now find everything they need in one request.
Published an Agent Skills index at /.well-known/agent-skills/index.json exposing three tool-worthy skills — generate-image, generate-video, and list-models — each backed by a markdown manifest with SHA-256 integrity. An agent runtime that speaks the Agent Skills Discovery RFC can now call PixelDojo image and video generation as first-class skills.
Added OAuth 2.0 Protected Resource Metadata at /.well-known/oauth-protected-resource (RFC 9728) with supported scopes and a pointer to the API key page. Agents can auto-discover how to get credentials for the PixelDojo API instead of needing bespoke setup docs.
Homepage structured data now identifies PixelDojo as a SoftwareApplication with a feature list, category, and offer details alongside the existing Organization and WebSite blocks. Cleaner ingestion by search engines and AI crawlers — and more accurate knowledge cards in answers about what PixelDojo actually is.
Declared our AI content preferences. The site robots.txt publishes Content Signals indicating the marketing site and documentation are open for AI training and inclusion in AI answers. User-generated gallery content and profile pages opt out of AI training via noai/noimageai — your posted creations are crawled for search and inspiration, not harvested for training.
Agents can now auto-discover how to sign in. Added OpenID Connect discovery at /.well-known/openid-configuration and the matching OAuth 2.0 authorization server metadata at /.well-known/oauth-authorization-server, mirrored from our Clerk-backed auth. Agents that speak either flavor can find the issuer, endpoints, and supported scopes without any bespoke config.
Markdown negotiation now works site-wide on every agent-facing surface — homepage, /agents, /api-platform, and /api-docs. Fetch any of them with Accept: text/markdown and you get the LLM-optimized reference; browsers keep getting the HTML pages. Moved the routing into middleware so it runs ahead of static-page caching, which the previous version got stuck behind.
Homepage Link headers are now served on every response, not just HTML ones. The original guard only emitted them when Accept: text/html was present, which meant agent crawlers using Accept: */* saw nothing. They now see the full RFC 8288 link graph (api-catalog, agent-skills, oauth-protected-resource, openid-configuration, service-doc, service-desc, and more) on any request to /.
Credit deductions are now fully transactional. Every debit and its ledger entry commit together as a single atomic unit, or not at all — so even if the database blinks mid-generation, your balance and our record of where it went stay perfectly in sync. Eliminates a whole class of "my credits got deducted but the job never ran and I never got a refund" edge cases.
API keys are now stored as one-way hashes, not the raw key. The full secret lives only in your hands — we cannot display it again after creation, and a database leak couldn't expose it either. Existing keys keep working seamlessly: they auto-upgrade to hashed storage the next time you use them. No rotation needed unless you've misplaced your copy.
Webhook HMAC signatures now use a dedicated signing secret separate from your API key. The signing secret is generated per-key and kept isolated — meaning webhook verification keeps working even if you rotate the API key, and vice versa.
Tool pages feel more like a professional studio and less like a consumer toy. Image and video output cards have subtler shadows and a lighter hover lift, so the art stays the hero and the chrome gets out of the way.
First-run delight: the first image, first video, and first Canvas chain step you generate in a session gets a small scale-in — then the animation goes quiet so it doesn't fire on every click.
This page (What's New) got an editorial redesign. Bigger display headlines, typewriter-style dates, magazine spacing. Designed to be read, not skimmed.
Per-second credit pricing on Seedance 2 Reference to Video and Seedance 2 Video Edit has been adjusted to reflect their higher compute costs. Seedance 2 Reference: Fast 5 cr/s at 480p and 9.5 at 720p; High 8 at 480p, 16 at 720p, 24 at 1080p. Seedance 2 Video Edit: Fast 6.5 cr/s flat; Standard 5 at 480p, 9.5 at 720p, 16 at 1080p. Plain Seedance 2 (text/image-to-video) pricing is unchanged.
1080p output is now available on Seedance 2, Seedance 2 Reference to Video, and Seedance 2 Video Edit — on the High / Standard quality tiers. Fast tiers stay at 480p / 720p.
Seedance 2 Video Edit brings back its Standard quality tier — higher fidelity for prompted video edits, especially at 720p and 1080p.
Pricing on all three Seedance 2 tools recalibrated to cleaner whole numbers. Seedance 2 / Reference: Fast 3 cr/s at 480p and 6 at 720p, High 5 at 480p, 10 at 720p, and 15 at 1080p. Seedance 2 Video Edit: Fast 4 cr/s flat; Standard 3 at 480p, 6 at 720p, 10 at 1080p.
Seedance 2 generations no longer hang on "queued" forever after the backbone migration — the status endpoint URL namespace was wrong in the first cut. New jobs now poll correctly.
Picking a video from "My Videos" in Seedance 2 Reference to Video no longer fails with "Failed to fetch" — the library picker was re-validating CDN-hosted videos and tripping browser CORS. Library videos skip the re-fetch now (they were already validated at upload time).
PixelDojo now respects your system-level "reduce motion" preference. If you have motion sensitivity settings turned on in your OS, decorative shimmer, pulse, and spin animations now pause, and transitions shorten. No setting to toggle in the app — your OS choice just works everywhere.
Seedance 2, Seedance 2 Reference to Video, and Seedance 2 Video Edit now run on a faster backbone — same prompts, same credit prices, faster and more reliable queuing. Expect fewer stalled jobs during peak traffic.
WAN 2.7 Video now accepts prompts up to 8,000 characters (previously 5,000). Long cinematic prompts and prompt-enhance expansions no longer trip the character cap.
Seedream 4 now retries transient ByteDance "service encountered an unexpected internal error" responses automatically with backoff. Generations that would previously fail after a provider hiccup now recover on their own.
Ideogram Character's resolution picker now lists only sizes the model actually accepts — the old options (512×512, 768×768, 1280×720, 1920×1080, 1024×768, 768×1024) were silently rejected by Replicate. Replaced with nine provider-valid resolutions covering square, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16, and ultra-wide/ultra-tall.
Veo 3.1 Standard no longer offers 1:1 as an aspect ratio — Google's API only accepts 16:9 and 9:16, so picking square was guaranteed to fail. The dropdown now matches what actually works.
Grok Image (xAI) now accepts prompts up to 5,000 characters (previously 2,000), matching P-Image, WAN 2.5, WAN 2.7 Video, and Seedance 2.
Canvas resolution picker for WAN 2.5, WAN 2.7 Video, and WAN Reference to Video now shows a proper 720p / 1080p chip selector instead of a free-text field — no more "Expected '720P' received '720p'" surprises. Stale sessions with lowercase values still submit successfully.
Prompt length limits unified across every generation tool. Modern image and video models accept 65K–131K tokens (hundreds of thousands of characters), so our per-tool caps of 1K–8K chars have been lifted to a single generous 250K-character ceiling. Detailed multilingual prompts, prompt-enhance expansions, and long cinematic descriptions no longer hit arbitrary limits on Ernie, Kling, Grok, Hidream, FLUX Edit, WAN, Seedance, or any other tool.
P-Image Edit now accepts prompts up to 5,000 characters (previously capped at 2,000), matching its sibling P-Image tool. Detailed edit instructions and longer Canvas chains no longer hit a surprise validation error.
WAN 2.2's public API schema now documents all LoRA parameters. Standard mode supports two LoRAs (lora_weights_transformer + lora_weights_transformer_2 with matching scale fields), and Extended mode supports full dual-LoRA with four slots — 1 and 3 target the high-noise transformer, 2 and 4 target the low-noise transformer. Scales accept 0–4.
Seedance 2 now accepts prompts up to 5,000 characters (previously capped at 2,000), matching the longer-prompt tools like WAN 2.5 and WAN 2.7.
The API Platform pricing page now shows real credit costs for 17 more models that were stuck displaying "0 credits" — Consistent Characters, FLUX, Flux Edit, Nano Banana Edit, Grok Imagine R2V, Ideogram Character, Kling Image V3, Kling Image Edit, Kling 2.6 Pro, Lip Sync, OmniHuman 1.5, Flux Redux, Seedance 1, Seedance 2 Reference, WAN 2.7 Image Edit, WAN Reference to Video, and WAN 2.7 Video Edit. Every variant now surfaces the correct per-request price.
When an image upload fails in an image-to-image tool or Canvas, you'll now see the real reason (like a session or subscription issue) instead of a generic "Failed to upload" message that made it look like the file itself was the problem.
Seedance 2 Reference to Video is live. Guide Seedance 2.0 with up to 9 reference images, 3 video clips, and 3 audio tracks — tag them in your prompt as [Image1], [Video1], [Audio1] to lock characters, motion, and sound in a single generation. Fast (3/5 credits/sec at 480p/720p) and High (3/6 credits/sec) tiers, native audio sync, 3–15 second clips.
Seedance 2 Reference in Canvas now shows proper Quality, Resolution, Duration, Aspect Ratio, and Generate Audio dropdowns instead of raw text inputs that could silently fail validation.
P-Image, WAN 2.5, and WAN 2.7 Video now accept longer prompts (up to 5,000 characters) — previously anything over 2,000 characters was rejected with a validation error.
GPT-Image 1.5, Qwen Image 2, and Ideogram Character now show proper dropdowns in Canvas More Settings — Size, Quality, Model, Speed, and Style controls replace raw text fields that sometimes caused silent failures.
Seedream 4 edit mode no longer rejects requests that pass a single reference URL instead of an array — the API now accepts both forms gracefully.
Pose Control (FLUX) works again — submissions without a LoRA were being rejected upstream because the empty strength value was sent as a string instead of a number.
Grok Imagine Video in Canvas now has proper Duration, Resolution, and Aspect Ratio dropdowns instead of raw text fields that silently failed validation.
Earlier validation on Veo 3.1, WAN 2.6 Image Edit, and Seedream edit. Veo 3.1 (standard) now catches 1:1 aspect ratio with an input image upfront and asks for 16:9 or 9:16. WAN 2.6 Image Edit caps reference images at 4 when selecting. Seedream 4 and Seedream 5 Lite edit cap reference images at 14 when selecting. All three avoid confusing post-submit errors.
Ernie — Baidu's multilingual text-to-image model is live. Write prompts in English, Chinese, or Japanese with built-in prompt expansion. HD is 1 credit, UHD is 3 credits, and the Turbo variant is 1 credit flat.
"Continue in Canvas" actually sends your image into Canvas now. If you'd ever used Canvas before, the button used to take you there empty-handed.
Seedream 4 and Seedream 5 Lite edit mode accept pasted or newly-uploaded images again — base64 uploads were timing out before reaching the backend.
Veo 3.1 Lite videos finish and display correctly. Jobs were completing on the provider but never got picked up on our side.
Pose Control (FLUX) with multiple LoRAs works now — was 0% success after the multi-LoRA update.
SeedEdit 3 is back online after a provider API change broke every request.
LTX 2 in Canvas now has proper Tier, Duration, Resolution, and FPS dropdowns instead of raw text inputs.
WAN Reference to Video in Canvas gets proper Duration slider plus Resolution, Aspect Ratio, Model, and Shot Type dropdowns — no more typing values that silently fail.
WAN 2.6 Image Edit in Canvas now has Size, Prompt Enhancement, and Negative Prompt as proper form controls. Previously free-text fields were tripping validation.
Magnific Upscaler now catches oversized images upfront with a clear "Image is too large to upscale (max 10 MB)" message instead of a cryptic upstream error.
Nano Banana edits with "Match input image" aspect ratio no longer 400 on the upstream API — always sends explicit width and height now.
Seedance 2 image-to-video no longer complains that you need an image after you already uploaded one — mode and image state stay in sync.
Omnihuman's Fearful, Disgusted, and Surprised emotions produce audio now (were sending wrong values to the voice engine).
Flux Edit's guidance slider in Canvas now stays inside the range the model accepts — no more mid-drag validation errors.
Empty-prompt submissions across 30+ tools now show a clean error. Advanced JSON mode, Enter-key submits, and rapid clicks all get caught client-side instead of hitting the API.
Content-moderation rejections across the Seedream and Seedance families are correctly labeled as content issues instead of provider outages — cleaner signal when something is actually broken.
The /api/v1/models listing and API Platform docs pages show real credit costs for dynamically-priced models instead of "0 credits".
Your Usage report loads noticeably faster — added indexes that make per-user lookups across credits and refunds much quicker, especially for longer time ranges.
Canvas's model picker now shows every model tier as its own option — pick Nano Banana Standard, Pro, or 2 directly instead of hunting through settings.
Multi-image editing in Canvas — models like Nano Banana Edit, P-Image Edit, and WAN 2.6 Image Edit accept up to 5 reference images. Click '+ Add reference' next to your source to attach more.
Canvas's More Settings panel is way friendlier. Aspect ratio and resolution are dropdowns. Image slots have a proper picker. Fields you've already filled from a canvas step get hidden so you're not duplicating work.
When you pick an image source for a video tool, mode auto-flips to image-to-video and the image slot is pre-filled. No more empty fields asking what to do.
P-Image Edit is now available in both the public API and Canvas — fast image editing with up to 5 reference images.
Fixed the bug that silently sent edits to the wrong endpoint — Canvas edits now land on the correct model every time.
Admin Tool Health dashboard — per-tool success rate, error breakdown, and latency for the last 7 days.
Light mode contrast fixes on the Canvas source indicator and temporary-output warning — readable in both themes now.
WAN 2.5 Image-to-Video now disables Generate until you've uploaded an image — no more silent failures.
WAN 2.7 Video Extend mode waits for your source clip's duration to finish processing before letting you Generate.
Nano Banana Pro and Nano Banana 2 work correctly via the public API again — requests now route to the right backend.
Image editing tools added to WAN 2.7 — paint over areas and regenerate with AI.
Usage history now shows which tool and provider were used for each generation.
All API error messages are now clearer and more consistent across every endpoint.
Strengthened security across the platform — all endpoints now require proper authentication.
Homepage loads 40% faster. Font loading improved by 97%.
Removed 90+ unused files and 48 unnecessary dependencies for a leaner, faster app.
Credits are now deducted upfront for LoRA training jobs — no more surprise charges.
Automatic refunds are more reliable — failed jobs always return your credits.