Seedream 4.5 is ByteDance's latest text-to-image model — photoreal output with strong aesthetic taste, accurate text rendering, and native 2K / 4K presets. Flat 1 credit per image, image-to-image supported, custom dimensions up to 4096px on either side.
Seedream 4.5 is the latest in ByteDance's Seedream image family. Where it stands out: aesthetic taste (the model picks lighting, color, and composition choices that read as designer-led, not template-driven), accurate text rendering, and native 2K + 4K presets that don't require a separate upscale step.
The model is synchronous — predictions complete in a few seconds rather than minutes. Custom dimensions go up to 4096px on either side. Image-to-image is supported for guided generation. Pricing is flat 1 credit per image regardless of size.
2K · 4K
Native resolution presets
1
Credit per image (flat)
4096px
Max custom dimension
Seedream 4.5 picks lighting, palette, and composition like a working photographer — natural depth-of-field, balanced color temperature, intentional negative space. You don't have to over-prompt the aesthetic.
In-image text comes out cleanly — multi-word, multi-line, and accurate to the exact string you specified. Wrap text in double quotes and Seedream preserves spelling and spacing reliably.
Pick auto_2K (default) or auto_4K and Seedream renders at that resolution natively — no separate upscale step that costs credits or smudges fine detail. Custom dimensions accept up to 4096px on either side for free-form aspect work.
Even a 10-word prompt produces aesthetically pleasing output — Seedream fills in lighting, palette, and composition decisions with taste. Compare to models that need 80-word prompts to look intentional.
Each example shows the exact prompt that produced the result. Copy any prompt with one click.
2K · landscape_16_9 · 1 credit
Cinematic portrait of an elderly Japanese fisherman mending nets on a sun-bleached wooden dock at golden hour, gnarled hands threading frayed twine, weathered skin, amber sunlight catching salt spray in the air, hyper-realistic 8k
For portraits: subject + action + specific sensory detail (skin texture, fabric, light quality) + format anchor ("hyper-realistic 8k"). Seedream's aesthetic-default kicks in here — you don't need to prompt the color palette explicitly.
2K · square_1_1 · 1 credit
Studio product shot of a matte ceramic vase with a single white peony, polished concrete tabletop, soft top-down rim light against a charcoal gradient backdrop, photoreal, advertising quality, no text or labels
Square 1:1 is Seedream's sweet spot for product. Specify the surface, the lighting setup, and add "no text or labels" to suppress unwanted typography on the product. 2K is enough for most social uses; bump to 4K only for print.
4K · landscape_21_9 · 1 credit
Wide cinematic landscape of a misty fjord at dawn, mountain peaks emerging through low fog, a single distant boat leaving a thin wake on glassy water, cool blue-green palette with a warm amber glow on one peak, hyper-detailed 8k
21:9 ultra-wide rewards atmospheric scenes. 4K resolution preserves the fog, water surface, and distant detail that compress to mush at 2K. Pair with a "warm spot" element to break the cool palette.
2K · landscape_16_9 · 1 credit
Long-shot of a woman in a red trench coat walking across a vast cracked salt flat at sunrise, silhouette razor-thin, casting a long shadow across the white earth, distant violet mountains, warm amber backlight, photoreal 8k
Seedream handles characters-at-distance with strong scale — small in frame, silhouetted, dramatic light. Pair "long shot" or "wide establishing shot" with environmental scale ("vast salt flat", "endless desert") to push the character into the landscape.
2K · portrait_3_4 · 1 credit
Vintage travel poster reading "TOKYO 1962" in bold typography at the top, smaller italic subtitle "neon and rain", central illustration of a neon-lit rainy alley, vintage poster color palette (deep indigo, hot pink, warm cream), art-deco border, the text must be exact
For posters: headline in quotes + subtitle in quotes + central illustration described visually + named palette + "art-deco border" or similar style anchor + "the text must be exact". Seedream renders text more reliably than most general-purpose models.
Seedream picks lighting, palette, and composition with taste even when you don't specify. Don't over-prompt — a 20-word prompt with strong subject + scene + format anchor often beats a 100-word one. Specifying every detail can actually limit the model's taste.
For in-image text, wrap the exact string in double quotes and add "the text must be exact" at the end. Seedream is one of the more reliable models for multi-word text, but explicit framing locks the rendering tighter.
2K (default) for social and web. 4K when you need print fidelity or atmospheric scenes with lots of fine detail. Custom width/height (up to 4096px) for free-form aspects not in the preset list.
auto_2K, auto_4K (model picks the aspect), and the named presets (square_1_1, landscape_16_9, portrait_9_16, etc.) are simpler than computing custom width/height. Use custom only when you need a non-standard aspect.
"Weathered skin", "salt spray in the air", "frayed twine" — concrete sensory anchors give Seedream something specific to render. Generic adjectives ("nice", "beautiful") don't guide the model.
Pass an input image with your prompt to use it as a guidance reference. Seedream keeps the composition and color hint of the source while applying your prompted changes. Useful for moodboarding from an existing shot.
| Setting | Values | Notes |
|---|---|---|
| Image size | auto_2K · auto_4K · square_1_1 · landscape_16_9 · portrait_9_16 · landscape_4_3 · portrait_3_4 · landscape_3_2 · portrait_2_3 · landscape_21_9 | Or custom width/height up to 4096px on either side. |
| Custom dimensions | 1296–4096 px (width and height) | Use when no preset matches your target aspect. |
| Outputs | 1–4 | Each output is 1 credit. Iterate at 1 first. |
| Output format | PNG (default) · JPG · WEBP | PNG for transparency, JPG for smallest file size. |
| Seed | 0 – 2,147,483,647 | Same seed + prompt + size reproduces the image. |
| Image-to-image | Pass image_url field | Uses input as guidance reference; composition and palette carry through. |
Aesthetic taste from sparse prompts. Seedream picks lighting, color, and composition like a working photographer even with a 20-word prompt. FLUX is more literal, GPT Image 2 has stronger text rendering. Pick Seedream when you want intentional-looking output without over-prompting.