EN 
Русский 
Español 
Français 
العربية 
Türkçe 
فارسی Seedream is ByteDance’s image model, and in the bot it is the one to pick when the look matters more than the literal reading. Name a medium, an era or a mood and it commits to it, instead of softening it toward something safe. That is also what you manage, because it commits just as hard to whatever you left vague, and because anatomy is where it gives way first.
The Seedream prompt formula
Six slots cover almost every prompt: subject, action, setting, light and lens, framing, style. The order matches the other models in the bot; the weighting does not, because the last slot moves a Seedream render further than anything else here.
A lone traveler in a heavy patched coat standing at the edge of a cliff road, one hand shading their eyes as they look out over a valley of floating stone islands, low gold hour light raking across the rock from behind, ultrawide 21:9 with the figure small on the left third, painterly matte illustration with visible brushwork in the manner of a classic adventure book cover
Subject, action and setting work as anywhere else, with one caveat: Seedream bends a loosely named setting toward the style you asked for, so anchor it with concrete nouns (a cliff road, stone islands).
Light and lens is read as mood first and physics second: “raking gold hour light from behind” lands more reliably than a lighting setup in equipment terms.
Framing matters more here because of the aspect ratios. At 21:9 there is a lot of width to fill, and width you do not assign gets filled with invented scenery. Say where the subject sits and what occupies the rest.
Style is the slot that pays. A named medium, era or technique moves the image: oil on canvas with visible brushwork, sun bleached 35mm film from the seventies, ink and wash, airbrushed science fiction paperback.
The under-specified prompt also fails differently here. Seedream does not hedge toward a safe average; it picks an interpretation and renders it with conviction, including for everything you never mentioned. Hence the three word prompt that returns a confident image of somebody else’s idea.
Attaching photos: @image1 and @image2
Up to 4 photos go in one message, with the instruction as the caption of that same message. Photos first and text after are two messages and will not be paired; the step-by-step guide walks through it with screenshots. Then address them by number.
Restyle @image1 as an oil painting with visible brushwork, keep the pose and the position of both hands exactly as in @image1, warm lamp light from the right, waist up
That combination is where Seedream is unusually good: change how a picture looks while keeping what is structurally in it. The structural instruction survives the restyle, so the reference holds the pose after the medium, the palette and the light have changed. With more than one photo attached, label each one (pose, palette, setting); left unlabeled, the model picks one interpretation for the whole set instead of asking.
Where Seedream is strong, and where it is not
| Strong at | Struggles with |
|---|---|
| Stylized, painterly and imaginative work | Hands, fingers, feet and unusual poses |
| Ultrawide 21:9, the only model in the bot that renders it | Crowded frames with several figures |
| Restyling a photo while keeping its structure | Exact or lengthy text inside the image |
| Art direction by medium, era and technique | Cheap iteration, since 2K at 2 credits is its floor |
For words inside the picture that must be exactly right, GPT-Image-2 is the specialist; for photorealism on a busy frame, NanoBanana Pro.
Fixing hands, feet and poses
Attach a reference and point at it. Photograph the pose yourself, or use any photo containing the hands you want, send it with the prompt and address it: hands posed exactly as in @image2, same grip, same finger positions. Pointing at a reference photo is the most effective anatomy fix there is: a referenced pose gets copied rather than invented, and inventing is the step that goes wrong.
Describe limb placement explicitly. Say where each hand is and what it is touching. “Holding a cup” is a request the model can satisfy with one hand and leave the other out of the picture, and it often does.
Keep hands inside the frame. Hands cropped at the border come back wrong more often than hands fully in view.
Render at 3K or 4K rather than the 2K floor. More pixels means more detail budget for small structures, and fingers are what a tight budget spends less on.
Do not crowd the frame. Crowded frames are where hands go wrong first. Render one or two people well, then add the rest as an edit.
The two pictures further down the page are the same scene run twice. The first came from “a barista holding a coffee cup”, and the model routed around the hard part: one hand and its arm hold the cup, the other is simply not in the picture, because nothing asked for it. The second names the grip and the framing.
A barista in a dark green apron over a white t-shirt holding a speckled ceramic cup out toward the camera with both hands, all ten fingers visible and wrapped around the sides of the cup, both wrists inside the frame, blurred espresso machine and cafe counter behind, soft daylight, square close-up
A looser filter, and where the line is
Seedream’s content filter is looser than GPT-Image-2’s and looser than NanoBanana’s. That is a real product difference, not a slogan: darker, stranger and more provocative ideas run here that the others refuse or flatten into something tamer, which for imaginative work is the point of the model.
Looser is not unrestricted, which is the honest answer when people ask how uncensored Seedream really is. It still refuses whole categories outright, and its filter is only the innermost layer: the bot’s rules and Telegram’s terms sit on top of it.
Use it responsibly
The extra room is for imagination, not for harming anyone. The rules are short and not negotiable.
- Do not generate images of real, identifiable people without their consent.
- Never intimate, sexual or defamatory imagery of a real person.
- Never anything sexual involving minors, in any style, under any framing.
- Do not pass a generated image off as a real photograph of a real event or a real person.
- Follow the law where you live, for images you make and images you share.
The bot’s rules and Telegram’s terms apply to every render, and accounts get banned for abuse.
Common fixes
| Symptom | Change to the prompt |
|---|---|
| Fingers are fused, extra or malformed | Attach a reference and write hands posed as in @image2, then re-render at 3K or 4K |
| A hand is cropped out of the frame | Ask for both hands whole and inside the border, and say what each holds |
| The pose keeps getting ignored | Give each limb its own clause, and attach a photo showing the pose rather than describing it twice |
| The composition falls apart with several figures | Cut the cast down to one or two, render that, then add the others as an edit |
Ready-made prompts
This Seedream prompt guide ends where the model starts teaching you: change one slot at a time on a prompt that already works. The Seedream prompt library has ready-to-paste prompts next to the images they produced; the model page lists resolutions, ratios and credit costs.
Open the bot, pick Seedream, and paste one in.