What People Recommend for Seedream 5.0 Pro: Prompts and Fixes

1. Write the prompt in layers, not as a keyword list

The most dissected example of the first weeks was a night street portrait with a prompt of about 170 words. It is long but not messy: six layers in a fixed order. Face and hair first, then a single piece of clothing, then the place and mood, framing and camera position, the "physics" of the shot (noise, light, blur), and finally exclusions with the format. People in the threads agree that the model reasons over the instruction, so connected sentences work better than a tag dump.

How to do it here. In Seedream 5.0 Pro, write in the same order and pick the format with the switch, not with words in the prompt:

A woman in her mid-twenties with dark wavy hair, in a plain white T-shirt. She stands by a pedestrian crossing on a quiet street at night, calmly looking into the camera. Behind her: warm streetlights, trees and parked scooters. Waist-up framing at eye level, as if a friend took it on a phone. Smartphone night mode, natural skin texture, light sensor noise, slight motion softness. No studio lighting, no beauty filters

2. Against plastic skin, ask for imperfections

The most common complaint in the launch thread on Reddit: people in Pro look less alive than in the previous version, with skin that is too smooth. The advice repeated by authors of successful portraits is not to ask for perfect sharpness but to describe the flaws of a real camera. Sensor noise, a handheld shot, slight defocus and a ban on beauty filters do more than any "ultra realistic".

How to do it here. End your Seedream 5.0 Pro prompt with a line like "natural pores and skin unevenness, light noise, handheld phone shot, no retouching". For editing a finished portrait from one photo, also try Nano Banana Pro: it is built for precise photo edits from one or two references.

3. Exclusions go inside the prompt, and only when needed

The model has no separate negative-prompt field. Two working techniques show up in forums: an "avoid: …" line at the end of the prompt with a list of concrete items, and short "no X, no Y" phrases inside the text. Experienced users add one more tip: don't start with a long list of bans. Run a clean prompt first, then add exclusions only for failures that actually appeared. A short list keeps the model focused on what you want.

How to do it here. Keep exclusions concrete: "no logos, no watermarks, no extra accessories" instead of "nothing bad". Where you can, replace a ban with a target: "hands resting calmly on her knees" works better than "no deformed hands".

4. Text in images: short, quoted, with a position

The developers themselves admitted at launch that small text still needs work, and early testers confirm it. Short headlines come out well; long paragraphs fall apart. People who get clean posters put every line in quotes and give it a place: "top center", "small, bottom right".

How to do it here. In Seedream 5.0 Pro, keep lettering to one or two short lines. If the poster has a lot of text, use FLUX 3 or GPT Image 2, where text is a strength. More on lettering in our FLUX 3 guide.

5. Dynamic scenes: tone down the action words

One of the first tests by an illustrator showed the flip side of the model's obedience. The stronger the action in the prompt, the more often the model overshoots: the character throws the katana instead of holding it. The advice from that discussion: if a pose breaks, don't add description, soften the action words and keep one action per character.

How to do it here. Write "raises the sword, gripping the hilt firmly" instead of "slashes furiously, everything flies". For real action in motion, make the frame in Seedream and animate it in Kling 3.0.

6. Full-body shots look pasted on: one light for everything

An early-test review noticed another weakness: in full-body portraits with detailed clothing and a busy background, the person sometimes looks cut out and pasted in. Edges are too clean, sharpness differs from the background, shadows don't match the surroundings. In close-ups this barely happens.

How to do it here. Describe one light source for the whole scene ("soft sunset light from the left, long shadows on the asphalt"), ask for clothing that responds to the pose, and don't load many different references at once. If full body matters, start with a knee-up shot and widen from there.

7. One edit per run, drafts at 1K

Veterans of editing models recommend a simple rule: change one thing per run. Five edits at once make it impossible to tell what broke or fixed the image. Another common finding: "4K" and "8K" in the prompt change the style, not the file size.

How to do it here. Tune the prompt at 1K, changing one layer at a time, then render the final at 2K. For reference edits, say what to take from each file by its upload number: "the character from image 1 in the interior from image 2, style as in image 3". If the result needs heavy enlargement, there is a separate Topaz upscaler.