How to Write a Seedance Prompt So the Face Doesn't Drift During a Look Transformation

The most common problem with look-transformation videos (the clothes change, the background changes, the person stays) isn't how good the result looks. It's that by the 5th second the face stops resembling the reference. A video generation model isn't obliged by default to hold a person's identity through the whole scene unless the prompt makes it do so explicitly.

A working Seedance prompt that spread through one niche AI channel in October 2026 solves this problem with four techniques. Let's go through them in order, without tying them to the specific plot of the example, so you can carry the structure over to your own scene.

Technique 1: the reference as an identity anchor, not as inspiration

The first line of the prompt doesn't describe the scene. It fixes a rule: the face, age, body proportions, skin texture, hair, and recognizable appearance of the person stay exactly as in the reference photo for the whole video. This is not a one-time mention of "looks like the photo," but an explicit instruction to keep the appearance unchanged even while everything else changes: clothes, surroundings, light.

In practice this means that at the start of the prompt, a separate block spells out exactly what must not be touched. Not the face in general, but specifically: face shape, skin color and texture, hairstyle, expression. The more specific the list, the less room the model has to drift in the middle of the clip.

Technique 2: the surroundings are described as physically real

The second part of the structure sets rules for the background and objects in the scene. If there is a forest in the frame, the trees must not bend, dissolve, or turn into something fantastical. If there is an interior, objects must behave like ordinary objects: no levitating, no changing shape for no reason.

This looks like extra caution, but this is where the model most often gets caught distorting: branches start moving in sync with the clothing transformation, fog fills half the frame, the background "plays along" with the magic instead of staying neutral. An explicit ban on animating the background solves this at the prompt level, before generation, not at the level of regenerating again and again.

Technique 3: the scene is cut by time, not described as a whole

Instead of one continuous description, the prompt is divided into time segments: 0 to 4 seconds, 4 to 10 seconds, 10 to 13 seconds. Each segment gets its own task: in the first, nothing changes, there is only a camera push-in and a setup of the mood. In the second, the transformation itself happens, with an explicit statement of what changes (only the clothes) and what does not (body, face, pose). The third shows the final result, with no transformation movement anymore, only the character's reaction.

This technique solves a problem that often spoils outfit-change or look-change videos: the model tries to show everything at once, and the transformation ends up looking choppy rather than smooth. Splitting by seconds with a concrete task for each segment brings back control over the pace of the scene.

Technique 4: the ban sits both next to the risk and in a separate summary list

The fourth technique is easy to mistake for an "either this or that" choice, but in a real prompt two levels of protection work at once. The first level: pinpoint bans are built right into the description of a specific risk. Right after the description of the forest comes "do not bend, dissolve, move, or stylize the trees." The clothing transformation has the same pinpoint ban: "do not fill the surroundings with smoke, do not cover the forest with smoke, do not distort the body, do not change the face and hair."

The second level: a separate summary block at the very end of the prompt, with a heading listing forbidden deviations, more than twenty items. Among them: no artificial trees, no moving trees, no deformation of the tree trunk, no cheap Halloween costume, no witch costume, no cosplay look, no distortion of the face, no deformation of the body, no beautifying, no sharp cuts, no teleportation, no text, subtitles, logos, or watermarks in the frame.

A pinpoint ban next to the risk catches what is most likely to go wrong in exactly that place in the scene. The final summary list catches more general and borderline cases that aren't tied to one specific moment: stray text in the frame, an abrupt editing jump, the general feel of a cheap costume instead of a quality transformation. It's more reliable to keep both levels at once than to choose one instead of the other.

How to build such a prompt for your own scene

If you carry this structure over to your own task, the order is this. First, fix identity from the reference, then the rules for the surroundings: what is realistic, what must not be touched. Next, split the scene into 3 or 4 time segments with one task for each, and a pinpoint ban inside each segment for the specific risk of that scene. At the very end, add a separate summary list of general bans for cases that aren't tied to one moment.

The more precisely a ban is worded, the fewer generations you spend redoing. General phrases like "make it beautiful but realistic" almost always give a less predictable result than a concrete description of what must stay unchanged.

I write about how I build working prompts and AI combinations for video in the channel «Растём с ИИ» ("Growing with AI"). If you need to set up such a system for your task, you can start with a free consultation.