GPT Image 2 Character Sheet Before Seedance

If your Seedance clips keep changing the face, hairline, clothes, or prop proportions, the problem usually starts before motion. A video prompt has to solve identity, staging, and movement together. That is why the practical workflow is to start at the Lem Gen prompt library, inspect the GPT Image prompt hub, and then build the base sheet in the image-editing Workspace with GPT Image 2 selected before you ever ask Seedance to animate anything.

When that sheet is ready, you can carry it into the Seedance prompt hub and the image-to-video Workspace with Seedance 2.5 selected. The point is not to make one magical master prompt. The point is to produce a reusable identity package that lets the video prompt focus on camera, timing, and action instead of re-inventing the character every time.

This guide shows how to turn GPT Image 2 into a practical pre-production step for Seedance: what the sheet must include, how to phrase the prompts, what to keep constant, which failures matter most, and how to connect the result back to repeatable motion prompts.

What a “character sheet before animation” really solves

A GPT Image 2 character sheet is not just a nice concept board. In a working video pipeline it acts as the identity contract for later shots. It records the details that should survive camera changes, motion, and story beats:

• face shape and eye spacing;
• hairstyle, hairline, and color anchors;
• outfit geometry and repeated accessories;
• the prop or tool that defines the character;
• the visual style you want repeated later;
• the clean silhouette that still reads when the camera moves.

That is different from the broader Character Consistency Prompt Workflow. That article explains how to protect identity across still-image revisions. Here the reader job is narrower and more urgent: build a sheet that survives the jump from still image to motion.

The distinction matters because a shot can fail even when the character still “sort of” looks right. In motion work, smaller mistakes become expensive:

• an apron becomes a different costume halfway through the clip;
• a mascot prop changes shape when the camera orbits;
• the face stays close enough, but the silhouette no longer matches the earlier scene;
• the first frame is usable, but the last frame cannot connect to the next shot.

That is why the pre-video sheet should be treated as a production artifact, not a mood board.

Build the sheet around shot needs, not around pretty poses

The first mistake is to ask GPT Image 2 for a “cool character sheet” with no production purpose. That often produces a visually attractive board that is weak as an animation input because it does not organize the identity cues a later video prompt needs.

Start with the shot family you expect to make. For short Seedance sequences, the useful questions are simple:

• Do you need a hero portrait, a torso crop, or a full-body silhouette?
• Will the character hold a stable prop?
• Does the scene require an expression change?
• Will the camera move around the subject or mostly push in?
• Is the motion built around a hand action, a body turn, or an environmental reveal?

Those answers tell you which sheet views matter. A social ad character might need:

• a clean front view;
• a three-quarter hero view;
• one expression set;
• one prop pose;
• one neutral body stance that can survive a cut.

A fantasy short or story trailer may need more environment notes, but the logic is the same. Build the sheet for downstream shots, not for gallery drama.

Start with one anchor prompt, then expand

The cleanest workflow is to use one anchor image or one tightly written anchor prompt first, then ask GPT Image 2 to expand that into a sheet. If your starting point is a reference image, the Image to Prompt workflow is still the right first move because it separates visible facts from guesses before you plan the animation pack.

The anchor prompt should be compact and factual. Avoid screenplay language at this stage. Your job is to lock the subject before you add action.

Create one production-ready character anchor.
Subject: young pastry chef mascot with round face, short dark bob haircut,
warm tan skin, white apron, red neck scarf, compact body proportions,
wooden spoon always visible.
Style: soft stylized 3D family-film look.
Background: neutral studio backdrop.
Goal: preserve facial structure, apron silhouette, scarf color, and spoon shape.

This is enough to establish the base. Once the identity is stable, expand the request:

Turn the approved anchor into a clean character sheet for animation planning.
Include: front view, 3/4 hero view, side view, one neutral standing pose,
one action pose with spoon raised, and three expressions
(focused, surprised, smiling).
Keep the same face shape, scarf, apron, and spoon geometry in every panel.
Neutral studio background. Clear spacing between panels.

The specific panel count is less important than the discipline: build from one approved identity rather than improvising six disconnected variants.

What the sheet should contain before it goes into Seedance

A character sheet prompt for Seedance should capture the details that motion prompts tend to lose first.

For most creator workflows, the minimum useful pack is:

1. one clean hero frame;
2. one side or three-quarter view;
3. one full-body or readable torso pose;
4. one expression variant;
5. one stable prop pose if the prop matters later.

If the later video is story-driven, add a timing-oriented state such as “calm before action” and “reaction frame after action.” Those extra states help the video model read the emotional direction of the shot instead of inventing it.

The sheet should also reveal what must not change. I usually phrase that directly:

Do not change the face proportions.
Do not replace the apron with another costume.
Do not add extra props.
Do not redesign the spoon.
Keep the same stylized 3D material language across every panel.

This kind of exclusion block is not glamorous, but it is practical. It tells GPT Image 2 what the pack is supposed to protect. The later Seedance prompt can then reuse those same invariants.

Separate the sheet prompt from the motion prompt

Once the sheet is approved, stop editing it unless identity is still wrong. Do not turn the sheet prompt into the motion prompt by simply appending camera instructions. That is where many workflows get muddy.

The sheet prompt answers:

• who is this character?
• what does the subject look like from the views we need?
• what props, outfit anchors, and expressions must survive?

The motion prompt answers:

• what happens in this shot?
• how does the camera move?
• what begins the clip?
• what resolves the clip?
• what is allowed to change?

Keep those responsibilities separate. The motion prompt should reference the approved sheet and then stay focused on shot behavior.

Use the approved character sheet as the identity source.
4-second image-to-video shot.
The pastry chef mascot stands at a wooden prep table,
raises the spoon once, leans forward, smiles, and settles into a hero pose.
Camera: slow push-in.
Keep face, apron, scarf, and spoon shape consistent.
No extra hands, no extra props, no costume changes, no sudden body stretching.
End on a clean centered hero frame.

That is already stronger than a giant paragraph that mixes story, style, motion, identity, and mood in no obvious order.

Plan one small shot before you plan the sequence

The fastest way to waste a good character sheet is to test it in a long, complicated scene first. A short controlled shot gives you better information.

For the first Seedance pass, use a clip that tests only one motion idea:

• a slow push-in;
• a single turn to camera;
• one hand action;
• a calm reaction shot;
• a simple walk-and-stop.

Why start small? Because the first test is diagnostic. If the result drifts, you need to know whether the failure came from the sheet or from the shot request.

If you ask for:

• a fast dolly move,
• a wardrobe change,
• a prop interaction,
• a location reveal,
• a dramatic expression shift,
• and a final hero frame

all in one first test, you will not know which demand broke the clip.

That is the exact problem described again and again in current creator discussions: identity often degrades when too many variables change at once, while reference-first workflows hold up better when the action is simpler and the camera direction is explicit.

Use the sheet to control the first and last frames

One of the most useful side effects of a good sheet is that it makes the first and last frames more intentional. For creators making short ads, reels, or cut-based story clips, that matters more than “general consistency.”

The first frame should inherit a stable identity state from the sheet. The last frame should be a deliberate landing point, not an accidental mush of motion blur and model improvisation.

That means your shot plan should state the frame logic directly:

Start state: neutral stance from the approved hero panel, spoon held low.
Middle action: one clear spoon raise with head turn.
End state: spoon settles at shoulder height, smile visible, body centered.

When the sheet and the motion plan agree, the clip is easier to review. When they disagree, the model has to guess which part matters more.

That is also why the older Seedance Video Prompt Patterns article still matters here. Seedance prompts work best when they behave like shot descriptions, not like adjective piles. The character sheet gives the shot a stable actor; the Seedance prompt gives that actor a readable beat.

Diagnose the drift by failure type

Once you run the first animation test, do not ask “is it good?” Ask what failed first? That tells you what to revise.

The most common failure types are:

1. Face drift

Usually the sheet did not define the face clearly enough, or the shot asked for too much head rotation too early.

Fix:

• strengthen the approved hero and three-quarter views;
• simplify the expression change;
• reduce fast motion until the face survives.

2. Outfit or prop drift

Usually the sheet under-specified the stable costume or object geometry.

Fix:

• add one clearer prop pose to the sheet;
• repeat the prop name and material in the motion prompt;
• explicitly forbid replacement props.

3. Body proportion drift

Usually the shot asks for too much action without enough full-body reference.

Fix:

• add one neutral full-body sheet panel;
• shorten the shot;
• remove unnecessary environmental movement.

4. Style drift

Usually the sheet and the motion prompt describe style differently.

Fix:

• use one stable style phrase in both prompts;
• stop adding extra visual metaphors in the video request;
• keep lighting and material language consistent.

5. End-frame collapse

Usually the shot had no explicit ending state.

Fix:

• define the last pose;
• end on a simple hero frame;
• reduce the amount of change in the final second.

A reusable prompt stack for Lem Gen

If you want a working loop inside Lem Gen, keep the stack simple:

• use the GPT Image prompt hub to inspect prompt patterns for identity and visual direction;
• use the portrait photography prompt hub when the subject needs clearer facial or pose language;
• build or tighten the sheet in the GPT Image 2 image-editing Workspace;
• then carry the approved pack into the Seedance image-to-video Workspace.

The key is that the “sheet” and the “shot” remain two separate assets:

• the sheet is the identity package;
• the shot prompt is the motion instruction;
• the review notes explain why the latest version was approved.

That review layer matters more than many people expect. A prompt without the review notes is hard to trust later because you no longer know which details were deliberate and which were accidents.

When to rebuild the sheet instead of forcing the video prompt

Sometimes the right answer is not to keep tightening Seedance. Rebuild the sheet when:

• the approved stills already disagree on the face or silhouette;
• the prop design changes across sheet panels;
• the outfit is too busy to survive motion;
• the style language is inconsistent across panels;
• the subject was never shown from the angle the shot requires.

In those cases, the video model is not the first problem. The pre-production pack is.

This is where many creators lose time. They keep editing the motion prompt even though the reference pack itself is unstable. If the identity source is noisy, the video prompt becomes a blame sink for failures it did not cause.

A short example workflow from still to motion

Here is the practical loop I would use for a short branded character clip:

1. Build one anchor in GPT Image 2.
2. Expand it into a sheet with front, three-quarter, side, and one prop pose.
3. Approve the sheet based on face, silhouette, costume, and prop geometry.
4. Write one 4-second Seedance shot with one action and one camera move.
5. Review only the first visible failure.
6. Either tighten the sheet or simplify the shot.
7. Save the approved pair: sheet + motion prompt + review note.

You can watch the same logic in motion below: the useful part is not that the clip moves, but that the subject remains readable while the movement stays limited and intentional.

The practical lesson is simple: motion quality improves when identity work happens first.

The working prompt templates

Use these as a starting point, then adapt them to your subject.

Template 1: character anchor

Create one clean identity anchor for later animation.
Subject: [who the character is]
Stable facial traits: [face shape, eyes, hair, skin, signature details]
Stable outfit anchors: [repeatable clothing pieces]
Stable prop: [object that must remain recognizable]
Style: [one visual style phrase]
Background: neutral studio setup
Goal: preserve identity for later character-sheet expansion

Template 2: sheet expansion

Convert the approved anchor into a production character sheet.
Include: front, 3/4, side, neutral standing pose, prop pose,
and three expressions.
Keep the same face, silhouette, outfit anchors, and prop geometry.
No background clutter. Clean spacing. Same style in every panel.

Template 3: first Seedance shot

Use the approved character sheet as the identity reference.
4-second image-to-video shot.
Action: [one simple action]
Camera: [one simple move]
Keep face, outfit anchors, and prop shape consistent.
No extra characters. No sudden costume changes. No new props.
End on a clean hero frame.

Template 4: repair note for drift

The clip drifted in [face / prop / outfit / silhouette / ending frame].
Revise only that failure.
Keep every other approved element from the previous test unchanged.

Template 5: approval record

Approved sheet version: [name]
Approved shot version: [name]
Why it passed: [specific identity and motion reasons]
Next shot should preserve: [list]
Next shot may change: [list]

Final checklist before you animate

Before you leave GPT Image 2 and switch to Seedance, verify this:

• one approved anchor exists;
• the sheet includes the views the shot actually needs;
• the face, costume, and prop are stable across sheet panels;
• the style phrase is consistent across still and motion prompts;
• the first video test uses one small motion idea;
• the ending frame is explicit;
• the review note explains what passed.

That checklist is more valuable than another paragraph of adjectives.

The shortest useful summary is this: make the character sheet first, make the shot second, and let each asset solve one job. Start from the Lem Gen homepage, compare structures in the GPT Image prompt hub and Seedance prompt hub, then build the identity pack in the GPT Image 2 image-editing Workspace before moving into the Seedance 2.5 image-to-video Workspace. If you need a broader still-image identity method, keep the character consistency workflow nearby; if you need tighter shot wording, pair this guide with Seedance video prompt patterns.