---
name: character
description: "Router for the person's own recurring AI character in stills — make the character sheet from their photo(s), keep the character the same across images and scenes (references with one job each, camera, light, mixing), put the character into a film frame (head / face / whole-character swap). Source experiment pick: GPT Image 2.5 edit at quality low, picture roles only, one pass. Classifies the job and hands off to character-* sub-skills with exact settings, prompts and times. Use when the user says character sheet, turnaround, pose sheet, expression sheet, consistent character, same face in every image, my character in this scene, replace the actor, face swap, head swap, лист персонажа, персонаж из фото, консистентный персонаж, тот же персонаж, вставить персонажа в кадр, заменить актёра, замена лица."
---

These creative references describe source experiments with external tools. Use your own supported account and tools; this download installs no automation. Settings and timings are experimental observations, not current FlashBoards availability.

# character · meta-skill (router)

**Work for the person's own character: their photo / sheet from their character card. The example runs belong to a demo character.**

Measured in source experiments (28.09–09.10.2026). Articles with the prompts and results:
- https://flashboards.yaroflasher.com/learn/acting/ai-character-sheet/
- https://flashboards.yaroflasher.com/learn/acting/consistent-ai-character/
- https://flashboards.yaroflasher.com/learn/acting/replace-character-in-film/

Slots used in all character-* skills: {character} (the person's character, by name) · {photo} (the person's own photo(s) of the character) · {character sheet} (face close-up + full body in one image, from their character card) · {frame} (a film frame or any picture to put the character into) · {film}.

## 🏆 Source experiment pick — identity and timing

One model for every step: **GPT Image 2.5 edit** on fal (`openai/gpt-image-2.5/sunburst/edit` or `openai/gpt-image-2.5/flare/edit`), **quality `low`**, one image per request, one pass.

1. **{character sheet} from {photo}** → picture 1 (and 2) = the person's photo(s), a two-sentence prompt: "Character sheet of {character} from pictures 1–2: an extreme close-up of the face and a full-body front view side by side on a plain white background. Same person and same clothes in both views, no text." → `character-sheet`.
2. **Every later image** → picture N = {character sheet}, every other reference with **one job** named by role ("the pose of picture 3", "the place of picture 4"), in **one pass** → `character-consistent`.
3. **Into a film frame** → picture 1 = {character sheet}, picture 2 = {frame}: "Shoot it again exactly the same… but [the mode]… Change nothing else." 20–26 s → `character-replace`.

Why not Nano Banana: it is faster (6–9 s for Nano Banana 2 Lite), but it keeps losing the character — the people from the references survive in pose and expression sheets (pose sheet 5–8 of 10, expression sheet 0–7 of 9 against 10 of 10 and 9 of 9 on GPT Image 2.5) and the actor survives in about one film frame in six (GPT: 10 of 10 frames). The first consistency tests ran on Nano Banana 2 Lite in three passes; the later canon is GPT Image 2.5 in one pass (see `character-consistent`).

## Classify

| The job | Go to |
|---|---|
| No sheet yet: make it from the person's photo(s); turnaround, pose sheet, expression sheet, a back view | `character-sheet` |
| The same character in new images: pose, emotion, place, haircut, make-up, prop, palette, light, camera; mixing references | `character-consistent` |
| The character in a film frame: the whole character, the head, the face, the face keeping paint / make-up; the stare fix; refusals | `character-replace` |
| Animate the result (H3) | `animate` meta-skill |
| A music video (every face replaced, look sheets) | `musicvideo` meta-skill (`musicvideo-replace`) |
| Upscale a sheet | `upscale-sheet` (upscale meta-skill) |

Pipeline: `{photo}` → `character-sheet` → (`character-consistent` for stills · `character-replace` for film frames) → `animate`.

## Global rules (break none)

0. **Blind prompts**: picture roles only — never the character's or the frame's looks. Exception (owner, 9.10.2026): the expression of the frame (gaze, brows, mouth) may be named in words.

1. **The character comes in as a picture, never as words.** Picture roles only: "{character} from picture 1: their face, hair and clothes exactly as in picture 1". Never describe the face, hair, clothes of the character or the content of the reference / film frame — the model sees them; text pulls it toward someone else.
2. **The person's own material only**: their {photo} and the {character sheet} from their character card. A photo of a real person needs that person's consent. Never use a demo character's sheet, prompt, faces or frames for a person.
3. **GPT Image 2.5, quality `low`, set explicitly** — the edit schema defaults to `high`. `high` gave the same look as medium in the measured frame comparisons.
4. **Size = the target's aspect**, width ~1920 (heights multiple of 16: 1920×1088 for 16:9, 1920×816 for 2.35:1). Automatic size follows the first input — a sheet first gives a frame in the sheet's proportions.
5. **One pass.** Do not split a frame into frame → look → light passes (later canon). If a reference is lost, cut roles, do not add passes.
6. **Check every output with crops**: the face, the clothes, the back, leftovers of the reference person or the actor.
7. Refusals (`content_policy_violation`) returned in 3–26 s in the tests; a famous look kept on another face may never pass. Respect moderation and choose another permitted frame or mode.
8. Heavy compute in the cloud (fal); local work only for crops and charts.

## Settings and times (measured)

| Step · settings | Time |
|---|---|
| GPT Image 2.5 Sunburst edit · low · sheet / pose / expression sheet | not recorded |
| GPT Image 2.5 Sunburst / Flare edit · low · actor replaced, 1920×1088 or 1920×816 | 20–26 s (successful Sunburst requests on one frame: 19–52 s) |
| GPT Image 2.5 Flare edit · low · four-view sheet 2048×1152 · 1 / 2 references | — |
| GPT Image 2.5 Flare edit · low · 2K / 4K | 19 s / 35 s |
| Nano Banana 2 Lite edit (the first tests) | 6–9 s a view, 6–19 s a mix |
| Qwen Image 3 edit (failed mode 4) | 106–172 s |

## External tool workflow

Use your own supported image-edit tool to upload the reference images, set the model, explicit quality and target dimensions, then compare the result with the source. The model endpoints above identify the external experiments; no local runner is included. Use the `animate` reference for the subsequent video step.

## Example run (reference only — never reuse its files, prompts, faces or frames for a person)

Demo character **Mara Vale** (invented, not a real person): face and front full-body anchors, a two-view sheet, a 19-view collection, pose/expression charts and a back-view sheet. Film examples used *The Shining*, *Don't Breathe*, *This Boy's Life*, *Do the Right Thing*, *The Color Purple*, *Mad Max: Fury Road* and *Joker*; head/face modes were compared on Jack Sparrow (*Pirates of the Caribbean*) and William Wallace (*Braveheart*). These are experimental examples, not input material supplied by this download.