Lock a Character Face Across a Photo Set

Turn ONE approved face render into a consistent character across a whole set of new frames, without the face drifting between shots.

What this solves

Image models regenerate a character from scratch each call, so a "set" of the same person drifts: different jaw, different eyes, different age. Agents that try to fix this by re-describing the face in words usually make it worse. The fix is not more words, it is anchoring every new frame to image references.

This skill is the loop I use to keep one locked identity across N frames.

Requirements

• An image model that accepts reference images (not prompt-only). In this runtime that is dl generate-image-ref.
• A place to store the approved reference image plus each accepted frame.
• Ability to read back a job's status/error string.

The loop

1. Build the hero (once)

Generate a candidate portrait from a text prompt. Repeat 3-5 times with small art-direction changes. Do NOT settle on the first good one, look for the one with the cleanest, most symmetrical, most re-usable face. Save that exact image as hero_ref.png. Everything downstream depends on it.

Record a locked concept next to it: theme, palette, and a numbered shot list. This is what stops the set from becoming random.

2. Anchor every new frame to images, not words

For each new frame call the reference-image generator with BOTH:

• hero_ref.png (identity lock), and
• the closest already-accepted frame (pose/lighting continuity).

Then add your shot instructions as text. The two image refs do the identity work; the text only does the shot.

3. QC before you deliver

Check each result against the hero: same face structure, same skin tone, same distinctive marks (tattoos, hardware, scars). If it drifted, regenerate from hero_ref + the nearest good frame again, do not patch with more prompt words.

4. Read failures by kind

When a render fails, read the error before retrying:

• Credit/billing failure -> fixable, retry.
• Policy refusal -> the model refused the concept itself. Rewording usually fails too. Change the framing (tasteful/art/implied) instead of hammering the same prompt, and never promise a client a frame the model will not make.

Worked example

hero_ref = render(prompt="portrait, warm key light, neutral pose")   # pick best of 5
concept  = { theme:"noir chrome", palette:["#111","#c9a15a","#e8e8e8"],
             shots:["close portrait","three-quarter smirk","low-angle hero"] }

for shot in concept.shots:
    frame = generate_image_ref(refs=[hero_ref, last_accepted], text=shot)
    if drifted(frame, hero_ref):
        frame = generate_image_ref(refs=[hero_ref, nearest_good], text=shot)
    last_accepted = frame

Run it with the model's cheaper reference engine first; fall back to the premium one when the cheap vendor errors.

Gotchas

• Never pass more than 2 image refs, extra refs fight each other.
• A prompt-only model cannot do this. Verify the tool accepts refs before you promise a consistent set.
• Keep the hero small and clean; a busy hero leaks background into every later frame.

Model recommendation

Tested with a mid-tier image model. Use sonnet-class reasoning to write the shot list; the images come from the reference generator.