Prompt breakdown: consistent character across 12 shots without training a LoRA

by @nikolaikim_studio in c/tutorials · · Prompt Breakdown

Everyone reaches for a LoRA for character consistency. You often do not need one for a short sequence. Here is the no-training method that got me 12 coherent shots of the same woman.

Lock the character in one hero still first. Generate until you have ONE image of her you love. This is your source of truth.

Use it as an image reference (--cref in MJ, or IPAdapter in Comfy) at moderate weight. Too high and every shot is the same pose. Around 50-60% weight keeps the face and lets the pose change.

Describe her the SAME way every single prompt. Same five identity tokens, same order, every time: 'woman, early 30s, sharp jaw, copper undercut, small scar left eyebrow, olive field jacket'. Consistency in = consistency out.

The scar/undercut/jacket are anchors. Give the character 2-3 unusual, specific features. Generic features drift. Weird-specific features the model clings to.

Result in comments. It is not LoRA-perfect but for a 12-shot sequence it is 90% there for zero training time.

344 points · 10 comments

Comments

  • @ethanchen1452 · · 73 points

    the '2-3 weird specific features as anchors' tip is gold, generic characters drift because there is nothing for the model to hold onto, the scar gives it a handle

  • @theohoffmann_fx · · 40 points

    exactly, I call them 'load-bearing details', one memorable scar does more for consistency than ten adjectives about her being beautiful

  • @sashapatellab · · 31 points

    cref weight around 50-60 is the sweet spot I landed on too, people crank it to 100 and wonder why every shot is the identical frozen face

  • @felixcosta_art · · 24 points

    for anyone in Comfy instead of MJ: this is IPAdapter at ~0.6 plus the same identity tokens, same principle exactly, works great

  • @emmadavis · · 16 points

    can confirm, IPAdapter FaceID is even better for the face lock specifically if you can run it

  • @noahpatel_studio · · 11 points

    does this hold up going into video or does the face drift the second you animate it

  • @harperfloreslab · · 14 points

    the face holds for the first frame then drifts over a 5s clip, for video you really do want the LoRA or a face-swap pass, this method is for still sequences