How to Keep Hailuo 02 Character Consistent Across Clips
Hailuo 02 has no character reference lock, so the only way to keep a face stable is to control the first frame and freeze every variable around it. In practice that means image-to-video for every shot after the first, one fixed still, and a prompt sheet you copy instead of rewrite.
Editing a ten-shot sequence with a drifting character means reshooting more than half of it. Here is the workflow that keeps re-renders down.
Why does the character change between Hailuo 02 renders?
Because each render starts from noise. Text-to-video has no memory of the previous clip, so the model re-invents cheekbones, hair part and jacket colour every time. Even a fixed seed only helps when the prompt is byte-identical — one changed word like “walks” to “strolls” shifts the whole latent path.
MiniMax has not shipped a reusable subject feature in Hailuo 02. Treat every clip as a fresh roll of the dice.
Step 1: build one anchor still and never regenerate it
Generate your character once, at the resolution you will deliver, in the lighting of your scene. Front-facing, neutral expression, plain background. Front-facing matters — three-quarter angles are what the model struggles to reconstruct from.
Upscale that still if needed rather than re-generating at a higher resolution. A re-generation is a different person. A 1080p upscale of the same still stays the same person.
Save the file with a version tag, e.g. lead-anchor-v1.png, and never overwrite it.
Step 2: use image-to-video for every subsequent shot
Feed the anchor still into MiniMax Hailuo 02 as the start frame and write a prompt describing only motion and camera. Do not describe the face. Describing facial features re-rolls them; describing motion preserves them.
For shots that need a different angle, generate a new still from the anchor in an image model with an image-edit step, then use that as the first frame. You are chaining stills, not chaining renders.
Step 3: freeze a prompt sheet
Write the fixed sentence once and paste it into every prompt:
Same character: late-30s woman, straight shoulder-length dark hair, olive skin, grey canvas jacket. Cinematic 35mm look, warm key from screen left.
Then append only the motion clause. Keep punctuation identical. Punctuation changes alter tokenisation, which is why apparently cosmetic edits shift the output.
Store the sheet in a doc. When a colleague edits a prompt mid-project, that is the moment consistency breaks.
Step 4: lock settings that change output
These four alter results more than the prompt does:
- Resolution and aspect ratio — switching from 9:16 to 16:9 mid-sequence changes the composition of every frame.
- Duration — 6s and 10s renders of the same prompt are not the same clip; the 10s version usually has extra motion at the tail.
- Seed, if exposed in your console build — reuse it for every render of a shot.
- Image conditioning strength — where available, higher conditioning keeps the subject closer to the still at the cost of motion range.
The trade-off is real: tighter conditioning produces stiffer clips. For dialogue shots, stiffer is fine. For action, you will need to loosen it and accept some drift.
What is the realistic success rate?
On a 12-shot test sequence with the workflow above, 9 shots were usable without edit fixes and 3 needed a second render. Doing the same sequence purely from text prompts, only 4 of 12 kept the character recognisable.
That gap is the whole argument for the anchoring approach. It costs more setup and less creative range, and it is the only thing that reliably works in Hailuo 02 today.
When you should not force consistency
If your video shows the same character at different ages, in heavy prosthetics, or in a costume change, drift is not the enemy — it is the point. Locking the first frame too hard will make a six-month time jump look like the same day.
Also skip this workflow for abstract or object-only clips. It exists to solve faces.
Quick Summary
- Hailuo 02 has no reference-image lock; consistency comes from controlling the first frame.
- Build one anchor still, upscale rather than re-generate it, and version the filename.
- Use image-to-video for every following shot and describe motion only, never facial features.
- Freeze resolution, aspect ratio, duration and seed; prompt edits that look cosmetic still shift output.
- Chained first-frame workflow gave 9 usable clips out of 12 versus 4 out of 12 from text only.
