Shot four and shot five were the same woman, roughly. She had lost about eight years and gained a different jaw. Nothing in the prompt had changed except the location, which was the clue: the model was not drifting randomly, it was re-deciding who she was every time I did not tell it.
The fix turned out to be boring. A sheet of pictures, numbered, and a sentence that assigns each one a job.
The short version. One H3 request takes up to nine image references and three audio clips, cited by number in the prompt. The first five images are free; on 2K each additional image costs 3 credits.
What can one H3 request carry?
One image-to-video request takes up to nine reference images and three audio clips. They are cited by number in the prompt, so the slot a picture occupies is part of the instruction:
Image 1 is the protagonist: her face, age and freckles. Image 2 gives the same face in three-quarter view. Image 3 is her wardrobe. Image 4 is the location. Match the camera move from Video 1. Use Audio 1 for her voice.
That sentence is doing more work than any adjective I had been adding.
How do you build the reference sheet?
Five slots carry the decisions; the rest are supporting frames.
- Hero portrait — neutral light, eyes to camera. Owns identity: face, age, bone structure.
- Three-quarter view — stops the flat, passport-photo result.
- Wardrobe, full length, worn — silhouette and colour, not fabric detail.
- Location, no people — set, horizon, orientation.
- A light reference — the mood, key direction, contrast.
Slots six to nine are for the shot's own problems: a prop, a second character, a texture, a frame that shows the camera angle.
The order matters because it is the vocabulary. Put a prop in slot one and every sentence about the protagonist has to call her slot five.
Three rules that came out of the failures
- Name the property, not the object. "Image 1 is the face" beats "Image 1 is the woman", which invites the model to re-cast the role.
- One authority per property. Two portraits in different light force a choice that changes between takes. That was the eight-year drift.
- Number the audio and video too. Video 1 and Audio 1 are part of the same numbering, and a shot that follows another shot's camera move should say so.
Do extra reference images cost more?
The first five images are free. On 2K, the sixth image onward costs three credits each, so a full nine-image 2K take adds twelve credits. Below 2K there is no per-image surcharge at all.
Which means the cheap place to experiment with references is 768P, not the top tier. I do the sheet work there, then move the approved take up.
What references will not do
They will not carry continuity across generations. Nothing passes between requests except the sheet itself, which is why it should be a saved set rather than something rebuilt every time.
They will not fix conflicting identity — two portraits of the same person at different ages still average into a stranger. Fewer, better pictures beat more pictures.
And they will not rescue a shot with no recurring subject. A landscape or a texture plate does not need a sheet; it needs a prompt and a cheap tier. References are continuity insurance, and continuity is only worth buying when there is a second shot.
The reference limits and the per-image pricing for 2K are on the H3 page.
I work on Callirra, so the slot rules and the free-five pricing are the shipped behaviour rather than a reading of someone's docs. If your sheet holds a character better than mine does, the useful part is how you assigned the slots — that is the thing worth publishing.