H3 Max: image to video

Give it a first frame, a prompt, and optionally a last frame or reference pictures; get a clip back in a fraction of the time and price of Kling or Seedance, with good motion and acting. The cost is fine detail: facial hair, skin texture and hair strands come out mottled and smoothed together, painted rather than photographic, and close-ups look plastic. Draft on it, keep it for wide shots and subjects with little texture, and move to kling-v3-image-to-video (close-up texture) or seedance-2-5-image-to-video (physics, 1080p grain) where the shot needs it. For a still speaking an audio clip use h3-max-lipsync-a-still; for a scene with a dialogue track, h3-max-dialogue. Always H3 Max, never plain H3: Max is its fine-tune.

Inputs

The first frame. The output canvas follows its shape.
A last frame (optional), continuous with the first: derive it from the locked first frame (nano-banana-derive-last-frame). Two independently edited stills made her hair morph mid-shot.
Reference pictures (optional) for subjects the frame doesn't show well or who enter: up to 9 subjects and 9 pictures in all. Every subject keeps its first picture, the rest are dealt round-robin. Single views from a character sheet, not the sheet.

Run it

Call describe on the model before the first run in a job. Before that first run, a test included, show the person the still that will move and ask (core skill: the person decides every look).

run("fal/minimax/h3-max/image-to-video", {
  "image_url": "<job>/stills/shot04_first_c.png",
  "prompt": "<H3 prompt, see below>",
  "duration": 3.2,
  "resolution": "768P",
  "prompt_expansion_mode": "disabled",
  "seed": 5000
}, into="<job>/video/raw", unit="shot04_h3", tag="shot04")

duration: the length you want, in seconds as a float, from about 1 s up to 15 s. The clip may come back a little longer than asked (ask for 1 s, get 1.2 to 1.4 s); trim it to the shot. H3 returns 24 fps. Shots longer than about 15 s aren't advisable on any video model: split the shot at a cut or a pause.
resolution: 480P, 768P or 1080P, uppercase P. 1080P is 768P plus latent refinement. Priced per second by tier whatever the shape: 768P $0.052/s and 1080P $0.104/s at the promo rate ($0.08 and $0.16 at list).
prompt_expansion_mode: "disabled": required by the schema, and fal's "balanced" rewrite adds a soundtrack.
end_image_url pins the last frame. seed is optional; vary it for takes.
With reference pictures, call fal/minimax/h3-max/reference-to-video instead, with the same fields plus "reference_image_urls": [...], in the order of the prompt's <Picture N> labels. The first frame is still image_url. Each picture must be 256 to 2,048 px a side with an aspect between 0.4 and 2.5, or fal rejects the whole request (wide logos are the usual offender): pad with the edge colour, never crop.

Prompt

Load h3-prompt-format first: the image-to-video shape without references, the reference-to-video shape with them. For this mode:

Say "big, clear motion" in the direction and spell out counts ("blinks exactly once").
H3 pushes in by default; state "static shot, locked on the opening framing" when the frame must hold.
Never name a removed accessory, even to forbid it: "no pin or badge" grew a pin by 1 s. Describe the clean state instead ("a plain, unadorned knit, the same from the first frame to the last").
Ask for the shot's own length. If the clip still runs longer than the shot, ask for the action in real time, then a hold.

What you get

An MP4 at 24 fps, opening on the first frame, a little longer than the duration asked, with generated sound. Trim to the shot (ffmpeg-frame-exact-trim), keeping the raw render.

Tips and traps

Plastic skin on an otherwise good take: add grain in post (ffmpeg-grain-for-plastic-skin) before re-rendering on another model.
Not for complex physical motion (papers falling, cloth in wind), 1080p photographic realism, or close-ups where beard, skin or hair texture must hold.
A draft approves composition and timing, not motion. Re-judge the motion on the final model.
H3 sometimes reaches for a raised index finger however it's directed; take the least-bad take rather than looping.

Check it

Frame 0 against the still (and the end against a pinned last frame): python-check-pinned-frames.
Close-up texture at 100% against the source, if the shot is a close-up.
Then a review page beside the original or the earlier round.

Hypernatural