If you make ads or story scenes with generative video, you already know the failure that kills projects: the subject falls apart. A face warps by the third second, a bottle label smears during a rotation, a jacket changes color between takes. Early diffusion models made consistency a scavenger hunt. Keeping one character stable across cuts meant training a custom LoRA, writing two hundred words of prompt chains, or rolling fifty seeds to find one usable clip.
The reference-to-video pipeline in MiniMax H3 Max, hosted on fal, replaces that lottery with evidence. You attach real photographs, movement clips, and audio tracks to a single generation as explicit conditioning files, and the model draws the shot from your own material instead of guessing from words. One run accepts up to twelve multimodal inputs: nine images, three video clips, and three audio files. This guide walks through the setup we use, step by step, based on forty benchmark runs across an actor dialogue sequence and a cold-brew bottle spot.
One boundary first, because it matters more than any parameter: reference conditioning requires complete ownership or written licensing of every human likeness and every trademark you upload. Feeding the endpoint a competitor's campaign still or a scraped actor photo is a rights problem no model setting can fix. None of this is legal counsel.
Step One: Pick the Shot References Can Actually Lock
The common mistake is loading ten files into a wild creative leap and wondering why the output drifts. Reference conditioning is a physics anchor, not a mood board. Choose one shot with one clear job.
For a brand spot, that might be a bottle rotating on a turntable while light sweeps across the label. For narrative work, a character seated at a cafe table delivering a line, with a specific hand gesture and a matching voice tone. Both share one property: the subject and the action are locked, and only the camera or environment moves. If your concept needs the subject to transform, that is a different tool.
Once the shot is chosen, assemble a reference package. The files you gather depend on what the scene is built from:
Brand or product shot: three to five high-resolution product stills from different angles, one two-second motion clip showing the physical action, and one clean audio file for the room tone or mechanical sound.
Character dialogue: four to six portrait photos covering frontal and profile angles, one short clip of the actor's real movement pattern, and one voice recording to lock the timbre.
Stylized scene: up to six images for wardrobe, fabric, and lighting references, one clip for camera choreography, one audio bed for pacing.
Keep every file sharp. Blurry uploads teach the model blur. And keep the package small on purpose: three good references beat twelve mediocre ones, because every extra file splits the model's attention.
Step Two: Load the Endpoint and Index Everything
On fal, open the H3 reference-to-video playground. The interface exposes dedicated slots for image, video, and audio inputs, and the order you drop files in is not decorative. The model assigns each file an index number, and that index is the vocabulary of your prompt.
Set the base parameters before uploading anything. Resolution at 768p for the structural hold, duration at five to eight seconds, aspect ratio matched to your delivery target, and prompt expansion left on its default so the engine tidies your motion description without a rewrite delay. Then upload in a deliberate order: identity files first, motion references second, audio last. Note each index as you go, because your prompt will refer to them by number.
Step Three: Write the Prompt as a Director, Not a Describer
Here is the mental shift that separates clean takes from mush. The references already say what things look like. Your prompt should stop describing static appearance entirely and start assigning roles to the files.
A working pattern looks like this: use the face from image 1 and the jacket texture from image 2, replicate the camera push-in from video 1, and have the character speak the line in the voice of audio 1 while the background hums at the level of audio 2. Every clause ties one numbered file to one job. The model is told what to copy, in what order, and what to leave alone.
Two rules keep this from falling apart. First, one action per take: walk in, sit, deliver a line, all in one five-second window and the model sprints through frames or drops the ending. Second, never re-describe what a reference already shows. Prompting a red jacket when image 2 is the jacket invites the model to repaint it and drift the color.
Step Four: Review for Drift Before You Trust the Take
Generation on the live endpoint lands in roughly fifteen to forty seconds depending on duration and load. The review pass is where the forty-run method earns its name. Scrub at half speed and check three gates:
Identity gate: does the face, logo, or product silhouette survive from the first frame to the last? Warping by second three means your reference angles were too narrow. Add a profile or top-down still and rerun.
Motion gate: does the movement read as continuous, or does the model snap between states? Snapping usually means the motion clip was too long or contained two actions. Cut it to the essential two seconds.
Audio gate: does lip or mechanical movement land with the sound? Reference audio locks timbre and pacing, not perfect sync, so expect to nudge the track in post.
Log what fails and why. After a few dozen runs, the failure patterns become predictable, and predictable is the whole point of a production pipeline.
Know the Limits
Reference conditioning is powerful but bounded. The model honors the files you give it, which means garbage in, garbage with better lighting: a single frontal face photo still leaves the profile to imagination, so supply the angles. Output tops out at 768p, so a 4K deliverable still needs an upscaling pass. Fine text and small labels can soften during fast rotations before snapping back, so keep critical typography out of heavy camera moves. And exact pose precision mid-motion is not guaranteed by anyone, model vendor or host. You are directing with evidence, not with a keyframe on every joint.
Questions That Come Up
Can I reuse one reference package across multiple shots?
Yes, and you should. Index the files once, keep the package in your asset library, and only change the motion clause per shot. Consistency across a campaign comes from reusing the same identity files, not regenerating them.
What happens if I upload more than twelve files?
The endpoint rejects the excess or errors on the request. Twelve is a hard ceiling: nine images, three videos, three audio. Curate within it.
Why does my character's face pass the first seconds and then drift?
Almost always narrow reference coverage. Frontal-only photos leave the model to invent every other angle. Add profile and three-quarter stills of the same person, same lighting, and the drift usually disappears.
Does a bad-looking result get refunded?
No. Pay-as-you-go platforms bill the raw compute seconds consumed. A generation that completes without a system error is billed whether or not it matches what you wanted.
Can the reference audio drive lip sync?
It locks voice timbre and pacing, and the model attempts to match mouth movement to it. Treat the result as a strong draft and expect final polish in an editor, not a broadcast-grade sync out of the box.
The Takeaway
Character and product drift is not a mystery to pray about, it is a missing-evidence problem. H3 Max on fal lets you hand the model the proof: the face, the fabric, the movement, the voice, all as indexed files in one call. Pick a shot with one locked subject and one action, build a small sharp reference package, write the prompt as role assignments instead of descriptions, and gate every take through three review checks. Do that and consistency stops being a lottery and starts being a pipeline.
