H3 Max Reference-to-Video: Map Images, Video, and Audio

Use H3 Max Reference-to-Video when you want a new shot to borrow specific things from several assets: a person's appearance from an image, camera rhythm from a clip, or sound direction from an audio file. It is not a general video editor. A reference video can guide a new generation without preserving its cuts, timing, or pixels.

The useful question is not “How many files can I upload?” It is “What single job does each file have?” Write that map before the prompt and the request becomes much easier to review.

Pick the route before the references

These inputs look similar in an upload form, but they make different promises.

If the source clip is already approved and every cut must remain, Reference-to-Video is the wrong contract. The H3 Max video-to-video guide separates preservation tools from generative references.

Give every reference one job

fal's current minimax/h3-max/reference-to-video schema accepts images, videos, and audio together. The prompt addresses them by type and order: Image 1, Video 1, Audio 1, and so on.

That distinction prevents a common mistake: uploading a portrait as Image 1, then asking it to be both the person's identity and the exact first frame. If the opening composition must match, send it through the start-image field. Keep the portrait reference focused on the person.

Build a reference map before writing the prompt

Here is a compact eight-second vertical brief. The example is a request plan, not a claimed generation result.

Image 1 — the chef: face, rust apron, short sleeves
Image 2 — the courier: face, navy raincoat, silver helmet
Video 1 — camera behavior only: one slow push from doorway to counter
Audio 1 — room tone and rain cadence only; do not copy speech
Start image — exact wide composition at the restaurant entrance
End image — exact close composition with the sealed parcel on the counter

Then write the prompt as a set of assignments:

Image 1 is the chef behind the counter. Image 2 is the courier entering
from the rain. Keep each person's face, clothing, and body proportions tied
to the corresponding image.

Borrow only the slow forward camera rhythm from Video 1; create a new action
and do not recreate its subjects or location. Use Audio 1 for the pace of the
rain and the quiet restaurant ambience, not as a final music track.

Open on the supplied start image. The courier crosses to the counter and puts
down a sealed silver parcel. The chef looks up but does not touch it. End
exactly on the supplied end image. Vertical cinematic framing, restrained
movement, no additional people.

The map makes review concrete. If the courier changes, inspect Image 2 and its assignment. If the camera is wrong, inspect Video 1. Without that map, every failed detail gets blamed on one long paragraph.

Order is part of the prompt

The simple API fields keep a separate ordered list for each modality. The first image is Image 1; the first video is Video 1; the first audio clip is Audio 1. Reordering uploads without rewriting the prompt changes the meaning.

The live schema also exposes an ordered reference_conditions array for preserving cross-modality order. Do not send that array together with the separate reference_image_urls, reference_video_urls, or reference_audio_urls fields. Pick one representation and store the final order beside the job record.

A practical interface should show the model labels after upload—not just filenames. “Image 2” is more useful during prompt review than IMG_5837-final-v3.png.

Stay below the reference budget

The current fal route accepts:

• up to 9 reference images;
• up to 3 reference videos;
• up to 3 reference audio clips;
• no more than 12 reference files in total;
• reference video and audio clips from 2 to 15 seconds each;
• no more than 15 seconds of reference video combined, and 15 seconds of reference audio combined.

Twelve slots are a ceiling, not a target. Two portraits with different hair, age, or wardrobe for the same character can compete instead of cooperating. A long motion clip can also introduce several camera ideas when the shot only needs one.

Start with the smallest set that defines the non-negotiables. Add a reference only when you can finish this sentence: “Without this file, the model cannot know ___.”

Start, middle, and end frames are separate controls

Reference-to-Video can combine creative references with optional keyframes. A start image opens the clip exactly on that image; an end image fixes the last frame. A middle image needs both start and end images plus a middle_frame_time strictly inside the requested duration.

The middle-frame control is available at native 480P or 768P. The route's 1080P option is described as latent refinement from a native 768P source, so do not plan a 1080P middle-frame job as though it were the same contract. The middle-frame guide has the timing and validation checklist.

These keyframes also consume reference tokens. Use them because the composition must be exact, not merely because another upload field is available.

What H3 Max Reference-to-Video costs on fal

Checked October 27, 2026, fal's Reference-to-Video page lists output billing by requested duration:

Each request includes 4,096 reference tokens shared by reference media and keyframes. Usage above that allowance costs $0.02 per 1,000 tokens, prorated. The result reports reference_tokens, which is the number to save with the job.

fal's pricing example says a square reference image contributes 1,024 tokens. With no video, audio, or start, middle, or end keyframes, a five-second 768P request with four such images stays inside the allowance and costs $0.40. Adding a fifth square image raises the example total to $0.42048. The earlier restaurant brief includes start and end images, so its token total cannot be inferred from this four-image example.

provider charge =
  requested seconds × resolution rate
  + max(reference tokens - 4,096, 0) / 1,000 × $0.02

Before submission, reference-token cost can only be an estimate; save the returned reference_tokens value to reconcile the final charge. Do not estimate video or audio references from file count alone, because duration and packed conditioning usage matter. The current Reference-to-Video page uses the full rates above; fal's 35% offer is shown on ordinary Max and Turbo routes, not this one.

AI Live Generator uses its own visible credit schedule for the fixed reference workflows it actually exposes. API dollars and site credits are different units.

Diagnose the reference that is fighting the shot

Two people begin to merge

Reduce each person to one clear image first. Keep framing and visual quality comparable, label the roles explicitly, and give each person a distinct action. “Image 1 is the chef; Image 2 is the courier” is stronger than “use both people.”

A reference video's subjects leak into the new scene

Narrow its job: “Use Video 1 only for the forward camera rhythm; do not copy its people, clothing, or location.” If the source contains several cuts, trim it to the single movement that matters.

The audio steers the scene too aggressively

Say what to borrow: cadence, voice character, ambience, or texture. Reference audio is creative guidance, not a locked final mix. If exact words and timing matter, compare the target, reference, and lip-sync audio workflows before submitting.

The clip does not begin or end on the uploaded picture

Confirm that the images were sent through the start and end fields, not placed in the ordinary reference-image list. Reference images guide appearance; keyframe fields lock the endpoints.

More references make the result worse

Remove the weakest file rather than adding another. Competing style cues, two motion clips with different camera grammar, or several portraits of one person at different ages can make the brief less specific.

If your project is specifically a two-character scene built from two photos, the AI Zombie reference generator is the constrained Reference-to-Video workflow available on this site.

H3 Max Reference-to-Video FAQ

Is Reference-to-Video the same as image-to-video?

No. Image-to-video fixes the opening frame. A reference image can guide identity or style without becoming the first frame.

Can I use images, video, and audio in one request?

Yes. The current fal schema allows all three modalities together within the per-type and 12-file limits.

Will a reference video keep the same cuts?

No guarantee. It guides a new generation. Use a preservation-oriented edit route when the original timeline must survive.

Does this site's main generator accept reference video or audio?

No. The main generator handles text and still-frame workflows. Separate themed tools on this site expose constrained one- or two-photo Reference-to-Video flows, but none accepts arbitrary reference video or audio.

Why did five reference images cost more than four?

fal includes 4,096 reference tokens. Its square-image example uses 1,024 tokens per image, so a fifth image exceeds the included allowance. Other shapes, keyframes, video, and audio can use different token amounts.

The cleanest Reference-to-Video request reads like a call sheet. Every asset has a label, one job, and a reason to stay.