Pika H3 Flex Takes 12 References in One AI Video Shot. Build the Reference Pack Like a Shot List
A video model that accepts twelve reference files sounds like a gift. Nine pictures, three clips and three sound files: your actor's face, the location, a camera move you shot on your phone, the voice you want, the room tone. Hand it all over and get one shot with synced sound back.
In practice, twelve files is twelve chances to confuse the model. Give it a photo of your actor in a red coat and a location photo with a red door, and which red belongs to whom? Give it a clip for camera motion that also has a person walking through it, and is that person now in your scene?
Pika released Pika H3 Flex on October 29, with downloadable files on Hugging Face and the same model in its paid API. It makes video with sound from text, from a first and last frame, or from that pile of references. The interesting part is not the number twelve. It is the prompt guide Pika published alongside it, which reads less like advice on wording and more like a production document. Every reference gets a label, a role and a statement of how much of it to keep.
That is the workflow worth learning, and it works whether you run Flex locally, in Pika's API or in the free demo page.
What Flex does with references
Pika H3 Flex is a retrained version of MiniMax H3, an open model that makes picture and sound together. Pika's version is cut down to six generation passes instead of the fifty its starting model used, which is why it is quick.
In reference mode you can supply up to 12 files per shot: up to nine images, three videos and three audio clips. At least one image or video is required. Each video or audio clip must run 2 to 15 seconds, and the combined length of all video references, and separately of all audio references, must stay within 15 seconds. Output is 480p or 768p, 5 to 15 seconds long, at 24 frames per second with stereo sound.
Pika's showcase examples hint at what it is for: a character from one image placed in a setting from another, moving like the person in a phone clip; outfit or prop swaps that keep the original performance; product films; short multi-shot scenes with dialogue.
Where to run it, and what it costs
There are three doors in, and one of them is closed to a lot of readers.
Pika's API. Pika lists reference-to-video at from $0.03 per output second, so a 15-second shot starts around 45 cents. Text or image to video starts at $0.01 per second. The API also offers things the downloadable version does not: finishing to 1080p, 2K or 4K, and splitting the sound into five separate stems (dialogue, score, effects, foley or recreated physical sounds, ambience), listed at $0.001 per second for stems. Pika's launch post says Flex reaches its regular creation app on November 2, though the published page itself flags that date as unconfirmed.
The free demo. Pika has a demo page on Hugging Face, the model-sharing site (Pika-Labs/pika-h3-flex), running on shared hardware. Expect a queue. It is the cheapest way to test whether your reference pack makes sense.
Your own machine. This is the closed door. The downloadable files are released under the MiniMax H3 Community License, which excludes the United States, the European Union, the United Kingdom and South Korea, and says you may not use the model "or any of their Outputs" there. Pika is a US company, so its home market cannot run these files under that licence. If you are elsewhere, the hardware is the next wall: Pika's own tests used an 80 GB data-centre card, the short 480p case used roughly 66 to 69 GB of graphics memory, and the base files need at least 211 GB of disk. ComfyUI, the free node-based app many people use to run image and video models at home, ships ready-made templates for Flex, but we found no one showing it running on a desktop card yet.
For most readers, then, this is an API workflow. The pack you build is the same either way.
The idea: one file, one job
Pika's prompt guidance names four kinds of label. <Picture 1> is an image you supplied. <Video 1> is a clip. <Audio 1> is a sound file. And <Subject 1> is a thing you define: a person, an object, a style, a motion, built from one or more of your files.
Then it asks you to say what each file is for, and how much of it to keep. The guide's own vocabulary for pictures runs from fully_preserved through partially_preserved and attribute_transfer down to weak_reference. For sound it runs from fully_copy to weak_reference. You can also mark something removed.
Read those terms like a director's notes. "Keep her face, hair and outfit exactly." "Take only the camera move from this clip." "Use this voice's tone, not its words." "Borrow the mood of this room, nothing else."
The guide's example sentence is the pattern to copy: "<Subject 1> is the woman from Picture 1, retaining her face, hair and outfit." And for motion: "<Video 2> provides the forward camera trajectory."
Every file earns its place by doing exactly one thing. If you cannot write its job in one line, cut the file.
Put this into practice: build a pack for one 10-second shot
Pick a real shot from something you are making. Here is the routine.
1. Write the shot first, in plain words. Two sentences, no AI vocabulary. "A courier in a yellow jacket pushes through a crowded night market toward the camera. The camera backs away at walking pace; at the end she stops and looks up."
2. List what the model cannot guess. Usually: who (face and costume), where (location), how the camera moves, what it sounds like. Those become your references. Everything else, let the text carry.
3. Gather the minimum. One clean, front-lit photo of your subject. One photo of the location. If the camera move matters, a short phone clip of the move with nobody important in frame. If the voice or room sound matters, one audio clip. Four files beat twelve.
4. Clean each file so it can only mean one thing. Crop the subject photo so the background is plain. Avoid character sheets with several poses and labels; if you must use one, Pika's guide has you say outright that alternate views do not add characters, and to show one character "without the panels, labels or alternative views." Trim the motion clip to the move alone, and check its soundtrack, because a video with sound also creates an audio reference.
5. Write the subject definitions. One line per subject, with a keep level. "<Subject 1> is the courier from <Picture 1>, retaining her face, hair and yellow jacket." "<Video 1> provides only the backward camera move; its people are not in the scene."
6. Write the scene in playback order. Pika's guide asks you to establish the subjects, where they stand and the composition before the actions and their visible results. Name who does each action. State how many people are present. For the camera, give direction, distance, speed and where it ends. Start with [Shot 1]; if you want a cut, write the time, such as [Shot 2] At 00:06.000.
7. Write the sound separately. One to four sentences of ambience and physical sounds in the soundscape field. One to three sentences on score in the music field, or N/A. Dialogue goes in the shot with a speaker label, like (S1).
8. Generate once at 480p, then change one thing. If the face drifts, strengthen that subject line. If the background ignores your location, check whether another file is fighting it. Change one reference or one line per run, the way you would adjust one light at a time.
9. Save the pack. A folder with the files, the subject lines and the prompt that worked is a reusable asset. Next week's shot with the same character starts at step 6.
Where it breaks
Markers are requests, not guarantees. Pika's own guide says its keep levels "express prompt intent, not deterministic editing, literal waveform copying, or exact voice reproduction." You will get close. You will not get a pixel-perfect costume every time.
Some shot types favour other models. In Pika's own blind comparisons, Flex led in reference mode on visual style, framing and camera motion, and in text and image modes on anatomy, composition and sound sync. People preferred rival systems on physics, multi-scene stories, lighting and action-heavy reference shots. Plan accordingly: a chase scene built from references is the hard case.
Camera words drift. Comfy's tutorial notes that timed camera moves can land 0.2 to 0.3 seconds off or go the wrong way, that "wide shot" alone often comes out medium, and that a hand meant to stop short of something usually touches it. Describe the subject as small in a large space if you need width, and give every move a stopping point.
Resolution is a draft. The open version stops at 768p. Finishing to 1080p and up is a paid API feature.
The licence follows the output. If you live in an excluded territory, the downloadable files are off limits, outputs included. The API is a separate service under Pika's own terms, which are worth reading before client work.
The real skill here
It is tempting to treat twelve reference slots as twelve chances to get lucky. Fill them all, write a paragraph, hope.
Flex rewards the opposite habit, the one an assistant director or a costume supervisor already has: decide what must stay fixed, name it, and say exactly how fixed. The model is quick and cheap enough now that the slow part of the job is preparation, and preparation is where you have the most control.
Pick one shot this week. Build a four-file pack, write one line of job description per file, and run it in the demo before you spend a cent. If you cannot write those four lines, the shot was not ready, and that is worth knowing too.