Pika releases H3 Flex, generating 15-second clips in as little as 7 seconds

• H3 Flex can generate a 15-second clip in seven seconds, per the launch announcement. Pika’s footnote specifies transformer compute at 768p, with additional processing outside that measurement.
• Five editable audio stems are the standout feature for early-access user hasantoxr, separating dialogue, score, SFX, foley and ambience.
• Code, weights, ComfyUI support and Diffusers support are available through the release thread. High-resolution finishing and stem separation remain Pika-hosted features.
• Creators are publishing detailed shot lists: Artedeingenio’s anime action experiment uses start and end images with five timed beats between them.

Pika’s launch gallery includes a knitted game world assembled from seven reference assets and a charcoal tightrope film guided by five sketches and two photographs. The ComfyUI guide contains an unusual warning: increasing the step count won’t add detail.

Pika released H3 Flex on October 29, post-training MiniMax H3 to generate in six steps instead of 50. The company’s controlled benchmark measures seven seconds of transformer computation for a 15-second clip with synchronized audio at 768p.

Total request time also includes processing such as reference inputs, decoding, higher-resolution finishing, stem separation and service overhead. Local performance depends on hardware and implementation.

At launch, Pika claimed first place among open-weight models and seventh overall on DesignArena’s text-to-video board, ahead of base H3, H3 Max and Seedance 2.5.

Pika H3 Flex launch film

Every generated clip comes with synchronized stereo sound. Through the API, Pika’s model specification lists output_mode: audio-stems for five separate WAV exports:

• Dialogue
• Score
• SFX
• Foley
• Ambience

Stem separation costs an additional $0.001 per output-second, or 1.5 cents for a 15-second clip. That gives an editor separate files for changing the mix without regenerating the picture.

For hasantoxr, the attraction was fixing one bad sound layer without losing an otherwise usable generation.

Flex generates natively at 480p or 768p. Pika’s upsampler supplies the 1080p, 2K and 4K finishing passes.

The draft-to-video endpoint preserves the selected take and its audio, charging only the difference between the draft and finished resolution. Charging for the upgrade alone is a welcome bit of sanity.

A generation can combine character imagery, environments, motion footage and voice samples. The published reference limits are more specific than the headline count:

• 12 files total, subject to individual caps of nine images, three videos and three standalone audio clips.
• Two to 15 seconds per video or audio clip, with combined duration capped at 15 seconds for each media type.
• At least one image or video in reference mode. Audio-only reference requests are unsupported.

Pika’s demonstrations assign different jobs to different assets: a character from one image, an environment from another and choreography from a video. Reference-guided editing can also take existing footage and replace an outfit, prop or setting while being directed to preserve performance and timing.

Artedeingenio created The Assassin’s Last Stand using image-to-video with both endpoint frames. His 15-second prompt divides the action into five three-second blocks:

1. 0–3 seconds: Opening gunfire, followed by a whip pan revealing attackers.
2. 3–6 seconds: Close-quarters combat.
3. 6–9 seconds: Sliding gunfire and an overhead action shot.
4. 9–12 seconds: The fight stops; a metallic sound draws the assassin’s attention upward.
5. 12–15 seconds: The camera rises into the supplied overhead end-frame composition.

The prompt separately locks the protagonist’s face, beard, suit, injuries and illustration style to the references. It asks the final camera angle to emerge through movement rather than an abrupt visual transformation.

During early access, Artedeingenio also tested anime, cinematic action, sci-fi and historical scenes. Three published prompts show different ways of directing the model:

• Bullet Ballet: A vertical 9:16 corridor fight, constrained to three characters and one pistol. The prompt uses the narrow architecture for framing and cuts the soundtrack for the final three seconds.
• The Forgotten Astronaut: Every camera position stays inside the station, while a mysterious girl remains outside its observation window. Her moving shadow motivates the astronaut’s turn toward the glass.
• Gravity Is Broken: Artedeingenio scripted five timed beats, progressing from frozen raindrops to floating cars and a folded city. He named Christopher Nolan films as a visual touchstone in a follow-up.

The GitHub inference package and Hugging Face release provide two task adapters on top of MiniMax H3:

• FL2VA: Text-to-video and first-frame, last-frame or paired-endpoint conditioning.
• Ref2VA: Multimodal image, video and audio references.

Local inference produces native 480p or 768p video with mixed stereo audio. Pika retains the high-resolution upsampler, audio-stem separation, prompt enhancement and accelerated serving stack as hosted features.

ComfyUI’s templates use Pika’s six-step schedule with lcm, ManualSigmas and CFG 1.0. Its guide reports about a minute for a five-second 1344×768 clip in a single-card example.

The adapter weights use the MiniMax H3 Community License Agreement; the example Python code uses Apache 2.0. Pika reports representative local CLI runs on an H100 80GB.

Some early posts described weights as forthcoming. Minchoi subsequently confirmed they had dropped.

API access is available at launch. Pika’s announcement schedules H3 Flex’s arrival in its creation product for November 2.

The published pricing lists these text, image and keyframe generation rates, with launch pricing for API Club members from October 29 through November 5:

| Output | Launch price per second | Regular price per second |
| --- | --- | --- |
| 480p | $0.01 | $0.02 |
| 768p | $0.02 | $0.04 |
| 1080p | $0.035 | $0.07 |
| 2K | $0.04 | $0.08 |
| 4K | $0.08 | $0.16 |

At launch rates, a 15-second 768p draft costs $0.30. Finishing that same take at 4K adds $0.90, bringing the video charge to $1.20 before optional stems.

Reference-to-video pricing starts at $0.03 per second during the promotion. Video references incur charges for input-video seconds plus output-video seconds; image and audio references carry no additional input charge.

Pika also announced 15 free generations per day for two weeks, separately from the one-week launch-price window.