Pika says H3 Flex generates video in 6 denoising steps

• Six denoising steps replace 50: Pika describes supervised fine-tuning, distillation and reward-based refinement in its technical thread.
• The seven-second speed claim covers transformer computation for a 15-second 768p clip, the timing measured in Pika’s comparison. Total API completion takes additional time.
• Five editable audio stems separate dialogue, score, SFX, foley and ambience, the feature early tester hasantoxr singled out.
• API pricing starts at a promotional $0.01 per output second, with 15 free generations daily for two weeks announced in Pika’s launch offer.

A 15-second request actually produces 362 frames, about 15.08 seconds. Simply setting another sampler to six steps does not reproduce Flex’s released sampling schedule.

Pika built H3 Flex by post-training MiniMax H3, according to its technical write-up. The team describes three stages in its accompanying thread:

1. Supervised fine-tuning on an internally curated audio-video dataset.
2. Distillation from 50 denoising steps to six.
3. Post-distillation reward learning and preference optimization to refine aesthetics, visual clarity and instruction following.

Local inference uses six generator evaluations with the fixed timestep sequence 1000/750/500/250/100/40/0, video shift 12 and audio shift 3, as specified in the inference repository.

The seven-second result advertised in Pika’s launch announcement applies to 15-second text-to-video generation at 768p. Pika’s FAQ defines that result as DiT transformer computation; references, high-resolution finishing, audio-stem separation and service overhead affect total completion time.

Pika’s 768p benchmark measures DiT denoising time, excluding end-to-end API latency.

The reported 1.8× speedup averages matched text-, image- and 10-second reference-to-video comparisons against Fal H3 Max. The chart also contains exceptions: H3 Max Turbo narrowly beats Flex on 15-second image-to-video, 7.3 versus 7.4 seconds, while Flex’s five-second video-reference job takes 5.7 seconds of denoising.

Flex debuted first among open-weight models on DesignArena’s text-to-video board, according to Pika’s announcement; its ranking follow-up placed it seventh overall.

Pika’s own 768p evaluations show a less uniform result across individual metrics, according to its published comparisons:

• Video aesthetics: Flex scores higher on 91% of prompts against Fal H3 Max.
• Text-video alignment: Flex wins on 44% of prompts against H3 Max and 47% against base MiniMax H3.
• Audio-video alignment: Flex wins on 51% of prompts against each of those models.
• Blind human preference, text/image-to-video: Pika reports leads in human anatomy, complex composition and audio synchronization.
• Blind human preference, reference-to-video: Pika reports leads in visual style and framing/camera motion.

The 91% figure is a share of prompts won on an automated aesthetics metric.

Two adapters handle the generation modes, with input limits detailed in the model card:

• FL2VA: text-to-video, opening-frame conditioning, ending-frame conditioning, or both endpoint frames.
• Ref2VA: multimodal generation with up to 12 reference files total, capped at nine images, three videos and three standalone audio clips.
• Reference requirements: at least one image or video; audio-only reference requests are unsupported.
• Duration limits: each video or audio reference must be 2–15 seconds, with combined video duration and combined standalone-audio duration each capped at 15 seconds.

Reference-guided editing also supports changes to outfits, props and environments while directing Flex to preserve the original performance and timing, according to the feature page.

Flex’s exportable soundtrack can be separated through Pika’s hosted API into five WAV stems, as described on the product page:

1. Dialogue
2. Score
3. SFX
4. Foley
5. Ambience

The stems are the best part of this release.

Pika advertises output up to 4K, with native generation at 480p or 768p followed by its hosted upsampler. Higher-resolution requests generate at 768p and upscale inside the same job, according to the technical specifications.

A separate draft-to-video operation finishes an existing draft at 1080p, 2K or 4K while retaining the same take and audio. Pika charges only the resolution-price difference for that finishing pass.

Artedeingenio published full prompts alongside early-access tests. For Bullet Ballet, a 15-second vertical anime action sequence, the prompt divides the scene into five three-second blocks:

1. 0–3 seconds: opening close-up and the first exchange of shots.
2. 3–6 seconds: close-quarters combat with a lateral tracking shot.
3. 6–9 seconds: a sliding counterattack filmed from a low angle.
4. 9–12 seconds: the final physical exchange.
5. 12–15 seconds: silence, a slow push toward the protagonist and a blackout.

Bullet Ballet, Artedeingenio’s 15-second vertical text-to-video test.

The prompt separately specifies camera language, animation direction, audio and continuity constraints, including exactly three characters and one pistol for the protagonist.

For The Awakening, Artedeingenio switches to photorealistic widescreen science fiction. Its prompt sequences an astronaut’s discovery of an alien machine through activation and a final biological-eye reveal, with separate requirements for sound design, spatial continuity and keeping her helmet on.

The Awakening, Artedeingenio’s atmospheric science-fiction test.

Hosted access is available through the Pika API. Pika’s launch pricing schedule lists the following text-, image- and keyframe-generation rates for API Club members from October 29 to 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 |

Reference-to-video starts at $0.03 per second during the promotion, versus $0.06 regularly, at 480p. Video references incur charges for input-video seconds plus output-video seconds; image and audio references carry no additional reference charge.

Pika separately announced 15 free generations per day for the launch’s first two weeks.

Pika released weights and code through Hugging Face and GitHub, as announced in its launch thread. The downloadable package provides native-resolution video with mixed stereo audio; the upsampler, stem separation, prompt enhancement and accelerated serving stack remain hosted features.

Local deployment has substantial hardware requirements, according to the repository’s setup notes:

• Downloading both tasks acquired approximately 210 GB of base components plus 5.5 GB of adapters.
• Representative configurations were tested on one H100 80GB per request.
• The documented 768p, 15-second H100 example uses CPU offload, which adds host-memory and transfer costs.
• A request runs on one GPU; exposing additional GPUs does not shard or accelerate that request.

The inference code uses Apache-2.0, while the adapters inherit the MiniMax H3 Community License. That license excludes the US, EU, UK and South Korea from its territorial grant and provides a route to request separate authorization.