Pika releases H3 Flex: reduces MiniMax H3 video-generation denoising steps from 50 steps to 6 steps
BIBIBI
AT A GLANCE
Pika released H3 Flex, stating that it can generate videos with 6 denoising steps, and made the adaptation weights and inference code public.
Article
Pika creates a high-speed version of MiniMax H3: generation steps reduced from 50 steps to 6 steps, with even better image quality
According to a report by Beating AI News, Pika released H3 Flex, a high-speed version of MiniMax H3。The team first fine-tuned H3 using self-built audio and video data, then used model distillation to compress the denoising steps required to generate videos from 50 steps to 6 steps. It subsequently improved image quality through reward and preference training, and further optimized underlying inference performance.
According to Pika’s DesignArena text-to-video leaderboard,H3 Flex ranked 7 overall and 1 among open-weight models, surpassing the original MiniMax H3 and H3Max。In official testing, generating a 15-second 768p video with sound took as little as 7 seconds of core model computation, excluding video decoding and service processing.
Pika has made the model adaptation weights and inference code public, supporting local execution through ComfyUI and Diffusers, with native generation of videos up to 768p。The advertised 4K requires Pika’s upscaling service; audio-track separation and accelerated inference services were not released together.
Original link https://m.theblockbeats.info/flash/374663
Key points
01
Beating AI News reports that Pika released H3 Flex。
02
The team first fine-tuned H3 using self-built audio and video data, then used model distillation to compress the denoising steps required to generate videos from 50 steps to 6 steps.
03
The report says that the team subsequently improved image quality through reward and preference training, and further optimized underlying inference performance.
04
According to Pika’s DesignArena text-to-video leaderboard,H3 Flex ranked 7 overall and 1。
05
Pika’s official testing states that generating a video with sound lasting 15 seconds took 768p video, with core model computation taking as little as 7 seconds; this time excludes video decoding and service processing.
06
Pika has made the model adaptation weights and inference code public, supporting local execution through ComfyUI and Diffusers, with native generation of videos up to 768p。
07
The advertised 4K requires Pika’s upscaling service; audio-track separation and accelerated inference services were not released together.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
H3 Flex is an accelerated version developed by Pika based on H3。Its primary approach is to use fine-tuning and model distillation to reduce denoising steps during video generation, then use reward and preference training to improve output quality. The public materials allow users to run it locally with ComfyUI or Diffusers, but the native output limit is 768p，4K and some supporting capabilities still depend on services that have not been fully released.
Why it matters to readers
If Pika’s reported speed, image quality, and rankings can be independently replicated,H3 Flex could reduce the computational steps required for video generation and provide more options for local execution; the public weights and code also facilitate testing and adaptation by developers.
Beginners can understand “50 steps reduced to 6 steps” as meaning that the number of denoising rounds the model performs when generating videos is reduced. However, this does not mean the complete service can necessarily deliver a video within 7 seconds, because the official timing excludes video decoding and service processing. Also note that the native local maximum is 768p; the advertised 4K requires an additional upscaling service.
Risks and unknowns
Leaderboard rankings, 6 denoising, 7 seconds of computation, and image-quality improvements are based primarily on materials published by Pika or official testing; the evidence contains no independent replication.
7 seconds counts only core model computation, excluding video decoding and service processing, and cannot directly represent the complete generation time.
Related Developments
Loading event timeline…
The evidence does not provide the specific licensing terms, hardware requirements, or actual accessibility of the adaptation weights and inference code.
4K Depends on the Pika super-resolution service; the audio-track separation and accelerated inference services are not available, and their usage methods and costs are unclear.
Related concepts
Fine-tuning
This term is not in the glossary yet. Browse related concepts in the glossary.
