768p draft
4K finishDraft at 768p, finish at 4K
Explore at lower resolution, then upscale to 1080p, 2K or 4K in finishing mode. The take, framing and audio are preserved.
- 768p
- 1080p
- 2K
- 4K

Our faster-than-realtime model.
Built for professional iteration.
Pika H3 Flex
Pika H3 Flex is a post-trained H3 model designed to give professional creators the speed to explore and the control to be precise. It keeps the quality you need while generating faster than real-time in almost every configuration.
Resolutions reach 4K, references reach twelve inputs, and editing happens through a reference plus a prompt. Exportable audio stems let you isolate and refine any layer of the soundscape.
Speed and price
faster than H3Max
cheaper at matched quality
On average, across matched text-to-video, image-to-video and reference-to-video comparisons. DiT time, not guaranteed end-to-end latency.

A 15 second clip in as little as 7 seconds. Creation at the speed of thought, so the next idea is always one take away.
*Best-case performance.Built for tinkering
Fast drafts are only useful when they can become final shots. Flex keeps the take, the audio and every reference intact from the first pass to the finish.
768p draft
4K finishExplore at lower resolution, then upscale to 1080p, 2K or 4K in finishing mode. The take, framing and audio are preserved.
Combine sketch, image, sound, video and moodboards in one request. Up to twelve inputs hold the look together.
Start with a clip and edit with a reference or a prompt. Change a character, camera move, voice or look without rebuilding the shot.
Separate audio stems
Export up to three audio stems to isolate and refine any facet of a clip's soundscape: dialogue, music and effects.
A flexible performer
Flex is built for a wide range of video types and genres, from product films to animated shorts.





Direct anything from orbits to dollies, cranes, tracking shots, whip pans, macro moves and changing lenses in natural language.
Make it move.Quality
Flex sits above base H3, H3Max and Seedance 2.5 on DesignArena's text-to-video board, and it is the highest ranked model with open weights.
DesignArena leaderboard snapshot · open weights highlightedFor active Pika H3 Flex members. Until November 15th.
Priced to play
Text, image and keyframe generation are charged per output-video second. Draft-to-video finishing charges only the difference between the draft and the finish.
Open weights
Pika H3 Flex is released with open weights and an inference implementation. Local inference will not necessarily match the optimised hosted API, but the model is yours to inspect, run and build on.
Open weights · research-friendly releaseWeights, licence and evaluation notes
Reference implementation and setup
A visual workflow for the model
Full model specs

FAQ
Everything else lives in the specs above. For billing, see the pricing page or your account settings.
Pika H3 Flex is a post-trained H3 video model built for professional iteration. Text, image, reference and edit workflows share one model, so a draft can become a finished shot without changing tools.
In its fastest configuration it can generate a 15 second clip in as little as 7 seconds. Across matched text, image and reference comparisons it averages 1.8x faster than H3Max.
No. Seven seconds is the best-case generation time for a 15 second clip. Queue depth, resolution, references and finishing mode all affect end-to-end latency.
Up to 12 assets in total: images, video clips, audio, sketches, moodboards or a written prompt. At least one image or video is required in reference mode.
Yes. Start from an existing video and add a reference or prompt to adjust a character, camera move, look or voice while keeping the rest of the take.
Yes. The request generates at 768p and upscales once inside the same job to 1080p, 2K or 4K, preserving the generated image content and audio.
When requested, every generation endpoint can return the video plus up to three WAV files for dialogue, music and effects.
The model weights and an inference implementation are published together. Local inference speed depends on your hardware and optimisation choices.
Use any Try Pika button on this page. You will be asked to sign in with Google, choose a paid plan and then enter the Studio queue.