Description
The painful part of AI video isn't one render, it's re-rendering after every small change: a different camera move, a reworded line, another long sampling run. Pika H3 Flex goes after that iteration cost by sampling in only six steps, and what comes out is a finished video with synchronized stereo audio, no separate soundtrack pass.
It is a set of LoRA adapters Pika post-trained on top of MiniMax-H3, with weights on Hugging Face. Write a prompt, add up to 12 reference files, and you can direct shots, camera moves and character consistency, which suits multi-shot stories, motion graphics and product films that need repeated polishing.
Six-step sampling: One generation runs six generator evaluations and produces 24 fps H.264 video with 32 kHz stereo audio, muxed into an MP4.
Several ways in: Text only, or pin the opening frame, the ending frame or both. A multimodal reference mode takes up to 9 images, 3 videos and 3 audio clips, 12 files in total.
Length and framing: Native local output is 480p or 768p, 5 to 15 seconds, across 11 aspect ratios from 21:9 to 9:21.
Camera work in the prompt: In the official examples one prompt lays out overhead, crane-down, whip-pan and orbit moves by time range, and the model follows them. Dialogue, ambience and music can be described in the same prompt.
Ways to run it: An official CLI package with pip, Conda and Docker paths, ComfyUI workflows for text-to-video, first/last frame and reference-to-video, plus a Diffusers pipeline.
Hosted API for higher quality: 1080p, 2K and 4K output, audio stems and prompt enhancement come from Pika's hosted API, not the local weights.
The local bar is high. One request uses one GPU; Pika tested on an H100 80GB, where default 480p/5s sampling takes roughly 66 to 69 GiB of VRAM and 768p/15s needs CPU offload. Base components total about 210 GB. It ships under the MiniMax H3 Community License, so read the terms before commercial use.
It is a set of LoRA adapters Pika post-trained on top of MiniMax-H3, with weights on Hugging Face. Write a prompt, add up to 12 reference files, and you can direct shots, camera moves and character consistency, which suits multi-shot stories, motion graphics and product films that need repeated polishing.
Features
Six-step sampling: One generation runs six generator evaluations and produces 24 fps H.264 video with 32 kHz stereo audio, muxed into an MP4.
Several ways in: Text only, or pin the opening frame, the ending frame or both. A multimodal reference mode takes up to 9 images, 3 videos and 3 audio clips, 12 files in total.
Length and framing: Native local output is 480p or 768p, 5 to 15 seconds, across 11 aspect ratios from 21:9 to 9:21.
Camera work in the prompt: In the official examples one prompt lays out overhead, crane-down, whip-pan and orbit moves by time range, and the model follows them. Dialogue, ambience and music can be described in the same prompt.
Ways to run it: An official CLI package with pip, Conda and Docker paths, ComfyUI workflows for text-to-video, first/last frame and reference-to-video, plus a Diffusers pipeline.
Hosted API for higher quality: 1080p, 2K and 4K output, audio stems and prompt enhancement come from Pika's hosted API, not the local weights.
The local bar is high. One request uses one GPU; Pika tested on an H100 80GB, where default 480p/5s sampling takes roughly 66 to 69 GiB of VRAM and 768p/15s needs CPU offload. Base components total about 210 GB. It ships under the MiniMax H3 Community License, so read the terms before commercial use.
