Original title: Pika-Labs/Pika-H3-Flex
A set of LoRA adapters post-trained by Pika on MiniMax-H3, generating videos with synchronized stereo audio; officially claimed to produce results in just six sampling steps. It consists of two adapters: FL2VA handles text-to-video and control over first, last, and first-last frames; Ref2VA handles multimodal references, supporting up to 9 images, 3 video clips, and 3 audio clips, totaling 12 reference files. Local open-source inference natively outputs 480p / 768p resolution at 24 fps for 5–15 seconds, supporting 11 aspect ratios; 1080p to 4K resolution, separated audio tracks, and prompt enhancement are only available via Pika's hosted API.
Why it matters Open weights allow local execution, with official ComfyUI workflows available, making it suitable for multi-shot short films, product videos, and animation based on reference images
Rising score 151 · Likes 528 · Downloads 1083
Reasons
- A single model generates both visuals and audio (ambient sounds, dialogue, music), eliminating the need for separate audio models
- Two control methods: first-last frame and multi-reference; listed use cases include multi-shot storytelling, motion graphics, product videos, and animation based on reference images
- ComfyUI already has official tutorials and three workflow templates for R2V, FL2VA, and T2V; the repository includes BF16 LoRAs in Comfy format
Usage
- Clone the GitHub repository Pika-Labs/Pika-H3-Flex, install via pip / Conda / Docker quick start, and use the pika-h3-pro command to select t2va, i2va, l2va, fl2va, or ref2va tasks
- ComfyUI: Use the LoRAs in comfy/loras/ with the MiniMax-H3 base model, applying the official Pika H3 Flex workflow templates
- Write prompts following the official guide, specifying camera movements, actions, dialogue, and sounds by time segment
Risks
- Adapters are released under the MiniMax H3 community license; review terms before commercial use
- Base model components total approximately 210 GB; official tests are only provided for H100 80GB, requiring CPU offload for large requests
- Demo videos are creative samples; the official statement does not guarantee exact reproduction using the released package
- This entry has not been run; refer to the model card for performance details