Pika H3 LoRA & Adapters
Pika H3 ships two task-specific adapters on top of MiniMax H3: FL2VA for frame-interpolation and text-to-video, and Ref2VA for reference-conditioned generation. People call them LoRAs informally; technically they are adapter heads with their own safetensors. Updated November 2026.
What each adapter does
| Adapter | Accepts | Best for |
|---|---|---|
| FL2VA | Text, first frame, last frame | Concept shots, interpolation, title cards |
| Ref2VA | Up to 12 refs: images, clips, audio | Identity preservation, product, music-driven |
Download
Both safetensors are at huggingface.co/Pika-Labs/Pika-H3-Flex. Place them under ComfyUI/models/diffusion_models/.
Minimum workflow (5 nodes)
- Checkpoint Loader (base MiniMax H3 weights)
- Adapter Loader (FL2VA or Ref2VA)
- CLIP Text Encoder
- KSampler (sampler LCM, steps 6)
- VAE Decoder
For Ref2VA, insert a Reference Loader node between the Adapter Loader and the sampler, pointing at up to twelve reference files.
FAQ
Does Pika H3 use LoRA?
Not in the classical rank-decomposition sense. The release ships two task-specific adapter modules — FL2VA and Ref2VA — that attach to the MiniMax H3 base. They are often called LoRAs informally because they are swappable modules.
Can I train my own adapter?
The open-weights license permits fine-tuning. Pika Labs has not published a training recipe for FL2VA/Ref2VA as of November 2026, so community training is still exploratory.
Where are the weights?
huggingface.co/Pika-Labs/Pika-H3-Flex. Both adapter safetensors are published alongside the base weights.