# Pika H3 Flex adapters for ComfyUI

These BF16 LoRA adapters use the native MiniMax-H3 module layout for ComfyUI. They are separate from the FP32 adapters used by the Pika inference package.

| File | Use with |
| --- | --- |
| `loras/pika_h3_flex_fl2va_comfy_bf16.safetensors` | The matching MiniMax-H3 FL2VA model for text and first/last-frame conditioning |
| `loras/pika_h3_flex_ref2va_comfy_bf16.safetensors` | The matching MiniMax-H3 Ref2VA model for multimodal reference conditioning |

## Download

Install `huggingface_hub` and set `HF_TOKEN` if repository access requires authentication. Resolve the current revision once, then use that immutable revision for the download:

```bash
python -m pip install huggingface_hub
export COMFY_REVISION=$(python -c 'from huggingface_hub import HfApi; print(HfApi().model_info("Pika-Labs/Pika-H3-Flex").sha)')
hf download Pika-Labs/Pika-H3-Flex \
  --revision "$COMFY_REVISION" \
  --include 'comfy/*' LICENSE NOTICE 'licenses/*' \
  --local-dir ./pika-h3-flex-comfy
```

Record `COMFY_REVISION` to reproduce the same download later. Verify the downloaded adapters from inside `pika-h3-flex-comfy/comfy` with `sha256sum -c SHA256SUMS`.

Place the two files from `comfy/loras/` in your ComfyUI `models/loras/` directory. Use a workflow with native MiniMax-H3 support and the matching FL2VA or Ref2VA base model. The files identify `comfyui-minimax-h3` as their target integration; this directory does not bundle that integration, base weights or a workflow.

## Tutorial and workflows

Follow the [Comfy tutorial](https://docs.comfy.org/tutorials/video/pika/pika-h3-flex), then choose the [R2V workflow](https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_pika_h3_flex_ref2va.json), [FL2VA workflow](https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_pika_h3_flex_fl2va.json) or [T2V workflow](https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_pika_h3_flex_t2v.json).

## Format and sampling

Conversion renames the modules, fuses Q/K/V LoRA projections and reorders the SwiGLU halves for the native model layout. Ordinary LoRA modules have rank 128; fused Q/K/V modules have rank 384. The conversion preserves unit scaling and omits explicit alpha tensors. Use LoRA model strength 1.0 to apply the intended adapter scale.

Use a Pika H3 Flex-compatible six-NFE workflow. The original sampling contract uses knots `1000/750/500/250/100/40/0`, video shift 12 and audio shift 3, with fresh intermediate noise. Simply choosing six steps in a generic sampler does not reproduce that contract. See the original [FL2VA manifest](../fl2va/manifest.json) and [Ref2VA manifest](../ref2va/manifest.json) for the release configuration.

The tensors were checked against the corresponding original released adapters after conversion and BF16 casting. File integrity and tensor conversion were validated; an end-to-end ComfyUI render has not been verified as part of this upload. Outputs can differ from the original inference runtime because of precision, sampling, conditioning and implementation differences.

The existing Pika inference package continues to use `fl2va/adapter.safetensors` and `ref2va/adapter.safetensors` at the revision pinned by the official inference repository's `release.json`. Do not substitute these Comfy-format files into those locations.

## License and notices

These adapters are Model Derivatives of MiniMax H3 and are distributed under the [MiniMax H3 Community License Agreement](../LICENSE). See [NOTICE](../NOTICE) and the [upstream agreement](../licenses/MiniMax-H3-LICENSE.txt). Keep these legal files with redistributed adapters.

Modified by Mellis, Inc. in 2026. Copyright © 2026 Mellis, Inc.
Model Derivative of MiniMax H3, distributed under the MiniMax H3 Community License Agreement.