# MiniMax H3 INT8 R2V Javawock Pack

## Summary

Same-day `MiniMax H3` reference-to-video workflow from the updated `javawock7618` pack. This is the newer reference-retention graph to keep alongside the older `2608` import.

## Model Family

- `minimax`

## Status

- `imported-only`

## Source

- Workflow pack page: <https://huggingface.co/javawock7618/comfy-MiniMax-H3-workflows>
- Workflow file: <https://huggingface.co/javawock7618/comfy-MiniMax-H3-workflows/raw/main/MiniMax_int8_R2V-javano2609.1.1.json>
- Original publisher: `javawock7618`
- Date imported: `2026-09-03`

## Developer notes

- The pack README describes this as the `Reference-to-Video` workflow.
- It is built around INT8 inference, Sol-Attn, Spectrum, Lightx2v LoRA support, optional Turbo-LoRA, native Extend support, and latent upscaling.
- The workflow uses the pack's dedicated reference LoRA path for appearance retention and also carries Fun ControlNet Union support for V2V control-video guidance.

## Our notes

- Best new MiniMax H3 import today if you care about reference retention more than flashy motion.
- Still heavy on an `RTX 3070`, but the INT8 path is exactly why it is worth keeping.
- Good comparison point against the current LTX reference-control and Ingredients-style consistency workflows.

## Required custom nodes

- `ComfyUI-SolAttn_triton`
- `ComfyUI-Spectrum-MiniMax-H3`
- `ComfyUI-H3-SLA-Attention`
- `ComfyUI-H3-FaceRefine`
- `ComfyUI MiniMax-H3 Latent Upscaler`
- `ComfyUI-Switchboard`

## Required models

- `Comfy-Org/MiniMax-H3`
- `minimax_h3_ref_lora_rank_256_bf16.safetensors`
- `Kijai/MiniMax-H3-experimental` for the Fun ControlNet Union path
- `lightx2v/Minimax-h3-Turbo` if acceleration is enabled
- `drbaph/MiniMax-H3-Turbo-Lora-ComfyUI` if the optional turbo LoRA path is enabled
- `LBH-123-AI/Minimax_h3_latent_Upscaler` if latent upscaling is enabled

## Notes

- This is the newer same-day R2V graph from the pack, not just a duplicate label.