51 lines
1.7 KiB
Markdown
51 lines
1.7 KiB
Markdown
# MiniMax H3 INT8 Reference-to-Video
|
|
|
|
## Summary
|
|
|
|
`MiniMax-H3` reference-to-video workflow with `INT8`, Motion Context, a dedicated reference LoRA path, and latent upscaling. This is the most practical same-day MiniMax consistency-oriented update.
|
|
|
|
## Model Family
|
|
|
|
- `minimax`
|
|
|
|
## Status
|
|
|
|
- `imported-only`
|
|
|
|
## Source
|
|
|
|
- Workflow repo: <https://huggingface.co/javawock7618/comfy-MiniMax-H3-workflows>
|
|
- Workflow file: <https://huggingface.co/javawock7618/comfy-MiniMax-H3-workflows/resolve/main/MiniMax_int8_R2V-javano2608.14.json?download=true>
|
|
- Original publisher: `javawock7618`
|
|
- Date imported: `2026-08-13`
|
|
|
|
## Developer notes
|
|
|
|
- This is the `R2V` branch from the source pack.
|
|
- The source README describes the pack as using `Sage-Attention`, `Spectrum`, `Lightx2v LoRA`, experimental `Turbo-LoRA`, `Motion Context`, and latent upscaling.
|
|
- The `R2V` branch expects `minimax_h3_ref_lora_rank_256_bf16.safetensors`.
|
|
|
|
## Our notes
|
|
|
|
- This is the strongest today-only `MiniMax-H3` workflow for actual consistency and reference control.
|
|
- It is still heavy for an `RTX 3070 / 8 GB`, but it is much more interesting than a novelty-only turbo drop.
|
|
- If you want one MiniMax workflow to test against `LTX 2.3`, this is the one.
|
|
|
|
## Required custom nodes
|
|
|
|
- `ComfyUI-Spectrum-MiniMax-H3`
|
|
- `ComfyUI-H3-Motion-Context`
|
|
- `ComfyUI-MiniMaxH3_LatentUpscaler`
|
|
|
|
## Required model files
|
|
|
|
- `Comfy-Org/MiniMax-H3`
|
|
- `minimax_h3_ref_lora_rank_256_bf16.safetensors` from `Kijai/MiniMax-H3-experimental`
|
|
- optional `Kijai/MiniMax-H3_comfy`
|
|
- optional `drbaph/MiniMax-H3-Turbo-Lora-ComfyUI`
|
|
|
|
## Notes
|
|
|
|
- This is the better `MiniMax-H3` pick when you care about reference fidelity, not just speed.
|
|
- The pack also ships FLF and TTS workflows, but this import is the dedicated R2V graph.
|