1.8 KiB
1.8 KiB
MiniMax H3 INT8 R2V
Summary
Same-day MiniMax-H3 reference-to-video workflow from javawock7618. This is the MiniMax drop to watch if you care most about keeping a subject stable from a reference image while still getting motion and camera movement.
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/resolve/main/MiniMax_int8_R2V-javano2608.12.json?download=true
- Original publisher:
javawock7618 - Date imported:
2026-08-11
Developer notes
- The source README positions this as the reference-guided workflow in the pack.
- It is built around INT8 inference plus the
ComfyUI-H3-Motion-ContextandComfyUI-Spectrum-MiniMax-H3ecosystem. - The pack also recommends
ComfyUI-MiniMaxH3_LatentUpscalerfor post-upscale. - The R2V lane requires
minimax_h3_ref_lora_rank_256_bf16.safetensorsfromKijai/MiniMax-H3-experimental.
Our notes
- This is the most directly useful MiniMax reference-control item from today.
- On a 3070 it is still heavy, but it is the kind of workflow that could actually be worth testing if you want stronger subject hold than a generic video model.
- Better fit than a novelty demo because it is explicit about the reference LoRA and the supporting nodes.
Required custom nodes
ComfyUI-Spectrum-MiniMax-H3ComfyUI-H3-Motion-ContextComfyUI-MiniMaxH3_LatentUpscaler
Required models
Comfy-Org/MiniMax-H3minimax_h3_ref_lora_rank_256_bf16.safetensorsfromKijai/MiniMax-H3-experimental
Notes
- Lightx2v LoRA and Turbo-LoRA are optional, but the source pack treats them as useful accelerators.
- This is the MiniMax workflow I would try before the more novelty-leaning releases.