MiniMax-H3 DenoMix — R2V pruned INT8 v3
A MiniMax H3 reference-to-video (R2V) diffusion model. It is built on the original MiniMax H3 R2V weights and was fine-tuned by merging the Singularity fine-tune together with a few additional weights into a single file.
Powered by MiniMax H3.
License
This model follows the MiniMax H3 Community License Agreement, since it is built on the original MiniMax H3 weights (modified by DENO, October 2026 — see NOTICE).
File
| File | MiniMax-H3_DenoMix_ref2va_pruned_int8_v3.safetensors |
| Size | 20,967,647,432 bytes |
| SHA-256 | 6fc735bb1d2e9e44951106b3ec6a4c063df1535c567bc7d44aaff4c9052add52 |
| Format | pruned INT8 (ConvRot) diffusion model, one file |
Use in ComfyUI
- Put the file in
ComfyUI/models/diffusion_models/. - In your MiniMax H3 R2V workflow, select it in Load Diffusion Model (
UNETLoader) in place of the MiniMax H3 R2V pruned INT8 diffusion model. The text encoder and VAEs stay the same.
A ComfyUI version that supports the MiniMax H3 pruned INT8 (ConvRot) format is required.
ComfyUI workflow
workflows/Deno_MiniMax_H3_R2V_Turbo_8plus2.json — a MiniMax H3 R2V workflow built around this model:
it first generates at a low resolution with the Alibaba turbo LoRA (8 steps), upscales the latent ×2, and then refines with the DMD turbo LoRA and NVIDIA's refiner (Kijai's light extraction) over the last 2 steps.
The workflow's note lists every model file it uses with its download link. Drag the file onto the ComfyUI canvas to open it.
Workflow license: GPL-3.0-only. The models and custom nodes it loads keep their own licenses.
Credits
- MiniMax H3 by MiniMax — base model.
- Singularity by WarmBloodAban (Apache-2.0) — merged into this model.
- Downloads last month
- 25,963