H3 Singularity: HDR Fusion Fine-Tune for MiniMax H3 Video
WarmBloodAban releases Minimax-h3_Singularity, a fused and pruned H3 fine-tune for ComfyUI with HDR clarity, distant-face restoration, cleaner skin, stronger action motion and VFX.
ref, fl and b25-49 variants) into one ComfyUI-ready file, then deep fine-tunes and prunes the result for three days to remove the artifacts high-step training introduces. The team behind it runs the AIGC-Singularity YouTube and Bilibili channels.
Showcase clip from the release repository: fast action with audio, generated in ComfyUI.
What the fine-tune changes
The model supports all four H3 pipelines natively: text-to-video, image-to-video, reference-to-video and video-to-video. The team lists these targeted fixes over stock H3:
| Area | Change |
|---|---|
| Image quality | HDR-trained to cut motion blur during high-speed action |
| Faces | Restores distant faces that collapse or blur in medium and long shots |
| Skin | Removes the heavy glossy "over-oiled" look for more natural lighting and materials |
| Motion | Stronger dynamic motion for sword fights and martial arts melee |
| VFX | Tuned for fantasy spellcasting, particle auras and magical combat effects |
| Camera | Better response to pan, tilt, zoom and tracking moves |
The authors state the merge preserves H3's original prompt adherence and style range, so existing prompts and workflows keep working without re-tuning.
Files and setup
Two INT8 checkpoints ship in the release:
| File | Size | Notes |
|---|---|---|
Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8.safetensors | 19.6 GB | Pruned build, recommended |
Minimax-h3_Singularity_ref2va_v1.3_int8.safetensors | 31.7 GB | Full build |
Both load as ComfyUI diffusion models in the same slot as the stock H3 INT8 files. For fast iteration the team recommends pairing with the minimax_h3_ref2v_turbo_4step_v0.1 acceleration LoRA, which brings inference down to 4 steps with minimal quality loss. If you do not have a local GPU, a Ready-to-run version is available on RunningHub.
Availability
The weights are on Hugging Face under Apache 2.0, so the fusion itself carries no usage restriction beyond what the underlying H3 community license already grants for the base model's behavior. Workflows and the online demo are linked from the model card.
Comments
Sign in with GitHub to join the discussion.