H3 Singularity: HDR Fusion Fine-Tune for MiniMax H3 Video

ComfyUI Wikinews

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.

Minimax-h3_Singularity (Hugging Face) is a community fusion fine-tune of MiniMax H3 from the AIGC-Singularity team. It merges several H3 checkpoints (including 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:

AreaChange
Image qualityHDR-trained to cut motion blur during high-speed action
FacesRestores distant faces that collapse or blur in medium and long shots
SkinRemoves the heavy glossy "over-oiled" look for more natural lighting and materials
MotionStronger dynamic motion for sword fights and martial arts melee
VFXTuned for fantasy spellcasting, particle auras and magical combat effects
CameraBetter 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:

FileSizeNotes
Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8.safetensors19.6 GBPruned build, recommended
Minimax-h3_Singularity_ref2va_v1.3_int8.safetensors31.7 GBFull 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.

Loading comments…
H3 Singularity: HDR Fusion Fine-Tune for MiniMax H3 Video | ComfyUI Wiki