MiniMax H3 Singularity V2: a community fine-tune against plastic skin and floating slow motion
Minimax-h3_Singularity V2 is an open community fine-tune of MiniMax H3 for ComfyUI that aims at H3's shiny skin, blurry distant faces and slow, floating motion. V2, released in INT8, adds wider action coverage, closer prompt following and a fix for audio reference conditioning.
What it does
It is a fusion of several H3 checkpoints with long fine-tuning on HDR video, and it supports text-to-video, image-to-video, reference-to-video and video-to-video in ComfyUI. The card says it reduces motion blur in fast action, face collapse in medium and long shots, and glossy skin.
Results and limits
Veteran AI's tutorial pairs it with the SelfLift node (a low-res pass, 3D latent upscale and high-res refinement in one sampler, at lowres_scale 0.4, transition_step 4 and rho 0 for 1376×768) and a YOLOv8 face pass with VOSR 2.0 at 1.0x. He tests it on fast moves like sprints and sword duels without speed-correcting LoRAs.
Worth knowing
The card recommends naming the style in the prompt (photorealistic, East Asian realism, 3D CG or cinematic CG) and keeping prompts under about 5,000 characters. The card lists Apache 2.0, while MiniMax H3 itself ships under the MiniMax H3 Community License, so check both before commercial use. The tutorial's workflows are free on GitHub and also run online on RunningHub.
Try it yourself
photorealistic, natural lighting, realistic textures. [Your character] sprints across [location] and grabs [object] in one fast move, handheld camera tracking alongside.