Mickmumpitz's one-model H3 short film and Singularity V2

The one-model pipeline. In his video posted this morning (8:30 AM MDT; NVIDIA sponsored it), Mickmumpitz builds 'SHIBUWAKA', a short about two would-be daredevils racing through rural Japan, using only the reference version of MiniMax H3 in ComfyUI. Every workflow shares one layout, so learning one teaches you the rest. There's a character-sheet generator (which he also uses for environments and car asset sheets), a turnaround workflow that saves a video plus an 8-view sheet, and H3 used as a reference-based image editor for start/end frames and style frames. A voice workflow clones from a short reference clip; he recommends turbo off there and on for video iteration. The shot workflow takes picture references by number, an audio reference or a locked voice line, start/end frames, and video references. That covers acting a scene out yourself, a storyboard, or a Blender camera layout for a move he couldn't get from prompts.

His honest notes. H3-generated music was 'hit or miss, and usually miss', so the film uses stock music; sound effects were decent. His high-res pass re-renders each shot with the same prompt and references, and he puts it at about 15 minutes per shot on an RTX 5080. The workflow templates are in a public Patreon post; the project's example files are for paying supporters. Clone only voices you have permission to use.

Singularity V2. WarmBloodAban's community fine-tune of H3 (the repo has about 900 likes and 900K downloads on Hugging Face) updated to V2 last week. Per the model card, V2 is trained on more action and varied motion, follows prompts more closely, fixes audio-reference conditioning, improves photoreal, 3D CG and cinematic CG styles, and ships as INT8. The full file is about 34 GB, and Abiray has posted pruned builds. ComfyUI Workflow Blog's tutorial today shows a 27-second continuous camera shot and a 30-second rooftop fight made from three reference images and the H3 Fight Flow LoRA, stitched from three 10-second parts. Its two-pass workflow pairs the LightX2V turbo LoRA with optional scene LoRAs (GunFu, SMACK, Motion Repair, Fight Flow), with the warning not to stack them all. The creator says some runs repeated hits or missed dodges, and that some Reddit users still prefer V1.3 for calm character shots. The model card lists Apache-2.0, but it's a fine-tune of H3, so check H3's community licence too.

One more speed note. Noncoder's update to the RefMod + OBVPM timeline workflow adds an 8-step turbo preset that, per the creator, cut a 10-second take from 372 to 161 seconds on his machine.

Our take: Mickmumpitz's template set is the cleanest free starting point we've seen for a whole H3 short, from character sheet to final shot, and it shows where H3 still falls short (music). Use Singularity V2 for action beats and keep the base model or V1.3 for quiet dialogue shots.