Sparse text-to-audio-video

Veda
MiniMax-H3

A learned attention predictor that selects the most useful 10% of tiles per head while preserving the MiniMax-H3 backbone, motion, visual style, and native audio path.

Free showcase mode

Watch the authors' outputs and prepare a text-only, first-frame, or Foodie multi-reference request in your browser. Your files stay on this device. Live inference requires dedicated GPU hardware and persistent model storage; it is not running in this Static Space.

90%attention tiles skipped
6.8×reported attention speedup
3.1×reported end-to-end speedup

Official comparisons

Dense and Veda renders use the same prompt, seed, and Turbo LoRA. The titles inside each video show the generation-time comparison.

Dense vs Veda · official checkpoint sample 01
Dense vs Veda · official checkpoint sample 03
Dense vs Veda · official checkpoint sample 04
Veda-only · official checkpoint sample 148

Build an H3 request

Compose a structured prompt, choose the reference mode, then download a role-preserving JSONL manifest.

Selected KOL character preview
Selected storyboard preview

Foodie keeps the roles separate: the KOL image owns visible identity and appearance, the storyboard owns composition and action, and the audio owns voice timbre only. Spoken narration still comes from the prompt's dialogue tags. The downloaded role manifest targets the DBee/ComfyUI Foodie handoff; Miowtion's current encoder CLI only accepts T2VA and FL2VA manifests. 


          

        

      


      

        

          
Deployment reality

          
The preview is a 33B audio-video DiT. Sparse attention saves compute, not the need to host the backbone.

        

        

          

            
Model storage

            
Roughly 70 GB for MiniMax-H3 plus the Veda predictor and Turbo LoRA. Persistent storage avoids downloading weights after every restart.

          

          

            
Runtime memory

            
Miowtion documents about 40 GB of free pinned host RAM per process. A 24 GB GPU can run with all transformer blocks offloaded.

          

          

            
Recommended GPU

            
H100/H200 avoids block streaming. Compatible Ada and Blackwell GPUs can use offloading when sufficient host RAM is available.