Local anime video gen: 20s clip with dialogue in ~4 min on RTX 4070 Super 12GB via ComfyUI + MiniMax H3

Raw_Emotion_ · reddit · 2026-11-12

A Redditor generated a 20-second anime video (481 frames, 24fps, 960×544) with dialogue, multiple camera angles and character movement in 200 seconds locally, using only an RTX 4070 Super 12GB, i5-14600K, and 16GB DDR5 RAM.
Workflow: MiniMax H3 MATLOW fused INT8 ConvRot model, 4-step resmultistep sampler, Comfy Kitchen Attention, BlockSparseAttention (SLA, 20% keep), H3 Chunk FeedForward (8 chunks), sigma shift 12/3, all in ComfyUI.
Tests included Death Note's Light and Ryuk plus two original characters walking home discussing a science experiment; dialogue, camera cuts, over-the-shoulder shots and background continuity held up well from a single reference image and prompt.
Weak spots: hands/fingers, lip-sync, some jarring transitions. Real-ESRGAN upscaling to 1080p pushed total time to 10 minutes.
Next steps: character sheets for consistency and VideoRef for motion transfer; author is soliciting tips on hand anatomy and temporal consistency.

More from Multimodal