Files

1.9 KiB
Raw Permalink Blame History

MiniMax H3 LongVideos candidate

This is the next R2V candidate after rejecting connected-chunk continuation. It generates the clip directly from the character reference and a paragraph-driven shot plan; it does not load or encode a previous video.

Fixed baseline

  • Reference: reference.png
  • Prompt: two-shot bright park scene: robins in shot one, then a squirrel joins in shot two
  • Resolution target: 960×544 landscape (0.55 megapixels, 16:9)
  • Duration: 10 seconds
  • Steps: 8
  • Sampler: res_multistep
  • Scheduler: simple
  • Seed: 20260927
  • Shift video/audio: 12 / 3
  • Reference noise augmentation: 0.999
  • SLA Attention: sparsity 0.85, block 32
  • Model Attention Backend: Comfy Kitchen attention
  • Low-VRAM Attention: 4 head chunks
  • Chunk FeedForward: 2 chunks, sequence threshold 4096
  • Turbo LoRA: enabled at strength 1.0
  • Latent and post upscaling: disabled for this baseline

Files

  • h3-longvideos-api.json — direct API workflow
  • h3-longvideos-ui.json — ComfyUI web-editor workflow
  • reference.png — fixed character reference
  • prompt.txt — shared reference-to-video prompt source

The live server must also provide the H3-compatible 2560-feature text encoder qwen3vl_32b_h3_ultra_uncensored_heretic_int8_convrot.safetensors. The currently listed 5120-feature qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors is not compatible with H3LongVideos.

Why this candidate

The installed H3LongVideos node supports paragraph-driven shot chaining, synchronized audio, identity references, and later latent/RTX upscale paths. It is the appropriate next workflow for multi-shot generation without loading the previous clip into VRAM.

The imported source workflow remains at models/minimax/minimax-h3-longvideos-smite79/workflow.json. This optimized graph now retains the proven memory/attention chain and explicitly wires the scene anchor and character memory fields.