gguf quantized version of mochi (incl. gguf encoder and gguf vae)🐷🐷🐷

  • run it straight with gguf-connector
  • opt a gguf file in the current directory to interact with by:
ggc m1

GGUF file(s) available. Select which one to use:

  1. pig-mochi-q3_k_m.gguf
  2. pig-mochi-q4_1.gguf
  3. pig-mochi-q8_0.gguf

Enter your choice (1 to 3): _

run it with gguf-node via comfyui

  • drag mochi to > ./ComfyUI/models/diffusion_models
  • drag t5xxl to > ./ComfyUI/models/text_encoders
  • drag vae to > ./ComfyUI/models/vae
  • drag demo video (below) to > your browser for workflow

review

  • new tensor fixed version; load faster with full set gguf (model + encoder + decoder)
  • upgraded encoder; from fp16/fp8 to fp32; but not affecting the file size and memory consumption; more compatible to old machine
  • new fp32 gguf vae decoder; similar size to fp16 safetensors; better quality; less ram requirement
  • q2 works but not usable; you could get q8 [here] pig architecture

reference