Can't this be used in Comfyui and what about GGUF or int8 model versions?

#3
by ryg81 - opened

Can't this be used in Comfyui and what about GGUF or int8 model versions?

JD.com Open Source org

Thanks for the interest!

ComfyUI — no official integration right now. JoyAI-Video-Edit is a streaming editor: the server keeps rolling KV state across chunks and edits frames as they arrive over a WebSocket, which doesn't map cleanly onto ComfyUI's stateless graph execution. A fixed-length clip→clip node on top of our pipeline is technically doable, but the cleaner path is the 🤗 Diffusers pipeline on our README TODO — once that lands, building ComfyUI nodes on top becomes much easier. Community ports are very welcome.

GGUF / int8 — no GGUF release planned: it's a custom 16B MMDiT with a custom streaming runtime, so a GGUF file alone wouldn't be runnable anywhere today. The quantization you'd actually want is already in the serving stack, applied at load time:

  • all DiT Linears run FP8 by default (per-channel weight + per-token activation quantization, cutlass kernels in deploy/joyomni_ops);
  • INT8 attention on GeForce via SageAttention (JOYOMNI_SAGE_ATTN=1);
  • JOYOMNI_LOW_VRAM=1 stages the model to the GPU block-by-block as FP8, so it fits a single RTX 5090 (32 GB) at 840×480 @ 24 FPS.

So there's no separate quantized checkpoint to download — grab the bf16 weights and the server quantizes at startup. Full setup: DEPLOYMENT.md. The local web UI also accepts uploaded clips, not just a webcam. And for a zero-setup try-out there's the live demo: https://huggingface.co/spaces/wxDai/joyai-video-edit

Sign up or log in to comment