ComfyUI-Leffa

ComfyUI-Leffa

Full-pipeline virtual try-on and pose transfer ComfyUI custom node, built on Meta's Leffa (MIT). Powers the mcp.baisoln.com/leffa/* API.

Unlike StartHua/Comfyui_leffa — which only runs the diffusion step and requires the caller to pre-compute a DensePose image and an agnostic mask — these nodes run the entire pipeline inside ComfyUI (human parsing → openpose → agnostic mask → densepose → Leffa diffusion), mirroring upstream LeffaPredictor.leffa_predict, and they support pose transfer.

Nodes

| Node | Purpose | |------|---------| | Leffa Loader (LeffaLoader) | Loads one task model: viton_hd, viton_dc, or pose_transfer. | | Leffa Virtual Try-On (LeffaTryOn) | person + garment IMAGE → tried-on image. garment_type ∈ {upper_body, lower_body, dresses}. | | Leffa Pose Transfer (LeffaPoseTransfer) | person + target_pose IMAGE → reposed image. |

A typical graph: LoadImage(person) + LoadImage(garment) → LeffaLoader → LeffaTryOn → SaveImage.

Models

Weights are expected under ComfyUI/models/Leffa/ (the whole franciszzj/Leffa HF repo, ~34 GB):

hf download franciszzj/Leffa --local-dir ComfyUI/models/Leffa

Subtree used: virtual_tryon.pth, virtual_tryon_dc.pth, pose_transfer.pth, stable-diffusion-inpainting/ (+ SDXL inpaint configs), densepose/, humanparsing/, schp/, openpose/.

In the cluster these are bind-mounted from /mnt/ai-models/comfyui/models/Leffa.

Notes

• The vendored detectron2/densepose route ROIAlign/NMS through torchvision.ops, so no detectron2 compilation is required for the densepose_rcnn_R_50_FPN_s1x model used here.
• Human parsing runs on CPU via onnxruntime; openpose/densepose/diffusion on GPU.
• Only one diffusion model is kept resident at a time (VRAM bound for the RTX 5090); switching task evicts the previous one.
• torch/torchvision/transformers come from the ComfyUI env and must not be upgraded; see requirements.txt.

License

MIT (inherits franciszzj/Leffa). Vendored upstream source under leffa/, leffa_utils/, preprocess/, detectron2/, densepose/, SCHP/.