[Review Request] ComfyUI Wrapper for FASHN VTON v1.5

[drphero]

Hi @danbochman, thanks for offering to review this implementation of your code.

[danbochman]

Hi @drphero,

First off, thank you for putting this together.

I’m not a regular ComfyUI user, so I may miss some community best practices. From what I can see, the
node largely wraps our repository as‑is, which is a valid approach, but it does limit some
ComfyUI‑side flexibility. For example, users might want to swap the keypoint detector, change the
clothing parser, or customize sampling/CFG options. Also, the current GPU offload logic only moves the
TryOn and human parser models; the DWPose ONNX sessions are still on GPU, and the pipeline device is
set at init based on torch.cuda.is_available(), which can mismatch ComfyUI’s selected device.

If you prefer to keep this as a thin wrapper, that’s totally fine. If there’s interest in a more
ComfyUI‑native integration, we can help, and I’m happy to answer questions about the
internal components if you want to expand the node’s controls.

[drphero]

Thanks for the suggestions. Looks like I missed some settings. garment_photo_type , segmentation_free, and skip_cfg_last_n_steps along with tooltips have been added in 7a50015.

I was having trouble trying to move the ONNX sessions, and from my understanding ONNX sessions generally lock to a device upon initialization, so you cannot simply call .to(device) on them. I made some changes to the pipeline that has them initialized in the __call__ method in 88690bd that allows for them to be unloaded after inference (if the option is selected) and loaded once again when needed. The pipeline device is also now set by ComfyUI.

As for supporting different keypoint detectors or clothing parsers, are you aware of any others that work well with your model?

[danbochman]

Keypoints

Perhaps there are other already existing great implementations of DWPose ComfyUI nodes you can leverage.
What's custom about our use of DWPose implementation is:

1. We greedily only take the best and largest pose candidate
2. We transform the keypoints into a grayscale image instead of an RGB image

Parsing

The parser is relatively simple to replace, because it is only used in two cases:

1. When maskless mode is disabled (the clothing on the person must be masked). You don't HAVE to support this path as it is generally not recommended.

2. When a garment image is provided as worn by a person (we need to isolate the garment and mask everything else).

In both cases, using the FASHN Human Parser is optional. You can replace it with any segmentation method that works, as segmentation is not core to the VTON model’s performance or results.

It is included primarily as a convenience. It closely matches what we run in production and was straightforward to package and distribute via Hugging Face.

[drphero]

Thanks again for the review!

I've implemented a major cleanup and extension of the node set to support external pose and segmentation tools while keeping default behavior intact. This has been merged to main. Handling the external segmentation was tricky, but I think it's working now.

What was added

• Keypoint-based external pose path using internal FASHN/DWPose grayscale rendering.
• Fashn Pose Keypoints Adapter node to convert POSE_KEYPOINT payloads into internal DWPose-style keypoints.
• Fashn Mask to Labelmap node to convert a merged garment mask into valid FASHN labelmap IDs for external segmentation backend.
• Parser backend routing with fallback:
  • fashn_human_parser
  • external_fashn_labelmap (with fallback to internal when external input is missing/invalid).
• Improved UI tooltips across inference inputs to clarify conditional behavior.