Kling Virtual Try-On

E-commerce got here first, but the use case is universal: you have a photo of a person and a photo of a garment, and you want a believable image of the person wearing that garment - seams, folds, fabric behaving like fabric. KlingDirect_VirtualTryOn does it in one API call using Kolors, the open image model Kuaishou trained (yes, Kling's parent) for exactly this clothing-transfer job. Kling runs it server-side, so you never download a Kolors checkpoint or fight its dependency stack.

For context, Kolors the open model is one of the "forgotten" releases of 2024 - real, capable, eclipsed by Flux for general image work. But virtual try-on is the niche where it stayed relevant, and Kling exposing it as a hosted endpoint is a reasonable deal: no local install, no 13GB checkpoint, one node.

What you set

Just three inputs, which is the appeal:

• human_image (IMAGE) - the person.
• cloth_image (IMAGE) - the garment, as clean a shot as you can get (a flat product photo works better than a crumpled one).
• model_name - kolors-virtual-try-on-v1, the only choice. It's a dropdown with one entry, so this input is essentially documentation.

No prompt, no sizing, no pose controls. The model figures out the drape and the fit. That's both the strength (it's genuinely one-shot simple) and the ceiling (you don't get to steer much).

Outputs

• image (IMAGE) - the try-on result, ready to preview, save, or feed onward.
• url - the hosted result.
• task_id - for status chaining.

Install and what it needs

Part of ComfyUI-Kling-Direct: ComfyUI Manager → search "Kling Direct" → install, or git clone https://github.com/IxMxAMAR/ComfyUI-Kling-Direct into ComfyUI/custom_nodes, then restart. No model downloads - Kolors lives on Kling's servers. You need Kling keys from https://app.klingai.com/global/dev (KYC required) in the Kling AI Authentication node or as env vars, with the correct region.

Where people get burned

• Clothing photo quality. The single biggest quality lever. A garment shot crumpled, angled, or on a hanger transfers worse than a flat, evenly lit product shot. The model faithfully transfers what it sees - garbage cloth in, garbage drape out.
• Body mismatch. A heavily posed human photo (extreme angles, cropped limbs) will fight the clothing transfer. A neutral standing pose is the model's comfort zone.
• It's a paid image job. Each call is metered against credits. Draft on one garment before batch-processing a catalog.
• Expecting edit-level control. You can't ask it to re-fit the sleeves or change the color mid-transfer. If you need that, you're looking at an image-edit model, not try-on.

The honest take: for the one job it does, this node is about as frictionless as it gets - two images in, one output, no prompt engineering. That's rare in this pack and worth leaning on when the job fits.