Garment-true virtual try-ons
A freelance retoucher built a ComfyUI pipeline that maps real garment detail — stitching, drape, print alignment — onto model shots, holding fidelity that off-the-shelf try-on tools lose. Their e-comm client now gets full-catalog coverage from one base shoot.
This is the fidelity bar that makes AI try-ons client-sellable, and one operator runs it.
The how
- One base shoot per model — neutral garments, controlled lighting, ~20 poses.
- Each product garment gets a dedicated LoRA trained on 30–40 flat-lay and detail shots (stitching, buttons, print).
- ComfyUI graph composites garment onto pose via IPAdapter + ControlNet depth pass, preserving drape physics from the base shot.
- A final detail pass re-projects the original print at full resolution so patterns never warp.
- QA grid renders every garment × pose combination overnight; the retoucher approves ~80% untouched.
Run it back
You’d need a controlled base shoot, product flat-lays, a ComfyUI install with IPAdapter/ControlNet, and an afternoon per garment for LoRA training. The skill that matters is the retoucher’s eye — knowing which 20% to reject.