tryon-FLUX.2-klein-4B-lora

Virtual Try-On — FLUX.2-klein-4B LoRA

A rank-32 edit LoRA for FLUX.2-klein-4B that dresses a person in a specified outfit. 3 ordered inputs → 1 output:

1. person — a full-body photo with the clothing region masked out (clothing-agnostic)
2. top — the upper-body garment (isolated, product-style)
3. bottom — the lower-body garment (isolated); for a one-piece dress, use the dress in both slots

Output: the person wearing the specified top + bottom, with identity, pose, and background preserved.

Usage

• Base: FLUX.2-klein-4B (image edit / multi-reference).
• Trigger / prompt: TRYON a full-body photo of a person. Replace the outfit with the top and bottom as shown in the reference images. The final image is a full body shot.
• Reference order (load-bearing): person, top, bottom.
• Note: the person input should be clothing-agnostic (the outfit region greyed/masked), matching how the model was trained — agnostic-ify the input before inference.

Training

Self-trained with ai-toolkit (v0.10.24) on black-forest-labs/FLUX.2-klein-base-4B (Apache-2.0). Rank 32 / alpha 32, adamw8bit, lr 1e-4, flowmatch, qfloat8, 3000 steps — this is the step-2000 checkpoint (best generalization in reconstruction + outfit-swap eval). Dataset: 273 full-body fashion tuples, each derived from a single photo via clothing-agnostic masking + garment-part segmentation (mattmdjaga/segformer_b2_clothes) + Qwen2.5-VL captioning. Source frames are Pexels (train-only) and are NOT redistributed — only the trained weights are published here.

Provenance

Known limitations (v1)

• Garment-part segmentation occasionally includes nearby objects (e.g. furniture the subject sits on) in the bottom crop, which can render as a faint artifact. A matting-refinement pass on the control crops is planned for v2.
• Best on full-body, standing/seated fashion framing; all-dark outfits are the weakest case.

Apache-2.0. Trained by xocialize.