Skip to main content

FLUX.2-klein-4B try-on LoRA (local)

xocialize/tryon-FLUX.2-klein-4B-lora is a rank-32 edit LoRA for black-forest-labs/FLUX.2-klein-4B that dresses a person in a specified top and bottom, preserving identity, pose and background. OpenTryOn runs it as klein-4b-tryon in the vton service (Path B, extra="local").

Why it matters: the base (FLUX.2 klein 4B, distilled, about 13 GB VRAM) and the LoRA are both Apache-2.0. The more popular klein 9B try-on LoRAs (for example fal/flux-klein-9b-virtual-tryon-lora) sit on a base released under the FLUX Non-Commercial License, which blocks commercial use. Compare with FASHN VTON 1.5 (Apache-2.0, maskless, single garment), Leffa and CatVTON.

v1 community LoRA — validate before relying on it

Trained on only 273 full-body tuples (Pexels frames, step-2000 checkpoint) with no published metrics. Known limits from the card: nearby objects can leak into the bottom garment crop (faint artifacts) and all-dark outfits are the weakest case. The card gives no sampler settings, so the defaults here are BFL's klein-4B defaults (4 steps, guidance 1.0) — tune them on your data. No GPU was available when this was integrated; the adapter is tested against a stubbed pipeline (reference order, masking, LoRA load order), not real weights.

Inputs (order is load-bearing)​

SlotFlagWhat to pass
1 — person--person-imageFull-body photo, clothing-agnostic (outfit region greyed)
2 — top--top-imageIsolated upper-body garment, product-style
3 — bottom--bottom-imageIsolated lower-body garment
(one-piece)--dress-imageA dress or jumpsuit: used for both the top and bottom slots

This adapter does not segment clothing. Either pass an already clothing-agnostic person image, or pass --mask-image (white = the clothing region to replace) and the adapter greys that area (RGB 128) before inference. Any parser can produce the mask; note that fashn-human-parser and segformer_b2_clothes (the one the LoRA author used) carry their own licences — check them before commercial use. The garments must be isolated crops; the LoRA was trained on both a top and a bottom, so both (or a dress) are required.

Requirements​

pip install opentryon[local]
pip install git+https://github.com/huggingface/diffusers.git peft accelerate

Flux2KleinPipeline is not in a stable diffusers release yet. CUDA GPU, about 13 GB+ VRAM (CPU offload is on by default). Overrides: KLEIN_TRYON_LORA_ID, KLEIN_TRYON_BASE_ID (or --lora-id / --base-model-id).

CLI​

# Separate top and bottom, person already agnostic
opentryon vton --model klein-4b-tryon \
  --person-image person_agnostic.png --top-image shirt.png --bottom-image jeans.png

# Let the adapter grey out the clothing region from a mask
opentryon vton --model klein-4b-tryon \
  --person-image person.jpg --mask-image clothing_mask.png \
  --top-image shirt.png --bottom-image jeans.png --seed 7

# One-piece garment
opentryon vton --model klein-4b-tryon --person-image person_agnostic.png --dress-image dress.png

Output size defaults to the person image's size, with the long side capped at 1024 and snapped to multiples of 16 (--height / --width override). Other flags: --prompt (must start with TRYON; the default is the LoRA's trigger prompt), --steps (default 4), --guidance-scale (default 1.0), --seed, --num-images, --dtype, --no-cpu-offload.

Python​

from tryon.models.klein_tryon import Klein4BTryOnAdapter

adapter = Klein4BTryOnAdapter()
images = adapter.generate_and_decode(person="person.jpg", mask="mask.png", top="shirt.png", bottom="jeans.png")
images[0].save("tryon.png")

Notes​

  • MCP tool: vton_klein_4b_tryon. Planner pins: klein 4b tryon, klein-4b-tryon. The vton default stays kling-ai.
  • The weight file is tryon-klein-4b.safetensors (rank 32, "BFL-fused" dialect); if your diffusers build cannot load it, update diffusers from GitHub.