Skip to main content

Extract-Outfit (local)

prithivMLmods/QIE-2511-Extract-Outfit is an Apache-2.0 LoRA for Qwen/Qwen-Image-Edit-2511 (also compatible with 2509) that does the reverse of try-on: it removes the person, pose and background and renders the clothing as a clean, flat studio mockup, keeping fabric texture, colour, stitching and patterns. Use it to turn on-model or street photos into catalogue-style garment images, or to produce garment input for try-on. OpenTryOn exposes it in the edit service as qie-2511-extract-outfit (Path B, extra="local").

Base modelQwen/Qwen-Image-Edit-2511 (Apache-2.0), Diffusers QwenImageEditPlusPipeline
LoRAprithivMLmods/QIE-2511-Extract-Outfit (Apache-2.0), loaded onto the pipeline before offload
Trigger promptExtract the clothing and create a flat mockup. (the default)
HardwareSame as qwen-image-local: CUDA GPU, ~40GB+ VRAM in bf16; CPU offload is on by default for smaller cards

The whole stack is Apache-2.0, so it is the commercially clean choice among the garment-extraction LoRAs. It is a community adapter: the card shows example images but publishes no metrics, so judge quality on your own photos.

Input must show the full outfit. A half-cropped photo only extracts the visible region.

Requirements​

pip install opentryon[local]
pip install -U diffusers transformers accelerate peft

peft is needed to load LoRA weights; the adapter raises a clear error if it is missing. Overrides: EXTRACT_OUTFIT_LORA_ID (LoRA repo or path) and QWEN_IMAGE_EDIT_MODEL_ID / QWEN_IMAGE_EDIT_PATH (base model, shared with qwen-image-local).

CLI​

opentryon edit --model qie-2511-extract-outfit --image model_wearing_look.jpg

# More control
opentryon edit --model qie-2511-extract-outfit --image look.jpg \
  --steps 40 --true-cfg-scale 4.0 --seed 7 --num-images 2 --no-cpu-offload

Flags: --image (exactly one), --prompt (defaults to the trigger prompt; other prompts may not activate the LoRA), --steps, --true-cfg-scale, --guidance-scale, --seed, --num-images, --lora-id, --edit-model-id, --dtype, --no-cpu-offload. Output is saved under outputs/ like any edit model.

Python​

from tryon.models.extract_outfit import ExtractOutfitAdapter

adapter = ExtractOutfitAdapter()
images = adapter.extract_outfit("model_wearing_look.jpg")
images[0].save("flat_mockup.png")

Notes​

  • MCP tool: edit_qie_2511_extract_outfit. Planner pins: extract outfit, extract the outfit, qie-2511-extract-outfit.
  • The adapter subclasses the repo's Qwen-Image local runner, so loading, dtype and offload behave exactly like qwen-image-local.
  • Similar models exist on the FLUX.2 klein 9B base (fal/virtual-tryoff-lora), but that base is non-commercial; see Integrate next.
  • No GPU was available when this was integrated; the adapter is tested against a stubbed pipeline (LoRA attached once and before offload, trigger prompt, validation), not real weights.