FLUX.2-klein-base-9B Virtual Try-Off LoRA

Virtual Try-Off: Given an image of a person wearing clothing and a garment category prompt, the model generates a clean image of the garment as if it were photographed alone.
The model reconstructs the clothing item while preserving its style, texture, color, and design from the input image.
1 input image (person wearing clothes) + text category → 1 output garment image
Built with fal.ai.

Usage

• Try the LoRA on fal.ai Playground
• Try the workflow (image-to-video) at fal.ai Workflow
• ComfyUI: download compatible weights virtual-tryoff-lora_comfy.safetensors
• Diffusers:

  import torch
  from diffusers import Flux2KleinPipeline
  from PIL import Image
  
  pipeline = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-base-9B", 
    torch_dtype=torch.bfloat16, 
    low_cpu_mem_usage=False
  ).to("cuda")
  pipeline.load_lora_weights(
      "fal/virtual-tryoff-lora", 
      weight_name="virtual-tryoff-lora_diffusers.safetensors", 
      adapter_name="vtoff"
  )
  pipeline.set_adapters("vtoff", adapter_weights=1.0)
  pipeline.fuse_lora(adapter_names=["vtoff"], lora_scale=1.0)
  
  image = pipeline(
      image=Image.open("<your_image>.jpg"),
      prompt="TRYOFF extract the full outfit over a white background, product photography style.  NO HUMAN VISIBLE (the garments maintain their 3D form like an invisible mannequin).",
      height=1024,
      width=768,
      num_inference_steps=28,
      guidance_scale=5.0,
      generator=torch.Generator("cuda").manual_seed(42),
  ).images[0]
  

Training

Trained with fal.ai trainer.

• Base model: FLUX.2-klein-base-9B
• Steps: 10000
• Learning Rate: 0.00005
• Dataset: 300 image pairs (model + garment) of shape 1024x1024

Author

Created by Riza Velioglu at fal.ai