Cycle-Consistent Tuning for Layered Image Decomposition (CVPR 2026)

We present an in-context image decomposition framework that leverages large diffusion foundation models for layered separation. Our method focuses on the challenging case of logo-object decomposition, where the goal is to disentangle a logo from the surface on which it appears while faithfully preserving both layers.

Our approach introduces:

• Cycle-Consistent Tuning: Joint training of decomposition and composition models with reconstruction consistency
• LoRA Adaptation: Lightweight fine-tuning of pretrained diffusion models
• Progressive Self-Improvement: Iterative training set augmentation with high-quality model-generated examples

The method achieves accurate and coherent decompositions and generalizes effectively across other decomposition types (intrinsic decomposition, foreground-background separation), suggesting its potential as a unified framework for layered image decomposition.

git clone https://github.com/edward3862/ImgDecom.git
cd ImgDecom
uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python -r requirements.txt
source .venv/bin/activate

The inference uses black-forest-labs/FLUX.1-Fill-dev with our LoRA weights.

Download the weights to the path expected by the inference scripts:

hf download picana/ImgDecom pytorch_lora_weights.safetensors --local-dir checkpoints

For the provided examples and test-set inference, download the logo-product test dataset:

hf download picana/ImgDecom --repo-type dataset --local-dir data/logo_product_test

For the example three-panel input, run:

bash scripts/inference_single_image.sh

You can also provide an image and decomposition prompt directly:

python inference_single_image.py \
  --input-image path/to/image.png \
  --input-format image \
  --decomp-prompt "A stepwise decomposition of branding elements. [LEFT]: {an object with a logo.} [MIDDLE]: {the isolated logo.} [RIGHT]: {the object without the logo.}"

Set --input-format grid for an existing 1536 × 512 three-panel image. With --input-format image, the script scales the image to fit a 512 × 512 panel, pads it if needed, and places it on the left of a three-panel grid.

bash scripts/inference.sh

The script processes the PNG inputs in data/logo_product_test/grid/. Each image_XXXXX.png needs matching image_XXXXX_decomp.txt and image_XXXXX_comp.txt files in the same directory. For each input, it saves a decomposition image (_decomp.png) and a recomposed image (_recomp.png). To set the input or output directory explicitly, run:

python inference.py \
  --folder data/logo_product_test/grid \
  --output-dir output/test_set

If you find this work useful for your research, please cite:

@inproceedings{gu2026cycle,
  title={Cycle-Consistent Tuning for Layered Image Decomposition},
  author={Gu, Zheng and Lu, Min and Sun, Zhida and Lischinski, Dani and Cohen-Or, Daniel and Huang, Hui},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}