This is a LoRA (Low-Rank Adaptation) model that has been trained for image overlay tasks, allowing you to seamlessly blend and place elements into existing images.

Model Details
Uses
Training
Usage Tips
Academic Use and Citation

Model Details

Model Description: This model is a LoRA fine-tune of black-forest-labs/FLUX.1-Kontext-dev specifically designed for image overlay workflows, enabling natural integration of new elements into existing scenes.

Trained by: ilkerzgi and gokaygokay
Model Type: LoRA for Image Overlay
Language(s): English
License: flux1-dev-non-commercial-license
Parent Model: black-forest-labs/FLUX.1-Kontext-dev
Resources for more information:
Download the weights
Try the model directly on fal.ai

Uses

Direct Use

This model is designed for image overlay tasks, allowing you to naturally integrate new objects, elements, or modifications into existing images. The model excels at understanding context and placement, creating seamless compositions.

The model works best when using the trigger phrase: "place it".

Recommended Settings

LoRA Scale: 1.0 or 0.9 for optimal results
Image Format: Square format images work best
Overlay Area: Clear definition of where elements should be placed
Context Understanding: Works well with natural scene composition

Training

Training Procedure

Training was performed using the FLUX Kontext Trainer on fal.ai.

The base model for training was black-forest-labs/FLUX.1-Kontext-dev, with specialized focus on image overlay scenarios and contextual element placement.

Usage Tips

Use square format images for best results
Set LoRA scale to 1.0 or 0.9 for optimal output quality
Use the simple prompt "place it" as the trigger phrase
Provide clear context about what you want to overlay and where
The model understands spatial relationships and lighting conditions for natural integration
Works well with both object placement and scene modifications

Academic Use and Citation

This model is cited and evaluated as a baseline in:

LayerComposer: Multi-Human Personalized Generation via Layered Canvas (2025 preprint)
Canvas-to-Image: Compositional Image Generation with Multimodal Controls (ACM SIGGRAPH 2026)

If this model is useful in your research, please cite the model and the relevant paper:

bibtex

@software{ilkerzgi_aydogan2025overlay_kontext,
  author = {{ilkerzgi} and Aydoğan, Gökay},
  title = {Overlay Kontext Dev LoRA},
  year = {2025},
  url = {https://huggingface.co/ilkerzgi/Overlay-Kontext-Dev-LoRA},
  note = {Hugging Face model}
}

@article{qian2025layercomposer,
  author = {Qian, Guocheng Gordon and Zhang, Ruihang and Chen, Tsai-Shien and Dalva, Yusuf and Goyal, Anujraaj Argo and Menapace, Willi and Skorokhodov, Ivan and Dong, Meng and Sahni, Arpit and Ostashev, Daniil and Hu, Ju and Tulyakov, Sergey and Wang, Kuan-Chieh Jackson},
  title = {LayerComposer: Multi-Human Personalized Generation via Layered Canvas},
  journal = {arXiv preprint arXiv:2510.20820},
  year = {2025},
  doi = {10.48550/arXiv.2510.20820}
}

@inproceedings{dalva2026canvas,
  author = {Dalva, Yusuf and Qian, Guocheng Gordon and Goldenberg, Maya and Chen, Tsai-Shien and Aberman, Kfir and Tulyakov, Sergey and Yanardag, Pinar and Wang, Kuan-Chieh Jackson},
  title = {Canvas-to-Image: Compositional Image Generation with Multimodal Controls},
  booktitle = {ACM SIGGRAPH 2026 Conference Papers},
  year = {2026},
  doi = {10.1145/3799902.3811219}
}

Co-trained by ilkerzgi and Gökay Aydoğan.

ilkerzgi/Overlay-Kontext-Dev-LoRA

Kontext Dev LoRA - Image Overlay

Table of Contents