Yiren Song¹, Cheng Liu¹, Haofan Wang², Mike Zheng Shou¹

¹Show Lab, National University of Singapore   ²Lovart AI

[Paper]   [Dataset]

OmniPSD is a unified diffusion-transformer framework for bidirectional conversion between raster images and editable PSD files with full transparency support. It addresses two tasks:

• Text-to-PSD: generates a layered PSD design (background, content layers, text layers) from a text description in a single forward pass, using a 2×2 spatial grid to capture inter-layer relationships.
• Image-to-PSD: decomposes a flat poster image into editable layers via an iterative extract-erase pipeline driven by Flux-Kontext.

Key components:

• RGBA-VAE — a transparency-preserving VAE that encodes/decodes RGBA images.
• Flux-Dev LoRA — fine-tuned for Text-to-PSD generation (4-panel grid layout).
• Flux-Kontext LoRA ×4 — separate expert models for foreground/background extraction of content and text layers.

A subset of our layered poster dataset is available on Hugging Face:

lc03lc/OmniPSD_Layered_Poster

git clone https://github.com/showlab/OmniPSD.git
cd OmniPSD
pip install -r requirements.txt

Base models required:

• FLUX.1-dev — for Text-to-PSD training and inference
• FLUX.1-Kontext-dev — for Image-to-PSD training
• RGBA-VAE weights (FLUX.1-dev-alpha) — replace /PATH/TO/RGBA_VAE/ in scripts

All scripts are run from the OmniPSD root directory. Edit the PATH placeholders in each script before running.

bash scripts/train_psd_flux.sh

Uses Flux-Dev with the RGBA-VAE. Training data is a 4-panel grid of [full poster | content layer | background | text-removed poster].

# Foreground extraction
bash scripts/train_psd_content_front.sh

# Background inpainting
bash scripts/train_psd_content_back.sh

# Foreground extraction
bash scripts/train_psd_text_front.sh

# Background inpainting
bash scripts/train_psd_text_back.sh

All Kontext-based scripts take (image, control) pairs where control is the input poster used as conditioning.

Edit inference/infer_psd_flux.py to set INPUT_TXT_DIR, OUTPUT_ROOT, LORA_PATH, and the RGBA-VAE path, then run:

cd /PATH/TO/OmniPSD
CUDA_VISIBLE_DEVICES=0 PYTHONPATH=$(pwd) python inference/infer_psd_flux.py

Each .txt file in INPUT_TXT_DIR is treated as one prompt. The script runs NUM_PASSES times per prompt with incrementing seeds and saves results under OUTPUT_ROOT/<stem>/.

If you find OmniPSD useful, please cite:

@article{Liu2025OmniPSD,
  title         = {OmniPSD: Layered PSD Generation with Diffusion Transformer},
  author        = {Liu, Cheng and Song, Yiren and Wang, Haofan and Shou, Mike Zheng},
  journal       = {arXiv preprint arXiv:2512.09247},
  year          = {2025},
  archivePrefix = {arXiv},
  eprint        = {2512.09247},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2512.09247},
  url           = {https://arxiv.org/abs/2512.09247}
}