FAPEIR_Uniworld — Initial Weights for FAPE-IR

Initial weights for FAPE-IR: Frequency-Aware Planning and Execution Framework for All-in-One Image Restoration.

📄 Paper (arXiv 2511.14099)   💻 Code   🏋️ Trainset   🧪 Testset

---

💡 What This Repo Is

This repository releases the initial weights required to start training FAPE-IR — i.e. all pretrained components consumed by the YAML config

scripts/denoiser/flux_qwen2p5vl_7b_vlm_512.yaml

in the FAPE-IR codebase. Concretely it bundles:

• the UniWorld-V1 initialization (Qwen2.5-VL-7B-Instruct + FLUX.1-dev re-organized weights),
• the SigLIP-v2 encoder used by the executor,
• a small set of projection / connector weights (mlp2, mlp3, SigLIP→FLUX redux),
• a VGG checkpoint used by the LPIPS loss.

⚠️ This is NOT the post-training checkpoint reported in the paper.

---

📂 File Layout

After downloading, the repository is meant to be placed under FAPE-IR/weights/ exactly as below:

weights/
├── flux/                                  # FLUX.1-dev backbone (re-organized)
├── siglip/                                # SigLIP-v2 encoder
├── uniworld/                              # UniWorld-V1 (Qwen2.5-VL-7B-Instruct + denoiser projection)
├── denoise_projector_params.bin           # planner-token → denoiser projector  (mlp2)
├── flux-redux-siglipv2-512.bin            # SigLIP-v2 → FLUX redux projector
├── vae_projector_only.bin                 # VAE high/low-frequency projector    (mlp3)
└── vgg.pth                                # VGG weights for LPIPS loss

These names match one-to-one with the fields of the YAML config:

model_config:
  pretrained_lvlm_name_or_path:    weights/uniworld
  pretrained_denoiser_name_or_path: weights/flux
  pretrained_siglip_name_or_path:  weights/siglip
  pretrained_mlp2_path:            weights/denoise_projector_params.bin
  pretrained_mlp3_path:            weights/vae_projector_only.bin
  pretrained_siglip_mlp_path:      weights/flux-redux-siglipv2-512.bin

training_config:
  lpips_weights_path:              weights/vgg.pth

If you change the layout, remember to update the YAML accordingly.

---

⬇️ Download

# inside the FAPE-IR project root
mkdir -p weights
huggingface-cli download David0219/FAPEIR_Uniworld --local-dir ./weights

Or in Python:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="David0219/FAPEIR_Uniworld",
    local_dir="./weights",
    local_dir_use_symlinks=False,
)

---

📝 Intended Use & Limitations

Intended use. Research on All-in-One image restoration with an MLLM-as-planner + diffusion-as-executor paradigm; reproducing or extending FAPE-IR; ablating individual components (LoRA-MoE routing, frequency regularization, adversarial training).

Limitations.

• Training requires substantial GPU memory because the executor is FLUX.1-dev (12B-class) and the planner is Qwen2.5-VL-7B-Instruct.
• These are initial weights only — running inference with them directly will not reproduce FAPE-IR's reported quality. Train first.
• The base models (FLUX.1-dev, Qwen2.5-VL, SigLIP-v2) keep their original licenses; in particular FLUX.1-dev is non-commercial. Users must comply with each license individually.

---

🔖 Citation

@article{liu2025fape,
  title   = {FAPE-IR: Frequency-Aware Planning and Execution Framework for All-in-One Image Restoration},
  author  = {Liu, Jingren and Xu, Shuning and Yang, Qirui and Wang, Yun and Chen, Xiangyu and Ji, Zhong},
  journal = {arXiv preprint arXiv:2511.14099},
  year    = {2025}
}

---

📜 License & Acknowledgement

Apache-2.0 for the connector / projector weights released here. The bundled UniWorld-V1, FLUX.1-dev, Qwen2.5-VL-7B-Instruct, SigLIP-v2 and VGG weights retain their original licenses, which users must respect.

We thank the teams behind UniWorld, FLUX.1-dev, Qwen2.5-VL, and SigLIP-v2 for open-sourcing their work.