DNF-Intrinsic: Deterministic Noise-Free Diffusion for Indoor Inverse Rendering.

Paper

This is official code of DNF-Intrinsic: Deterministic Noise-Free Diffusion for Indoor Inverse Rendering.

Structure

├── test_input                  <- Place the input images here
├── test_output                  <- The output results are saved here
├── infer.py                   <- Code for DNF-intrinsic prediction
├── requirement.txt       <- Env file
└── README.md

Installation

pip install -r requirement.txt

Download stable-diffusion-3

Please manually download the Stable Diffusion 3 model weights from the official Hugging Face repository at: https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers/

├── stable-diffusion-3-medium-diffusers                  
├──── model_index.json
├──── vae
├──── transformer
├──── tokenizer
├──── tokenizer_2
├──── tokenizer_3
├──── text_encoder
├──── text_encoder_2
├──── text_encoder_3
├──── scheduler

Download DNF-Intrinsic Lora model

Download the checkpoint folder from https://drive.google.com/drive/folders/1mTHbkTY-58iewCMp_KOxD5wq98PpwoJe?usp=drive_link

├── checkpoint                  
├──── adapter_config.json
├──── adapter_model.safetensors

Inference

python infer.py \
--pretrained_model_path "./stable-diffusion-3-medium-diffusers/" \
--input_dir ./test_input \
--output_dir ./test_output \
--peft_model_path ./checkpoint \
--num_inference_steps 10 \
--device cuda:0