TransNormal-2

Official model weights for TransNormal-2: Geometry-Grounded Rectified Flow with Edge-Aware Decoding for Precise Normal Estimation.

GitHub · Hugging Face · Project Page · arXiv

TransNormal-2 estimates surface normals from one RGB image using a single-step, FLUX.2-based rectified-flow predictor and geometry-aware refinement.

Weights

File	Component
lora_core_predictor.safetensors	Core predictor LoRA (rank 256, alpha 256)
lcm_normal.safetensors	Local Continuity Module (LCM)
grm.safetensors	Geometric Refinement Module (GRM)
config.json	Architecture and loading configuration

The three weight files use BF16 Safetensors and total 1.402 GB. They contain model tensors only. The base model is distributed separately at black-forest-labs/FLUX.2-klein-base-9B; follow its access instructions and license terms.

Download and use

Install the download utility:

pip install -U huggingface_hub

Download the weights and configuration without authentication:

from huggingface_hub import snapshot_download

weights_dir = snapshot_download(
    repo_id="Longxiang-ai/TransNormal-2",
    local_dir="TransNormal-2-weights",
    allow_patterns=["*.safetensors", "config.json", "README.md", "LICENSE", "NOTICE", "licenses/*"],
    token=False,
)
print(weights_dir)

Inference

Inference code and complete instructions are available on GitHub.

git clone https://github.com/longxiang-ai/TransNormal-2.git
cd TransNormal-2
python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python inference.py --input path/to/transparent_image.jpg --domain transparent --output_dir outputs/glass --save_npy

Replace path/to/transparent_image.jpg with your own image. Use Python 3.10 and a CUDA GPU with BF16 support. Follow the base model's access instructions before its first download; run hf auth login if authentication is required. The task weights in this repository download without authentication. For ordinary scenes, use the default --domain opaque. The same --input option also accepts a folder of images. See GitHub for CPU offload and the Python API.

This repository contains the weights and configuration. The custom TransNormal-2 pipeline is provided in the GitHub repository; a generic Diffusers image-generation pipeline does not implement surface normal estimation.

License

The authors' model contributions are licensed under CC BY-NC 4.0, subject to the underlying rights described in LICENSE. As a modified derivative of FLUX.2 [klein] base 9B, this model is also subject to the FLUX Non-Commercial License v2.1, including its non-commercial and non-production restrictions. See NOTICE for the required attribution. The base model is not redistributed here.

Citation

@misc{li2026transnormal2,
  title = {TransNormal-2: Geometry-Grounded Rectified Flow with Edge-Aware Decoding for Precise Normal Estimation},
  author = {Mingwei Li and Yi Yang and Hehe Fan},
  year = {2026},
  eprint = {2609.06665},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2609.06665}
}

Questions: @longxiang-ai or GitHub Issues.