Marigold Intrinsics (IID Lighting v1.1)

This model estimates intrinsic image decomposition from a single RGB image using the Marigold IID Lighting pipeline from ETH Zurich. It produces three physically-meaningful outputs:

Albedo — surface color without lighting effects

Shading — illumination & geometry-based lighting

Residual — information that cannot be fully decomposed

This is useful for:

Room visualization & material replacement

Relighting workflows

AR/VR scene understanding

Normal map / lighting estimation pipelines

🚀 How it Works

This implementation wraps the official model: prs-eth/marigold-iid-lighting-v1-1

When you upload an RGB image, the model runs a short diffusion process and outputs the decomposed intrinsic layers.

📥 Input Parameter Type Default Description image File (PNG/JPG/WebP) required Input RGB image num_inference_steps Integer 4 Denoising steps (higher = better quality, slower) return_zip Boolean false If true, all outputs returned as one ZIP file 📤 Output

You will receive 3 separate images (or ZIP if selected):

Filename Description albedo.png Pure surface color shading.png Lighting-only component residual.png Residual signal

These files can be brought into Blender/Photoshop for texture editing, or combined for relighting.

🧠 Model Details Property Value Framework PyTorch + Diffusers Device GPU if available Precision FP16 on GPU, FP32 on CPU Source Model Marigold IID Lighting v1.1 📝 Example Usage Basic Inference

Upload any clean, well-lit indoor scene image:

image: input.png

Higher Quality / Slower num_inference_steps: 15

ZIP Output return_zip: true

💡 Tips for Best Results

✔ Indoor scenes with defined surfaces ✔ Avoid extreme motion blur or very noisy images ✔ Higher resolution inputs give more detail

📚 Credits

🔬 Research: 📄 Marigold: Intrinsics-Guided Diffusion for Inverse Rendering ETH Zurich — Visual Computing Group

Model Source: Hugging Face — prs-eth/marigold-iid-lighting-v1-1