lefze/Marigold-V2-roughness-test

Marigold V2 roughness test

Experimental roughness adapter for the Marigold V2 inference architecture, fine-tuned from the albedo task on the Qwen-Image-Edit-2511 backbone. This is a community experiment, separate from the official Marigold release.

Checkpoint

• Completed training step: 10,500.
• Training resolution: 512 pixels.
• Transformer adapter: rank-128 LoRA.
• Includes the checkpoint's trained VAE decoder.
• Output: linear scalar roughness, 0 = smooth, 1 = rough.
• Single fixed transformer timestep: 499/1000.
• Prompt embeddings: qwen_edit_2509_qwen_albedo_rgb_dummy512 from the official Marigold V2 checkpoint repository.

trainables.safetensors contains the Diffuser.* LoRA tensors and VAE.* decoder tensors. roughness.json describes the compatible architecture and encoding. Both are required. This is not a standalone Diffusers pipeline or a generic LoRA that can be loaded without Marigold's task inference code.

Substance Designer

In the WarNeru Designer AI plugin, enable Roughness on AI Marigold 2, then press Generate AI. The plugin downloads its pinned release on demand and outputs a single-channel 16-bit linear grayscale bitmap at the input size. The shared backbone, VAE, and prompt embeddings are installed by plugin setup. Use Roughness Adapter Strength to adjust the transformer LoRA (default 1).

Plugin source: https://gitlab.com/Londo001/WarNeru/-/tree/tools/designer-ai-nodes/tools/Plugins/Substance%20Designer/AI%20Nodes

Limitations and provenance

This is a test checkpoint. No quantitative benchmark results or detailed training-dataset description are published with this release. Predictions are image-conditioned estimates and can be ambiguous under unknown lighting.

No optimizer state, training images, credentials, or machine-specific paths are included. See LICENSE and NOTICE for upstream attribution.