Albertchen96/mess
MeSS TUM2Twin v2 ControlNet
Single-DSN checkpoint (checkpoint-10000) used by the TUM2Twin v2 inference workflow. The ControlNet receives nine channels: depth RGB, semantic RGB, then normal RGB. Normal conditioning uses packed world normals with channel permutation (2, 1, 0). This is the no-reference model; it is not ControlNetModelWithRef.
Code and installation: https://github.com/Albertchen98/mess-code
Paper: https://arxiv.org/abs/2508.15169
Project: https://albertchen98.github.io/mess/
Download from the code repository root:
hf download Albertchen96/mess config.json diffusion_pytorch_model.safetensors --local-dir mess_weights/single-dsn/checkpoint-10000/controlnetUse with Stable Diffusion 1.5 and LCM-LoRA for SD 1.5 as documented in the code repository. This repository contains only the custom ControlNet, not those base models. The v2 workflow is an adaptation and does not reproduce every component of the paper's full method.
The authors have not yet specified a license for these custom weights. Refer to the respective upstream model repositories for their terms.
