RGDM

RGDM, the renormalization group-based diffusion model, is a generative diffusion framework built by Kanta Masuki and Yuto Ashida at the University of Tokyo Department of Physics. Its subject is the noise schedule — the sequence of corruptions a diffusion model learns to reverse. Conventional diffusion models fix that schedule by heuristics tuned per dataset, ignoring the multiscale structure natural data possesses. RGDM instead derives it from the exact renormalization group, the framework physicists use to relate descriptions of a system at different length scales.

The consequence is a generation process that runs coarse-to-fine by construction. A flow equation progressively erases information from fine-scale detail up to coarse-grained structure during the forward process; reversing that flow builds a sample by resolving global organization first and filling in fine features last. Because the schedule comes from theory rather than tuning, it transfers across data types without re-derivation.

Protein structure is one of two domains the paper uses to demonstrate this — the other is natural image generation on CIFAR-10 and FFHQ. Within the protein task, RGDM samples from a distribution of alpha-carbon backbones given a sequence, a framing the authors position as complementary to deterministic single-structure predictors and relevant to systems such as intrinsically disordered proteins where one static answer is the wrong output.

#Key Features

• Theory-derived noise schedule: The colored-noise schedule follows from a renormalization group flow equation rather than data-dependent hyperparameter search, removing a tuning step conventional diffusion models require per dataset.
• Coarse-to-fine generation: Because the forward flow removes high-wavenumber modes progressively, sampling constructs the global fold before local geometry rather than creating most meaningful structure in the final few steps.
• Sequence-conditioned sampling: Node and edge features generated from an amino acid sequence by a protein language model condition the reverse diffusion, so the model samples a sequence-specific structural distribution.
• Length-adaptive step count: The number of generation steps scales with sequence length as T = T0 + tau ln(N/N0), keeping the renormalization scale at the final step proportional to 1/N across proteins of different sizes.
• Fewer steps for equivalent quality: The margin widens as the step budget shrinks, letting the model reach comparable sample quality with roughly an order of magnitude fewer iterations.

#Technical Details

The denoising network is an e3nn equivariant neural network that learns the colored noise as a function of the corrupted structure, the diffusion time, and the sequence feature tensor. Sequence features — separate node and edge embeddings — come from OmegaFold, whose language model supplies positional and pairwise information about each residue. The protein implementation builds on the EigenFold codebase. Training used every Protein Data Bank structure deposited before 30 April 2020, 232,646 in total, validated on the 14,201 structures deposited between 1 May and 30 November 2020. Each model trained for 10 epochs with Adam and an exponential moving average ratio of 0.9999, taking about two weeks on a single consumer NVIDIA GPU.

Evaluation uses the 182 CAMEO targets released between August and October 2022, excluding targets of 750 or more residues, for which the feature generator cannot produce a tensor. Sample quality is scored by RMSD, TM-score, GDT-TS, and GDT-HA. Against a DDPM using the identical denoising network and step count, RGDM produces sharply unimodal quality distributions where the DDPM's are bimodal and frequently inaccurate, and it wins across all four metrics on single-structure accuracy. Both remain below the AlphaFold 2 reference values on the same targets. One behavior is worth knowing before use: sampling fails outright roughly once in every fifteen attempts, and the reported CAMEO figures assign zero score to those failures.

#Applications

The immediate audience is methods research on generative models, where the contribution is a principled alternative to schedule tuning applicable to any diffusion setting. For structural biology, the relevant use is sampling conformational ensembles rather than committing to a single predicted structure — the regime that matters for flexible loops, disordered regions, and alternative conformations. The repository provides training and sampling scripts for both the model and its DDPM control, and the trained checkpoints, configurations, and sampled structures used in the paper are distributed for download.

#Impact

RGDM's contribution is conceptual as much as empirical: a construct from statistical physics replaces an empirically tuned component of diffusion models, improving sample quality and sampling efficiency at once across data as different as protein backbones and photographs. Scope should be read carefully. Protein structure is one of two validation domains for a diffusion technique, benchmarked against a matched DDPM rather than against the structure-prediction field, and accuracy does not approach AlphaFold 2. It remains a preprint. Code is released under the MIT license; trained weights are distributed through a Google Drive folder linked from the repository rather than a versioned archive, and no model card accompanies them.