Extended Data Fig. 2. RFdiffusion learns the distribution of the denoising process, and inference efficiency can be improved.

A) Analysis of simulated forward (noising) and reverse (denoising) trajectories shows that the distribution of Cα coordinates and residue orientations closely match, demonstrating that RFdiffusion has learned the distribution of the denoising process as desired. Left to right: i) average distance between a Cα coordinate at X_t and its position in X₀; ii) average distance between a Cα coordinate at X_t and X_t-1; iii) average distance between adjacent Cα coordinates at X_t; iv) average rotation distance between a residue orientation at X_t and X₀; v) average rotation distance between a residue orientation at X_t and X_t-1. B-C) While RFdiffusion is trained to generate samples over 200 timesteps, in many cases, trajectories can be shortened to improve computational efficiency. B) Larger steps can be taken between timesteps at inference. Decreasing the number of timesteps speeds up inference, and often does not decrease in silico success rates (left) (for example, on an NVIDIA A4000 GPU, 100 amino acid designs can be generated with 15 steps, in ~11s, with an in silico success rate of over 60%). When normalized for compute budget (center) it is often much more efficient to run more trajectories with fewer timesteps. This can be done without loss of diversity in samples (right). For harder problems (e.g. unconditional 300 amino acids), one must strike an intermediate number of total timesteps (e.g., T = 50) for optimal compute efficiency. Note that for all other analyses in the paper, 200 inference steps were used, in line with how RFdiffusion is trained. C) An alternative to taking larger steps is to stop trajectories early (possible because RFdiffusion predicts X₀ at every timestep). In many cases, trajectories can be stopped at timestep 50–75 with little effect on the final in silico success rate of designs (left), and when normalized by compute budget (center), success rates per unit time are typically higher generating more designs with early-stopping. Again, this can be done without a significant loss in diversity (right).