Equivariant Diffusion · In Your Browser
Generate 3D molecular conformers from a shape constraint — no server, no upload. The model runs entirely client-side on your own machine via ONNX Runtime Web.
The weights are downloaded once from Membrizard/ml_conformer_generator on Hugging Face and kept in this browser’s cache, so later visits load them without downloading. Inference runs on your machine — nothing is uploaded.
Load a matching pair of ONNX files from your disk — to work offline, or to try your own fine-tuned weights. They are read into this tab; nothing is uploaded. The published files are on the Hugging Face Files tab.
The shape constraint is a triplet of moment-of-inertia eigenvalues. Presets approximate common ligand geometries. Each also sets the atom count and variance below — edit any of them and the shape becomes Custom.
Molecules per run — in Animate, one trajectory each. Runtime scales with this.
More steps → better quality, proportionally slower.
Heavy atoms per molecule (15–39).
Allowed spread around the atom count.
Same seed → identical output.
Animate streams every denoising step — about twice the time of Generate. “Live bonds” also predicts bonds on every frame, so you see connectivity form; that adds only a few percent.
Principal moments defining the target shape envelope.
Random shape picks a molecule shape from a sample of ChEMBL bundled with the package and generates for it. It uses Samples, Diffusion Steps and Seed; the shape preset, MOI eigenvalues, atom count, variance and mode do not apply — the atom count comes with the shape, so the runtime estimate above, which is for the slider settings, does not either.
This snippet shows the package API calls for your current weights, settings and last action.