chrispie/mojidiff-checkpoints

MojiDiff checkpoints

Research checkpoints from MojiDiff: small models, trained from scratch on an RTX 4080 (and, from 2026-10-30, on single B300 GPUs of a hackathon cluster), that write OpenMoji-style emoji as SVG programs in a typed codec. These are research artifacts with honest, mostly negative results, not a polished release. Loading them needs the MojiDiff code (mojidiff.learning.*) from github.com/cpietsch/emojidiff, which also holds every run record and the research log (weblog). In-browser demos are in progress.

Layout

• runs/<run_id>/best.pt: the checkpoint selected by the run's declared rule; latest.pt is the resumable end state (with optimizer); best-any.pt is the best without the selection constraint. run.yaml and result.md record the config, hashes, criteria and results of each run.
• legacy/<name>/checkpoint.zip: first-phase models (manifest plus numpy tensors).
• checkpoints-hf.json: sha256 and size of every file.

Main models

run	what	validation pixel error (72 px)
r2s-cur-d02-s120k-seed2-*	final transcriber: v9's recipe with dropout 0.2 and 120k steps, best of nine seeds on validation	0.064 greedy, 0.045 best of 8 (test: 0.062, 0.040)
vt-v3-d02-s5-*	variational canvas latent (c = 16) on the final transcriber	0.076 from the posterior mean
lfp-v2x-exploratory-*	rectified-flow prior over vt-v3-d02-s5's latent	samples: 17% fragments, CLIP recall 0.41
r2s-full-v7-systems-*	render-to-SVG transcriber, 9.0M parameters	0.077 greedy, 0.057 best of 8
r2s-full-v9-colour-*	v7 with palette permutation cut from 0.5 to 0.1	0.072
vt-v1-*, vt-v2-c16-*	variational canvas latent on v9 (c = 8, 16)	0.083, 0.081 from the posterior mean
lfp-v1x-exploratory-*	rectified-flow prior over vt-v1's latent	samples: 18% fragments, CLIP recall 0.42
latent-v1/v2/v3-*	single-vector program latents	baselines

Retrieval of the nearest training icon scores 0.090 on the same split, and 0.085 when mirrored training renders are allowed (the baseline for the vt rows).

Data and licence

Trained on OpenMoji 17.0.0. All emojis designed by OpenMoji – the open-source emoji and icon project. License: CC BY-SA 4.0. r2s-full-v8-twemoji-* was also trained on adapted Twemoji graphics (CC BY 4.0). The checkpoints, like the MojiDiff code and documentation, are shared under CC BY-SA 4.0 with attribution to both projects.