timm author ships first CLIP model trained with both FlexiViT patch jitter and NaFlex seq-length jitter
wightmanr · x · 2026-09-06
Ross Wightman (timm) pushed a test CLIP model from OpenCLIP trial runs, believed to be the first trained with both patch size jitter (FlexiViT) and sequence length jitter (NaFlex style)—a small NaFlexViT CLIP on CC12M, compared against fixed patch size with variable sequence length. Recent timm 1.0.29 and OpenCLIP releases added patch jitter support, with timm gaining pinv-matrix prewarm caching and pre-patchified NaFlex pipeline support (flat nn.Linear embed, 5D tensors upstream).
Related event: timm Author Releases NaFlexViT CLIP Model with Dual Jitter(2 posts)→
More from Research
- DeepMind paper shows cheating spreading like an epidemic across ~100 AI agents — jackclarkSF · 2026-09-06
- Context Compaction Theory: first formal proof linking agent compaction to communication complexity — lateinteraction · 2026-09-06
- Last theorem on Freek Wiedijk's famous list has been formalized — satnam6502 · 2026-09-06
- New theory shows how to scale residual network updates when layer weights are correlated — burkov · 2026-09-06
- Terence Tao post sparks buzz as a cool mechanistic interpretability application — Sauers_ · 2026-09-06
- GPT-6 Astra tops surgical AI leaderboard but loses to models 1000x smaller — ddonoho · 2026-09-06