AFFMAE brings 5x faster high-res vision pretraining to desktop hardware, at ECCV 2026
CSProfKGD · x · 2026-09-07
Researchers present AFFMAE at ECCV 2026: a masking-friendly hierarchical pretraining framework built on AutoFocusFormer's adaptive, off-grid token merging, aimed at high-resolution microscopy segmentation.
Key points:
- David Smerkous heavily optimized AutoFocusFormer code, yielding 5x inference speedup at 1024x1024 and 50% lower memory cost;
- The paper introduces numerically stable mixed-precision Triton kernels and a lightweight point-based decoder reusable as a segmentation head;
- At equal parameter counts, AFFMAE matches MAE fine-tuning on foot process width estimation with a ViT backbone, while pretraining 2x faster with halved peak memory, and up to 5x fine-tuning throughput at 1024px — enabling high-res pretraining on desktop hardware;
- Paper and code are open-sourced; presented at ECCV Poster Session 3, #74.
More from Infra
- Intel and AMD unite PyTorch CPU optimizations in oneDNN for a common x86 path — PyTorch · 2026-10-23
- Google and Meta open-source TorchTPU, a native PyTorch backend for TPUs — PyTorch · 2026-10-23
- Memory-bound at batch 1, compute-bound at batch N: how GPUs actually run LLMs — techNmak · 2026-10-23
- Intel demos AI Playground Home Agent: control local PyTorch compute from your phone via Slack — PyTorch · 2026-10-23
- TorchTPU and TorchNeuron Backends Land as PyTorch Becomes AI's Hardware Battleground — PyTorch · 2026-10-23
- TokenGate e2e-learns to drop tokens on the fly, opening a new frontier for LLM prefill — CShorten30 · 2026-10-23