Quantizing Cellpose-SAM for stem cell imaging: W4/W8 hits 6.76x compression with zero failures
capicu-ai · hf · 2026-09-21
A deployment-oriented evaluation of compressed Cellpose-SAM for iPSC culture segmentation on lab CPUs and edge hardware, using a pre-specified retention criterion: the 95% cluster-bootstrap interval of mean change from FP32 must stay above -0.02 for every imaging modality.
On a stratified 176-field panel spanning BBBC038 nuclei, BBBC039 U2OS fluorescence, and NIST iPSC images:
- W8A16 weight-only quantization preserves instance F1 across all modalities.
- A sensitivity-guided mixed W4/W8 scheme (four INT8 exceptions) achieves 6.76x weight-storage reduction with 0/176 catastrophic failures, matching W8A16 at this sample size.
- Ternary weight-only quantization reaches 12.08x compression but fails catastrophically on 169/176 fields.
The authors argue compression should be evaluated by modality-stratified downstream retention rather than single-number accuracy, and establish a reproducible protocol for auditing compressed foundation models in regulated stem-cell imaging.
Related event: Cellpose-SAM Quantization Achieves 6.76x Compression With Zero Failures(2 posts)→
More from Infra
- Colibrì's Brio mode scores candidate answers instead of generating text locally — Just_Vugg_PolyMCP · 2026-09-21
- Rothschild Redburn goes bearish on AI compute, rates NBIS and CRWV Sell — JOBhakdi · 2026-09-21
- PC memory module prices keep climbing daily as Meta's Muse holds #1 on the App Store — firstadopter · 2026-09-21
- Report: DeepSeek training a 2T-param model with 8T planned, Huawei chips due late 2026 — Hesamation · 2026-09-21
- CXMT's wafer capacity to grow 10x by 2030, nearing Micron parity in late 2026 — Terminator857 · 2026-09-21
- Building a 4-GPU local LLM rig: why Threadripper beats LGA1700 on PCIe lanes — El_90 · 2026-09-21