BigCell Generates Coherent Gigapixel Whole-Slide Pathology Images, Boosting Few-Shot Classification 18%
StonyBrook-CVLab · hf · 2026-10-15
Stony Brook's BigCell is the first method to generate coherent whole-slide histopathology images at native gigapixel resolution. It trains a flow-matching transformer at 2048×2048 on 120,000+ public WSIs, and with inference-time sliding-window synthesis produces images up to 32,768×32768, outperforming prior tile generators. An adaptive scheduling strategy speeds generation 3× with minimal quality loss, and a separate model enables the first 8k-resolution text-to-image from pathology reports. Augmenting with synthetic slides improves few-shot slide-level classification by up to 18%.
More from Multimodal
- Claymation-style AI video of a carnival ring toss charms Reddit — Ok-Salamander-3477 · 2026-10-15
- AI portrait thread shows photoreal red-haired fashion shots rivaling studio photography — LudovicCreator · 2026-10-15
- AI video fumbles a handshake: hands still break in generated clips — misovalko · 2026-10-15
- RAESR Does Super-Resolution in Frozen DINOv3 Latent Space at 37ms per 512px Image — RTP-LLM · 2026-10-15
- Austin AI Film Festival screens 700+ submissions across 11 categories this week — lmoroney · 2026-10-15
- Fish Audio's Drama-3 voice-cloning Space trends on Hugging Face — fishaudio · 2026-10-15