Biogod/spatial-banana

Spatial Banana biomedical segmentation adapters

This repository stores Patho-Banana task LoRAs for adapting Qwen-Image-Edit-2511 to direct biomedical instance-mask generation. The adapters are not standalone pipelines: inference requires the frozen Qwen base, the official four-step Qwen-Image-Edit-2511 Lightning LoRA, task-specific prompts, aspect-aware geometry, and deterministic RGB-to-instance decoding.

Recommended release

Use Mix448 E6 rank 128 for the current multimodal segmentation protocol:

experiments/multimodal/mix448_screen_s560_v1/runs/s560_e6_mix448_r128_he_source/final/pytorch_lora_weights.safetensors

Field	Value
Size	1,699,695,208 bytes
SHA-256	73ec67c27686b2746e6e328a7b27a61246806b6f6e55d0e6580f8203554ac4bb
Training mixture	128 PanNuke H&E plus 320 source-balanced non-H&E images
Task LoRA	rank 128, alpha 128, dropout 0
Inference	BF16, 4 steps, seeds 42/314159/271828
Decoder	qwen_topology16_majority_v1

Results

On the previously viewed 822-image CellSAM non-H&E cohort:

Method	Image-mean F1	Pooled F1	PQ	AJI+	FG Dice
Mix448 E6, current decoder	0.7538	0.7049	0.5943	0.6229	0.8099
Same outputs, legacy decoder	0.5048	0.3430	0.3969	0.4132	0.7222
CellSAM v1.2 official	0.8927	0.8926	0.7465	0.7692	0.9070

This is retrospective same-protocol evidence, not blinded SOTA or clinical validation. Qwen remains below CellSAM overall.

Download

The source repository provides pinned HF-Mirror commands, SHA-256 verification, local path placement, and preflight instructions:

• Weight download guide
• Full model card
• Current results

Archive contents

The other LoRAs cover nuclei few-shot experiments, epoch-aligned training, Data-Frontier 256, MM144 adaptation, source-balance screening, and all eight Mix448 screen arms. They are development or recovery artifacts unless their evidence package explicitly says otherwise. The presence of a weight does not mean it passed a scientific gate.

ARCHIVE_MANIFEST.tsv records every archived weight path, byte size, and SHA-256. Resolved configs, prompts, metrics, per-image rows, and completion audits remain in the GitHub source repository.

Limitations and terms

The recommended adapter requires the Qwen base and Lightning LoRA and remains subject to their licenses and terms. It is not intended for clinical diagnosis or treatment decisions. This release grants no rights to upstream models or training/evaluation datasets.