Silviase/QuIC-360
Viewer • Updated • 23k • 20 • 1
How to use wfwefw/cora-gemma3-4b-densecl with PEFT:
from peft import PeftModel
from transformers import AutoModelForCausalLM
base_model = AutoModelForCausalLM.from_pretrained("google/gemma-3-4b-it")
model = PeftModel.from_pretrained(base_model, "wfwefw/cora-gemma3-4b-densecl")Seed-42 LoRA adapter for the ECCV 2026 CORA release. It must be used with
google/gemma-3-4b-it at revision 093f9f388b31de276ce2de164bdc2081324b9767 and the matching CORA
config included as cora_config.yaml.
QuIC-360 test set (5349 query-caption pairs):
| BLEU-4 | METEOR | ROUGE-L | CIDEr | SPICE |
|---|---|---|---|---|
| 0.0445 | 0.1149 | 0.2455 | 0.3422 | 0.1667 |
Three-seed aggregate results and exact split hashes are maintained in the CORA-360 repository.
git clone --branch v2.0.0-eccv2026 https://github.com/wooseungw/CORA-360.git
cd CORA-360
./reproduce.sh evaluate gemma3-4b-densecl /path/to/test.csv
CORA was evaluated on English query-focused captions from QuIC-360. Performance outside that domain, on non-ERP imagery, or in safety-critical settings is not established. This adapter is subject to the Gemma Terms of Use in addition to the CORA source-code license.
@inproceedings{woo2026cora,
title={Overlap-Consistent View Decomposition for Adapting Vision--Language Models to 360-Degree Panoramas},
author={Woo, Seungwoo and Jung, Daewon and Youm, Sekyoung},
booktitle={European Conference on Computer Vision},
year={2026}
}