Cross-subject view-aligned generation on Objaverse-XL 1k-pair (test)
15.8PSNRRAVA
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| RAVAstrategy=Retrieval, backbone=fine-tuned Flux.2, K=32026.06 | 15.8 | 0.8398 | 0.1829 | |
| Top-Kstrategy=Retrieval, backbone=fine-tuned Flux.2, K=32026.06 | 15.4 | 0.8308 | 0.2009 | |
| CLIPstrategy=Retrieval, backbone=fine-tuned Flux.2, K=32026.06 | 15.17 | 0.8278 | 0.2091 | |
| DINOv3strategy=Retrieval, backbone=fine-tuned Flux.2, K=32026.06 | 15.03 | 0.8282 | 0.2104 | |
| Randomstrategy=Retrieval, backbone=fine-tuned Flux.2, K=32026.06 | 15.01 | 0.8249 | 0.2143 | |
| Flux.2mode=Zero-shot, input=one anchor image of subject A and one reference image of subject B2026.06 | 13.22 | 0.792 | 0.3042 | |
| Qwen-Imagemode=Zero-shot, input=one anchor image of subject A and one reference image of subject B2026.06 | 10 | 0.6915 | 0.4726 | |
| UNOmode=Zero-shot, input=one anchor image of subject A and one reference image of subject B2026.06 | 9.79 | 0.6866 | 0.5435 | |
| OmniGenmode=Zero-shot, input=one anchor image of subject A and one reference image of subject B2026.06 | 8.03 | 0.5756 | 0.6167 | |
| MS-Diffusionmode=Zero-shot, input=one anchor image of subject A and one reference image of subject B2026.06 | 7.26 | 0.6214 | 0.7869 |