#ModelScorePaper
1Full framework (Ours)79.40link
2LoRA (default)79.40link
3w/o QA fidelity scoring79.10link
4w/o cycle-consistent captioning79.00link
5w/o prompt perturbation (temp. sampling)78.80link
6QLoRA (4-bit)78.60link
7w/o self-consistency (STE only)78.30link
8SFT (self-generated data)78.10link
9w/o STE (self-consistency only)77.90link
10BLIP3o-8B (base)77.50link
11Full fine-tuning77.00link
12ViT-SO400M (pretrained)56.80link
13SBER-MoVQGAN-270M34.10link
SEED seed Leaderboard