modification text
Reference image + modification text → top-10 from the validation gallery. Green = ground-truth target. FashionIQ: [EOS] retrieval vs Efficient rerank ([EOS] top-100 re-scored by cos([EOS]) + cos(mean latents)), Qwen2.5-VL-7B backbone. CIRR: Efficient rerank of three FactEmbed backbones side by side.