ABaldrati/CLIP4Cir

Implements a two-stage training pipeline: task-oriented fine-tuning of CLIP's vision and text encoders using contrastive loss, followed by training a learnable Combiner network that fuses multimodal features through adaptive weighting and residual composition. Targets composed image retrieval on FashionIQ and CIRR datasets, where users query by combining a reference image with textual modifications, and achieves state-of-the-art results by bridging the gap between CLIP's general pre-training and task-specific requirements.

192 stars. No commits in the last 6 months.