LLaVA-Alternating-Attn

Alternating Attention Models:

Model Name	Base Model	Description
llava-alternating-attn-within-modality-qwen2-0.5b-ov	lmms-lab/llava-onevision-qwen2-0.5b-ov	Alternating attention architecture that restricts attention within modalities in alternating layers of the transformer blocks.
llava-alternating-attn-cross-modality-qwen2-0.5b-ov	lmms-lab/llava-onevision-qwen2-0.5b-ov	Alternating attention architecture that restricts attention to self-only and other modalities in alternating layers of the transformer blocks.

Instruction Tuned Models for Visual Correspondence Task:

Model Name	Base Model	Description
llava-alternating-attn-within-modality-qwen2-0.5b-ov-instructiontuned-visualcorres	llava-alternating-attn-within-modality-qwen2-0.5b-ov	Instruction-tuned variant of the within-modality alternating attention model, fine-tuned specifically for visual correspondence tasks (matching corresponding regions across images).
llava-alternating-attn-cross-modality-qwen2-0.5b-ov-instructiontuned-visualcorres	llava-alternating-attn-cross-modality-qwen2-0.5b-ov	Instruction-tuned variant of the cross-modality alternating attention model, fine-tuned specifically for visual correspondence tasks (matching corresponding regions across images).
llava-onevision-qwen2-0.5b-ov-instructiontuned-visualcorres	lmms-lab/llava-onevision-qwen2-0.5b-ov	Instruction-tuned baseline model (standard LLaVA-OneVision architecture) fine-tuned for visual correspondence tasks, provided for comparison with alternating attention variants.