SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2501.01986 · 2025

FrameFusion: Combining Similarity and Importance for Video Token Reduction on Large Vision Language Models

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 13 functions out of this paper's own repositories and ran 10 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
thu-nics/framefusion canonical 10 of 13
FunctionStatusWhere it lives
repeat_kv Ran thu-nics/framefusion/framefusion/models/internvl/modeling_internlm2.py
code served (permissive licence) · get_code("3c76e52815c5401d")
Qwen2DecoderLayer_merge_then_prune_by_cost_forward Ran thu-nics/framefusion/framefusion/models/qwen2/modeling_qwen2.py
code served (permissive licence) · get_code("12a855af63d45bc2")
Qwen2VLDecoderLayer_fastv_forward Ran thu-nics/framefusion/framefusion/models/qwen2/modeling_qwen2_vl_baseline.py
code served (permissive licence) · get_code("682eafc6f1711f42")
Qwen2VLDecoderLayer_merge_then_fastv_cost_given_forward Ran thu-nics/framefusion/framefusion/models/qwen2/modeling_qwen2_vl.py
code served (permissive licence) · get_code("3153f92d89ff16c1")
Qwen2VLModel_merge_then_fastv_cost_given_forward Ran thu-nics/framefusion/framefusion/models/qwen2/modeling_qwen2_vl.py
code served (permissive licence) · get_code("5661336a3d308111")
compute_density_overhead Ran thu-nics/framefusion/framefusion/models/qwen2/modeling_qwen2_baseline.py
code served (permissive licence) · get_code("264c5b9a2f25be7e")
get_attr_by_name Ran thu-nics/framefusion/framefusion/utils.py
code served (permissive licence) · get_code("ccac3bd0c453b847")
get_vllm_embedding Ran thu-nics/framefusion/framefusion/models/minicpmv/modeling_minicpmv.py
code served (permissive licence) · get_code("6f3669639f7ffa02")
rotate_half Ran thu-nics/framefusion/framefusion/models/internvl/modeling_internlm2.py
code served (permissive licence) · get_code("e03d53ba9d4f9ae5")
scaled_dot_product_attention Ran thu-nics/framefusion/framefusion/utils.py
code served (permissive licence) · get_code("2e002934a51f4e51")
Qwen2VLModel_fastv_forward Not yet run thu-nics/framefusion/framefusion/models/qwen2/modeling_qwen2_vl_baseline.py
code served (permissive licence) · get_code("18116389c0c9cc0c")
apply_rotary_pos_emb Not yet run thu-nics/framefusion/framefusion/models/internvl/modeling_internlm2.py
code served (permissive licence) · get_code("c028aebfc37921e4")
generate Not yet run thu-nics/framefusion/framefusion/models/internvl/modeling_internvl_chat.py
code served (permissive licence) · get_code("421deb414feba11c")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

The increasing demand to process long and high-resolution videos significantly burdens Large Vision-Language Models (LVLMs) due to the enormous number of visual tokens. Existing token reduction methods primarily prune tokens based on importance metrics, such as cumulative attention scores. However, even important tokens may exhibit high redundancy caused by similarity among adjacent video frames and repetitive visual elements. To address this limitation, we propose FrameFusion, a novel token reduction approach integrating similarity-based merging with importance-based pruning. We conduct a thorough study on token similarity characteristics, revealing three key insights: (1) spatially corresponding visual tokens between adjacent frames have higher cosine similarities compared to other token pairs; (2) high token similarities prominently decrease in deeper model layers; and (3) token similarity rankings are highly consistent across different layers. Guided by these observations, FrameFusion computes token similarities exclusively between corresponding visual tokens from adjacent frames, applies token merging at initial successive layers followed by pruning in deeper layers, and adopts a cascaded merging strategy to further enhance efficiency. We evaluate FrameFusion comprehensively across six diverse LVLMs, ranging from 2B to 72B parameters, using five video benchmarks encompassing video retrieval, question-answering, and spatial-temporal understanding tasks. Experiments show that FrameFusion reduces visual tokens by 70%, achieving 1.6-3.6x end-to-end speedups, with an average performance impact of less than 3%. Our code is available at: https://github.com/thu-nics/FrameFusion.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2501.01986")
get_code_for_paper("2501.01986")
have("2501.01986")

Connect an agent — have() is free.