SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2607.03612 · 2026

SAF3R: Dynamic Sparse Attention for Feed-Forward 3D Reconstruction Transformers

Tianlong Chen, Huanrui Yang, Jialu Wang, Song Wang, Yuanzhe Li, Jianing Deng, Jingtong Hu, Merge Split

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 12 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
jndeng/SAF3R canonical 10 of 12
FunctionStatusWhere it lives
broadcast_anchorframe_attention Ran jndeng/SAF3R/saf3r/patch/sparse_attention.py
pointer only (licence: NONE) · get_code("352f830a936276d7")
compute_qk_topk_indices_batched Ran jndeng/SAF3R/saf3r/patch/utils.py
pointer only (licence: NONE) · get_code("54ee3709aac5da15")
get_separate_indices Ran jndeng/SAF3R/saf3r/patch/sparse_attention.py
pointer only (licence: NONE) · get_code("f4b1536fd995efc5")
getattr_nested Ran jndeng/SAF3R/saf3r/patch/utils.py
pointer only (licence: NONE) · get_code("3397d0878a853bf9")
load_config Ran jndeng/SAF3R/saf3r/models/official_da3/cfg.py
pointer only (licence: NONE) · get_code("61fc3feab9b77a2e")
rel2abs Ran jndeng/SAF3R/saf3r/patch/patch_da3.py
pointer only (licence: NONE) · get_code("357cd1b8cc2b4d9a")
rel2abs Ran jndeng/SAF3R/saf3r/patch/patch_pi3.py
pointer only (licence: NONE) · get_code("3c7eb2d7ebaabac4")
resolve_inheritance Ran jndeng/SAF3R/saf3r/models/official_da3/cfg.py
pointer only (licence: NONE) · get_code("e670cca85ad64e35")
resolve_recursive Ran jndeng/SAF3R/saf3r/models/official_da3/cfg.py
pointer only (licence: NONE) · get_code("a09fd8ce00f2afb2")
sample_frames Ran jndeng/SAF3R/saf3r/profile_models.py
pointer only (licence: NONE) · get_code("eec68b77b0f0e53d")
build_model Not yet run jndeng/SAF3R/saf3r/utils/model_utils.py
pointer only (licence: NONE) · get_code("25c975b9fde6e203")
sparse_attention Not yet run jndeng/SAF3R/saf3r/patch/sparse_attention.py
pointer only (licence: NONE) · get_code("dcd87d59a89a40d5")

Repositories linked to this paper

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

Abstract

Feed-forward 3D reconstruction (F3R) transformers have recently achieved remarkable success. However, scaling them to long image sequences remains challenging, as the quadratic complexity of cross-view global attention quickly becomes the dominant computational bottleneck. While recent efforts attempt to improve efficiency through compressed or sparse attention, they fail to fully exploit the inherent sparsity and dynamic behavior of global attention. In this work, we present a comprehensive analysis of global attention across multiple F3R transformers and reveal that attention patterns are highly heterogeneous, dynamic, and extremely sparse across layers and attention heads. Motivated by these findings, we propose SAF3R, a training-free dynamic sparse attention framework tailored to F3R transformers. SAF3R integrates tailored sparse attention mechanisms with offline head profiling and an efficient online adaptation strategy to match input-dependent attention behaviors. Extensive experiments demonstrate that SAF3R achieves high sparsity ratios while preserving camera pose estimation and 3D reconstruction quality, translating into substantial end-to-end speedup on F3R transformers compared to existing methods. Code is available at https://github.com/jndeng/SAF3R.

For agents

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

get_harvested_code_for_paper("2607.03612")
get_code_for_paper("2607.03612")
have("2607.03612")

Connect an agent — have() is free.