SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2503.19903 · CVPR · 2025

Scaling Vision Pre-Training to 4K Resolution

Jan Kautz, U Berkeley, Song Han, Trevor Darrell, Han Cai, Baifeng Shi, Marco Pavone, Sifei Liu, Yao Lu, Boyi Li, Hongxu Yin, Pavlo Molchanov, and 1 more

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 14 functions out of this paper's own repositories and ran 11 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
efficient-large-model/vila canonical 10 of 12
nvlabs/vila — 1 of 2
FunctionStatusWhere it lives
check_dot_in_model_path Ran efficient-large-model/vila/llava/remote_code/modeling_vila.py
code served (permissive licence) · get_code("8dd6b65f72c07727")
repeat_kv Ran efficient-large-model/vila/llava/eval/vision_niah_vila/zigzag_ring_attn/modeling_qwen2.py
code served (permissive licence) · get_code("30d7eec482ebf6b1")
VisionTower Ran nvlabs/vila/llava/model/multimodal_encoder/ps3_encoder.py
code served (permissive licence) · get_code("232455d9c7b1d476")
apply_rotary_pos_emb Ran efficient-large-model/vila/llava/eval/vision_niah_vila/zigzag_ring_attn/modeling_qwen2.py
code served (permissive licence) · get_code("5d81bffdb6022427")
get_vila_version Ran efficient-large-model/vila/llava/remote_code/modeling_vila.py
code served (permissive licence) · get_code("554c742ab091b991")
load_haystack Ran efficient-large-model/vila/llava/eval/vision_niah_vila/eval_vision_niah.py
code served (permissive licence) · get_code("70759df34aae94c5")
load_image Ran efficient-large-model/vila/server.py
code served (permissive licence) · get_code("1c69e761fdf05a26")
load_video Ran efficient-large-model/vila/server.py
code served (permissive licence) · get_code("fca9a6d2afa23288")
rotate_half Ran efficient-large-model/vila/llava/eval/vision_niah_vila/zigzag_ring_attn/modeling_qwen2.py
code served (permissive licence) · get_code("b99eea6376d1e212")
safe_tokenize Ran efficient-large-model/vila/llava/eval/vision_niah_vila/eval_vision_niah.py
code served (permissive licence) · get_code("300bb78b66d35115")
with_opencv Ran efficient-large-model/vila/data_prepare/panda_split.py
code served (permissive licence) · get_code("7af38f0a2846676a")
PS3VisionTower Not yet run nvlabs/vila/llava/model/multimodal_encoder/ps3_encoder.py
code served (permissive licence) · get_code("ef6066f5b1710c5d")
build_mm_projector Not yet run efficient-large-model/vila/llava/remote_code/modeling_vila.py
code served (permissive licence) · get_code("aa7f4da9de0dc7cc")
get_literal_values Not yet run efficient-large-model/vila/server.py
code served (permissive licence) · get_code("f8e6d235cb12b138")

Repositories linked to this paper

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

Abstract

Figure 1. Left: Regular vision models such as SigLIP [44] processes images at a low resolution (e.g., 378 × 378 pixels), which is not enough for many daily tasks such as spotting the stop sign while driving. In contrast, PS3 is able to both encode low-res features and efficiently process high-res information of 4K-resolution images via top-down patch selection, i.e., selectively processing relevant patches based on any text prompt. Top Right: SigLIP is pre-trained by contrasting global vision features and global captions, which is costly for high-resolution images. PS3 is pre-trained with additional contrast between local high-res features with local captions, enabling pre-training at 4K resolution with 79× less cost than SigLIP. Bottom Right: VILA-HD uses PS3 to selectively process high-res regions based on the user prompt, outperforming state-of-the-art MLLMs such as Qwen2-VL [38] on the proposed 4KPro benchmark while achieving 2.96× speedup.

For agents

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

get_harvested_code_for_paper("2503.19903")
get_code_for_paper("2503.19903")
have("2503.19903")

Connect an agent — have() is free.