AI Paper Atlas

Research map / Image generation / Text-to-image diffusion

Unleashing Text-to-Image Diffusion Models for Visual Perception

Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, Jiwen Lu

ICCV 2023 · Image generation › Text-to-image diffusion · approach: Compositional text-to-image guidance · 250 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add class-name text prompt (ablation)VPD · ADE20K (8K iters)mIoU: 46.9 -> 47.1Improved
add class-name text prompt (ablation)VPD · ADE20K (4K iters)mIoU: 41.5 -> 42.0Improved
add text adapter after CLIP text encoder (ablation)VPD · ADE20K (4K iters)mIoU: 42.0 -> 42.9Improved
add text adapter after CLIP text encoder (ablation)VPD · ADE20K (8K iters)mIoU: 47.1 -> 48.0Improved
add cross-attention maps as semantic guidance (up+down) (ablation)VPD · ADE20K (8K iters)mIoU: 48.0 -> 48.7Improved
use cross-attention maps from middle blocks (ablation)VPD · ADE20K (8K iters)mIoU: 48.0 -> 47.8Hurt
use cross-attention maps from up blocks rather than down blocks (ablation)VPD · ADE20K (8K iters)mIoU: 48.2 -> 48.5Improved
combine up and down cross-attention maps (ablation)VPD · ADE20K (8K iters)mIoU: 48.5 -> 48.7Improved
use cross-attention maps of larger size (A64 vs A32) at 4K iterations (ablation)VPD · ADE20K (4K iters)mIoU single-scale: 43.1 -> 43.9Improved
use A32 instead of A64 maps at 8K iterations (ablation)VPD · ADE20K (8K iters)mIoU single-scale: 47.7 -> 48.7Improved
more pre-trained iterations of Stable-Diffusion (SD-1-1 to SD-1-5) (ablation)VPD · ADE20K (8K iters)mIoU: improved by more than 4Improved
swap pre-trained backbone for VPD under 1-epoch fast schedule (vs LAVT)LAVT · RefCOCO valoIoU: 52.56 -> 63.04Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy