AI Paper Atlas

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

Exploring Phrase-Level Grounding with Text-to-Image Diffusion Model

Danni Yang, Ruohan Dong, Jiayi Ji, Yiwei Ma, Haowei Wang, Xiaoshuai Sun, Rongrong Ji

ECCV 2024 · Image generation › Text-to-image diffusion · approach: Compositional text-to-image guidance · 4 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 Locate-to-Segment Processor (self-attention aggregation from cross-attention anchors) (ablation)DiffPNG cross-attention baseline · PNG val (zero-shot)Average Recall overall: 30.24 -> 33.54Improved
add SAM-based Mask Refinement (ablation)DiffPNG · PNG val (zero-shot)Average Recall overall: 33.54 -> 38.49Improved
add DenseCRF refinement (ablation)DiffPNG · PNG val (zero-shot)Average Recall overall: 33.54 -> 23.49Hurt
select anchors by threshold instead of top-K percentage (ablation)DiffPNG · PNG val (zero-shot)Average Recall overall: 30.13 -> 33.54Improved
use top-K percentage anchors rather than cross-attention alone (ablation)cross-attention baseline · PNG val (zero-shot)Average Recall overall: 30.24 -> 30.13No clear effect
use cross-attention resolution 16 rather than 32 or 64 (ablation)DiffPNG · PNG val (zero-shot)Average Recall overall: 16.54 -> 33.54Improved
use subject-focused aggregator instead of average aggregator (ablation)DiffPNG · PNG val (zero-shot)Average Recall overall: 32.40 -> 33.27Improved
use subject-focused aggregator instead of multiplication aggregator (ablation)DiffPNG · PNG val (zero-shot)Average Recall overall: 28.67 -> 33.27Improved
aggregate cross-attention rather than self-attention-enhanced maps (ablation)DiffPNG · PNG val (zero-shot)Average Recall overall: 33.27 -> 33.54Improved
use DatasetDiffusion or DiffSeg style attention merging instead of cross-attention only (ablation)cross-attention baseline · PNG val (zero-shot)Average Recall overall: 30.24 -> 23.46 and 24.08Hurt
tune binarization threshold alpha to 0.3 instead of 0.5 (ablation)DiffPNG LSP output · PNG val (zero-shot)Average Recall overall: 28.20 -> 33.54Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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