AI Paper Atlas

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

Glyph-ByT5: A Customized Text Encoder for Accurate Visual Text Rendering

Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, Yuhui Yuan

ECCV 2024 · Image generation › Text-to-image diffusion · approach: Personalized and region-controlled generation · 48 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 glyph-alignment pretraining (Glyph-ByT5-S) over plain ByT5-S (ablation)SDXL · VisualParagraphy (<=20 chars)word precision (%): 60.52 -> 92.58Improved
use character-aware ByT5-S instead of T5-L (ablation)SDXL · VisualParagraphy (>=100 chars)word precision (%): 26.09 -> 42.05Improved
scale glyph-text pairs from 500K to 1M (ablation)Glyph-ByT5-S · VisualParagraphy (>=100 chars)word precision (%): 83.17 -> 89.17Improved
use DINOv2 instead of CLIP as frozen visual encoder (ablation)Glyph-ByT5 pretraining · VisualParagraphy ablation (>=100 chars)word precision (%): 66.34 -> 73.29Improved
use box-level contrastive loss instead of image-level (ablation)Glyph-ByT5 pretraining · VisualParagraphy ablation (>=100 chars)word precision (%): 69.42 -> 73.29Improved
add image-level loss on top of box-level loss (ablation)Glyph-ByT5 pretraining · VisualParagraphy ablation (>=100 chars)word precision (%): 73.29 -> 70.54Hurt
add glyph augmentation (1:16 ratio) (ablation)Glyph-ByT5 pretraining · VisualParagraphy ablation (>=100 chars)word precision (%): 65.40 -> 73.29Improved
raise glyph augmentation ratio from 1:16 to 1:32 (ablation)Glyph-ByT5 pretraining · VisualParagraphy ablation (>=100 chars)word precision (%): 73.29 -> 72.16No clear effect
add ByT5-to-SDXL mapper (ablation)Glyph-SDXL · VisualParagraphy ablation (>=100 chars)word precision (%): 65.02 -> 73.29Improved
scale text encoder from Glyph-ByT5-S to B and L (ablation)Glyph-SDXL · VisualParagraphy ablation (>=100 chars)word precision (%): 73.29 -> 73.72No clear effect
use region-wise cross-attention instead of concatenating text embeddings (ablation)Glyph-SDXL · VisualParagraphy ablation (>=100 chars)word precision (%): 3.91 -> 73.29Improved
use Glyph-ByT5 instead of ControlNet-style glyph conditioning (ablation)SDXL · VisualParagraphy (>=100 chars)word precision (%): 32.75 -> 73.29Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 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