@inproceedings{you-etal-2026-formalimg,
title = "{F}ormal{I}m{G}: Evaluating Structural Compositional Generalization for {T}2{I} Models",
author = "You, Hong-Jie and
Shao, Jie-Jing and
Yang, Xiao-Wen and
Wu, Zhi-Fan and
Jia, Lin-Han and
Guo, Lan-Zhe and
Li, Yu-Feng",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1243/",
pages = "26308--26342",
ISBN = "979-8-89176-462-0",
abstract = "As natural language becomes a key interface for controlling text-to-image generation, users specify not only which visual concepts should appear but also how entities are constrained by attributes, relations, quantifiers, and cross-entity dependencies. Reliable controllable generation therefore requires generalization over semantic structure, not merely over novel combinations of concepts. Existing benchmarks mainly evaluate conceptual compositional generalization, while leaving the internal semantic structure of instructions underexplored. We study structural compositional generalization: whether text-to-image models can generalize across increasingly coupled semantic dependencies. We introduce FormalImG, a first-order-logic-based benchmark with 4,000 instructions across three complexity levels in two scenarios, together with computable structural compositional complexity and {\ensuremath{\varepsilon}}-structural compositional generalizability. Experiments on advanced text-to-image models show consistent degradation as structural complexity increases; even the strongest evaluated model generalizes only up to $G_{0.7} = 6$ in Natural and $G_{0.7} = 4$ in Knolling. Further analysis suggests that the main bottleneck lies in language-to-vision instantiation and can be partially mitigated by explicit intermediate layouts. The benchmark and code are available at https://github.com/yhj137/FormalImG."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="you-etal-2026-formalimg">
<titleInfo>
<title>FormalImG: Evaluating Structural Compositional Generalization for T2I Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hong-Jie</namePart>
<namePart type="family">You</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jie-Jing</namePart>
<namePart type="family">Shao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiao-Wen</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhi-Fan</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lin-Han</namePart>
<namePart type="family">Jia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lan-Zhe</namePart>
<namePart type="family">Guo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yu-Feng</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>As natural language becomes a key interface for controlling text-to-image generation, users specify not only which visual concepts should appear but also how entities are constrained by attributes, relations, quantifiers, and cross-entity dependencies. Reliable controllable generation therefore requires generalization over semantic structure, not merely over novel combinations of concepts. Existing benchmarks mainly evaluate conceptual compositional generalization, while leaving the internal semantic structure of instructions underexplored. We study structural compositional generalization: whether text-to-image models can generalize across increasingly coupled semantic dependencies. We introduce FormalImG, a first-order-logic-based benchmark with 4,000 instructions across three complexity levels in two scenarios, together with computable structural compositional complexity and \ensuremathǎrepsilon-structural compositional generalizability. Experiments on advanced text-to-image models show consistent degradation as structural complexity increases; even the strongest evaluated model generalizes only up to G₀.7 = 6 in Natural and G₀.7 = 4 in Knolling. Further analysis suggests that the main bottleneck lies in language-to-vision instantiation and can be partially mitigated by explicit intermediate layouts. The benchmark and code are available at https://github.com/yhj137/FormalImG.</abstract>
<identifier type="citekey">you-etal-2026-formalimg</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1243/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>26308</start>
<end>26342</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T FormalImG: Evaluating Structural Compositional Generalization for T2I Models
%A You, Hong-Jie
%A Shao, Jie-Jing
%A Yang, Xiao-Wen
%A Wu, Zhi-Fan
%A Jia, Lin-Han
%A Guo, Lan-Zhe
%A Li, Yu-Feng
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F you-etal-2026-formalimg
%X As natural language becomes a key interface for controlling text-to-image generation, users specify not only which visual concepts should appear but also how entities are constrained by attributes, relations, quantifiers, and cross-entity dependencies. Reliable controllable generation therefore requires generalization over semantic structure, not merely over novel combinations of concepts. Existing benchmarks mainly evaluate conceptual compositional generalization, while leaving the internal semantic structure of instructions underexplored. We study structural compositional generalization: whether text-to-image models can generalize across increasingly coupled semantic dependencies. We introduce FormalImG, a first-order-logic-based benchmark with 4,000 instructions across three complexity levels in two scenarios, together with computable structural compositional complexity and \ensuremathǎrepsilon-structural compositional generalizability. Experiments on advanced text-to-image models show consistent degradation as structural complexity increases; even the strongest evaluated model generalizes only up to G₀.7 = 6 in Natural and G₀.7 = 4 in Knolling. Further analysis suggests that the main bottleneck lies in language-to-vision instantiation and can be partially mitigated by explicit intermediate layouts. The benchmark and code are available at https://github.com/yhj137/FormalImG.
%U https://aclanthology.org/2026.findings-emnlp.1243/
%P 26308-26342
Markdown (Informal)
[FormalImG: Evaluating Structural Compositional Generalization for T2I Models](https://aclanthology.org/2026.findings-emnlp.1243/) (You et al., Findings 2026)
ACL