UFO: Chain-of-Evaluation for Omni-Condition Alignment in Multi-Modal Image Generation
Danning Zhang, Yijing Lin, Shuhan Zhuang, Mengqi Huang, Shaojin Wu, Shancheng Fang, Zhendong Mao
University of Science and Technology of China · Independent Researcher · Shenzhen University · Institute of Artificial Intelligence
Published September 17, 2026 · on the feed September 18, 2026 · ▲ 43 · arXiv 2609.12397
The abstract
Multi-modal image generation, particularly subject-driven customization, has garnered growing attention in recent years. Despite the rapid advancement of generative models, their evaluation remains largely lagging. Existing methods, whether embedding-based or Multi-modal Large Language Model (MLLM)-based, evaluate alignment with each modal condition in isolation, which contradicts the simultaneous condition alignment objective of multi-modal image generation, leading to poor consistency with human judgments. To address this challenge, we propose UFO, the first unified framework for omni-condition alignment simultaneous evaluation. Specifically, UFO introduces a novel Atomized Chain-of-Evaluation paradigm, i.e., it first decomposes omni-condition alignment into a sequential chain of fine-grained, disentangled Atomic Evaluation Units (AEUs), categorizes them into distinct modality-relevance classes, and then employs general or dedicated functional calls for accurate verification of different AEU types. Experimental results demonstrate that UFO achieves the highest correlation with human evaluation preferences, delivering an average improvement of 15.25%. Furthermore, we present UFO-Bench, a dedicated benchmark designed to holistically evaluate the performance of existing customization models under the diverse mutual interactions of textual and visual conditions.
The Brief
2:18A short podcast summary of the paper. Written and voiced by AI.
Transcript
UFO. Chain-of-Evaluation for Omni-Condition Alignment in Multi-Modal Image Generation. By Danning Zhang and six others, from the University of Science and Technology of China, Independent Researcher, and two other institutions. Published on the seventeenth of September. Its code repository is worked out: about forty-five source files, with a readme.
Multi-modal image generation, and subject-driven customization in particular, has drawn growing attention, but its evaluation has lagged behind. Existing evaluators, whether based on embeddings or on multimodal large language models, check alignment with each condition in isolation: the text prompt on its own, the reference subject on its own. Yet the task asks for all conditions to be satisfied at once. Scoring them separately contradicts that objective, and the result is poor consistency with human judgments. An image can score well against the text and well against the reference, each on its own, and still fail to satisfy both together. A check that looks at one condition at a time has no way to see that.
UFO is presented as the first unified framework for evaluating alignment with all conditions simultaneously, which the authors call omni-condition alignment. Its core is an Atomized Chain-of-Evaluation paradigm.
The chain works in three moves. First, omni-condition alignment is decomposed into a sequential chain of fine-grained, disentangled Atomic Evaluation Units. Second, those units are sorted into distinct classes by which modality they concern. Third, each type is verified with a call suited to it, either a general one or a dedicated functional one. A single overall judgement becomes a series of small checks, each handled by the tool best placed to answer it.
UFO achieves the highest correlation with human evaluation preferences, with an average improvement of fifteen point two five percent. The authors also present UFO-Bench, a dedicated benchmark for evaluating existing customization models holistically, under the many ways that textual and visual conditions interact with each other. The benchmark carries the same premise over to testing the generators themselves: conditions that interact, rather than conditions scored one by one.
The framework's premise is that a customized image succeeds or fails as a whole. Breaking that whole into atomic checks, rather than scoring each input separately, is what brings the automatic judgement closer to a human one.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A generated image can look right twice over and still be wrong. Start with the judge's dilemma.
You are judging a model that takes a text prompt and photos of a subject, say a woman, and must draw her again as asked: same person, but with red hair.
The road ahead
- 1 · The problem ~7 min Scoring an image against the text and the reference photo separately misses whether it satisfies both at once, so the scores disagree with people.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Break the judgement into small check units, tag each with the condition that governs it, score each with a fitting tool, and combine with weights.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min The method runs four stages in order: decompose alignment into atomic units, sort them by modality, verify each with a general or dedicated call, then aggregate with weights.
- 6 · The experiment ~6 min You’ll guess: On average, by how much did the new evaluator improve correlation with human evaluation preferences over existing evaluators?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.