Paper deep dive

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

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 48

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 9/16/2026, 8:37:39 AM

Summary

The paper introduces UFO, a unified framework for evaluating omni-condition alignment in multi-modal image generation, specifically subject-driven customization. It proposes an Atomized Chain-of-Evaluation paradigm that decomposes alignment into Atomic Evaluation Units (AEUs), categorizes them by modality relevance, and uses VQA or function calls for verification. The authors also present UFO-Bench, a benchmark with diverse interaction scenarios, demonstrating that UFO achieves higher correlation with human judgment than existing metrics.

Entities (10)

UFO · framework · 98%UFO-Bench · benchmark · 97%Atomized Chain-of-Evaluation · methodology · 96%Atomic Evaluation Units (AEUs) · concept · 95%Multi-modal Large Language Model (MLLM) · technology · 92%CLIP · model · 90%DINO · model · 90%VIEScore · model · 90%ArcFace · model · 88%DreamBench++ · benchmark · 85%

Relation Signals (7)

UFO → uses → Atomized Chain-of-Evaluation

confidence 98% · Specifically, UFO introduces a novel Atomized Chain-of-Evaluation paradigm

Atomized Chain-of-Evaluation → decomposes → Atomic Evaluation Units (AEUs)

confidence 97% · it first decomposes omni-condition alignment into a sequential chain of fine-grained, disentangled Atomic Evaluation Units (AEUs)

UFO-Bench → evaluates → subject-driven customization models

confidence 96% · UFO-Bench, a dedicated benchmark designed to holistically evaluate the performance of existing customization models

UFO → outperforms → existing metrics

confidence 95% · delivering an average improvement of 15.25%

CLIP → istypeof → embedding-based evaluator

confidence 90% · Traditional embedding-based evaluators, such as CLIP [7]

VIEScore → istypeof → MLLM-based evaluator

confidence 90% · MLLM-based evaluators, such as VIEScore [12]

ArcFace → usedfor → identity verification

confidence 88% · leveraging ArcFace [5] as a dedicated functional call for high-precision identity verification

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

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, \emph{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.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

44,391 characters extracted from source content.

Expand or collapse full text

arXiv:2609.12397v1 [cs.CV] 11 Sep 2026 UFO:Chain‑of‑Evaluation for Omni‑Condition Alignment in Multi‑Modal Image Generation Danning Zhang 1,* Yijing Lin 1,* Shuhan Zhuang 1,* Mengqi Huang 1, Shaojin Wu 2, Shancheng Fang 3 Zhendong Mao 1,4 1 University of Science and Technology of China 2 ByteDance Seed 3 Shenzhen University 4 Institute of Artificial Intelligence * Equal contributionCorresponding authorProject lead 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 proposeUFO, the firstUniFied framework forOmni‑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. Codegithub.com/UFOTeamwork/UFO Hugging Facehuggingface.co/UFOTeamwork/UFO‑Bench 1 Introduction Multi‑modal image generation [1,3,15,29,31] has achieved significant progress in recent years, driven by advances in large‑scale diffusion models [13,18]. These models are capable of generating high‑quality visual content based on diverse cross‑modal conditions, including textual conditions and visual conditions from reference images. The advancements have unlocked a plethora of creative applications, spanning personalized content generation, visual storytelling,etc.As generative models become increasingly expressive and controllable, reliable evaluation has emerged as a critical bottleneck for both research and real‑world deployment. Among various multi‑modal image generation tasks, subject‑driven customization [9,16,25,26] has drawn growing attention in recent years. It takes a text and images of a specific subject as input, aiming to generate new images that align with bothtextual semanticsand thesubject’s visual attributes that are not edited by the text. In contrast to other text‑image conditioned generation, such as image editing [10,20], subject‑driven customization requires regenerating the target visual subject in afree‑form manner, i.e., the generated subject may vary in pose, position,etc., making it more challenging to conduct accurate evaluations. Specifically, such variations not only make it difficult to judge the subject’s visual alignment, but also further complicate the assessment of its adherence to both textual and visual alignment simultaneously. Existing evaluation for subject‑driven customization can be categorized into two streams,i.e., the traditional embedding‑based evaluators [ 8,23] and Multi‑modal Large Language Model (MLLM)‑based evaluators [6,19,27]. Traditional embedding‑based evaluators, such as CLIP [7] and DINO [32], calculate the feature similarity between generated and reference images and assume that higher visual similarity implies better subject preservation. MLLM‑based evaluators, such as VIEScore [12] and 1 DINO/VIE-Score :Good HumanJudgement:Bad reference image prompt “The person withredhair fishingby the ocean.” False Positive Ignoring "red hair" in reference text DINO/VIE-Score :Good Human Judgement: Bad DINO/VIE-Score :Bad Human Judgement: Good False Positive Missing braidin reference image False Negative Penalizing validedit Figure 1Limitations of isolated evaluation in subject‑driven image generation. The top two rows show that existing evaluators fail to detect errors (ignoring the “red hair” prompt or missing the “braid”). The bottom row shows that evaluators penalize the successful edit that is correctly rated by humans. Dreambench++ [19], typically query the MLLM to assign a holistic score for evaluating textual and visual alignment. The common limitation of existing evaluators is that they all evaluate alignment with each modal conditionin isolation. In this study, we argue that the existing isolated evaluation paradigm is inherentlycontractive to the simultaneous condition alignment objectiveof multi‑modal image generation, resulting in poor consistency with human evaluation, and thereby substantially restricts the advancement of subject‑driven customization research. The reason for this is that subject‑driven customization requires generated subjects to adhere to both textual and visual conditions, where textual conditions invariably involve modifications to specific attributes of the subject in the original reference image. Isolated evaluation, by contrast, only compares the visual appearance of subjects in reference and generated images, and thus ignores the intended modifications specified by textual conditions, ultimately leading to false positive and false negative issues. As shown in Figure 1, existing evaluators assign high scores to generated images that closely resemble the reference subject even when they ignore textual requirements (e.g., failing to change hair color), resulting in false positives. Furthermore, they fail to capture fine‑grained identity modifications, leading to other false positives. Conversely, valid generations that correctly apply the textual edit while preserving the subject’s identity are often penalized due to reduced visual similarity, leading to false negatives. To address this problem and establish a reliable, automatic, and reusable evaluation paradigm with benchmarks for effective measurement of the progress in subject‑driven customization, we propose UFO, aUniFied framework forOmni‑condition alignment simultaneous evaluation, presented here for the first time, along withUFO‑Bench, which encompasses diverse multi‑modal conditions’ interaction scenarios to comprehensively assess the performance of existing customization models. Specifically, to tackle the challenges posed by the diverse potential interactions between distinct modality conditions (i.e., textual and visual conditions), UFO introduces a novelAtomized Chain‑of‑Evaluation framework. This framework first decomposes omni‑condition alignment into a sequential chain of fine‑grained, disentangledatomic evaluation units (AEUs)across two hierarchical levels,i.e., the local concrete component and global abstract attribute levels. Second, each AEU is categorized into a modality‑relevance class, specifying whether its alignment is governed by visual conditions, textual conditions, or their integrated dual‑modality combination. Next, UFO further quantifies the alignment score of each AEU either via general Visual Question Answering (VQA) queries or specialized function calls,e.g., leveraging ArcFace [5] as a dedicated functional call for high‑precision identity verification. Finally, all AEU‑level alignment scores are aggregated using adaptive importance weighting to yield an 2 interpretable holistic omni‑condition alignment score. Meanwhile, existing benchmarks for subject‑driven customization typically contain limited interactions between textual and visual modalities and few conflicting condition pairs, making them insufficient for comprehensive evaluation of complex multi‑modal alignment. To address this gap, we proposeUFO‑Bench, a dedicated benchmark tailored for subject‑driven generation under conflicting multi‑modal conditions. UFO‑Bench comprises 86 reference images across 7 major categories (e.g., humans, rigid objects, anime characters). Crucially, we design a hierarchical prompting strategy for each category, incorporating 3 tiers of editing difficulty with 81.97% conflicting condition pairs. Spanning from simple background shifts to complex edits that simultaneously modify local accessories and global styles, this strategy explicitly tests the ability of evaluation metrics to distinguish between text‑guided attribute modifications and core identity preservation. Our contributions are summarized as follows: •Conceptual contribution.We propose a novel formulation for multi‑modal image evaluation, defining consistency assessment as atomic, condition‑aware reasoning under free‑form multi‑modal conditioning. •Technical contribution.We introduce UFO, a chain‑of‑evaluation framework that integrates fine‑grained atomic decomposition, condition‑aware VLM evaluation, and expert tools into a unified metric. •Benchmark contribution.We construct UFO‑Bench, a comprehensive benchmark featuring diverse subject categories and hierarchical editing scenarios, addressing the lack of fine‑grained evaluation testbeds for subject‑driven generation. •Experimental contribution.Extensive experiments demonstrate that UFO achieves the highest correlation with human evaluation preferences, improving over existing metrics by 15.25% on average and showing particular advantages in complex scenarios. 2 Related Work 2.1 Subject‑Driven Image Generation Recent advances in subject‑driven image generation focus on creating unified models that enable diverse generation and editing tasks. For example, OmniGen2 [30] separates the decoder and image tokenizer, which supports text‑to‑image generation, image editing, and context‑aware synthesis. In parallel, UNO [31] applies a model‑data co‑evolution strategy to guide high‑resolution data generation and ensure semantically context‑consistent outputs. Likewise, the closed‑source Nano Banana [3] can jointly process text and image inputs within a single framework, thereby supporting both image generation and editing tasks. In the field of open‑source models, Qwen‑Image [29], built on the backbone of multimodal large language models, has achieved high‑quality and flexible subject‑aware generation and editing capabilities. BAGEL [4] introduces a Mixture‑of‑Transformer‑Experts (MoT) architecture that models multimodal understanding and generation separately. It enables collaboration through a shared self‑attention mechanism, improving cross‑modal interaction efficiency and generation consistency. 2.2 Subject‑Driven Image Generation Benchmarks In subject‑driven image generation research, early efforts mainly focused on benchmarks for general editing tasks. For instance, DreamEditBench [14] and ImagenHub [11] provide standardized datasets and protocols for subject‑driven editing, serving as foundational references. Subsequently, evaluations extended to multi‑task and 3D scenarios, such as DreamBooth3D [22], which adapts subject‑driven methods to geometrically consistent generation. Recognizing the limitations of metric‑based evaluations, recent works have shifted towards human‑aligned assessments. DreamBench++ [19] incorporates strong VLMs to approximate human judgment. To address the complexity of customization tasks, DSH‑Bench [28] introduces a hierarchical taxonomy with varying difficulty levels (easy/medium/hard), enabling more detailed performance diagnosis. Similarly, focusing on 3 Figure 2Overview of UFO‑Bench.The benchmark comprises 660 diverse cases spanning hierarchical categories—from rigid objects to full‑body characters—and organizes four editing paradigms (Non‑Editing, Global Editing, Local Editing, Complex Editing) into three tiers of editing difficulty. identity preservation, Beyond the Pixels [24] proposes using hierarchical VLM prompting (e.g., feature decision trees) to move beyond superficial similarity scores. However, significant gaps remain. Existing approaches typically evaluate textual and visual conditions in isolation, failing to capture the inherent conflict between identity preservation and complex semantic modifications. They lack a unified framework to reason about the trade‑offs in free‑form generation, whereas our approach is explicitly designed to assess these conflicting multi‑modal interactions through an atomized, condition‑aware paradigm. 2.3 Evaluation Metrics Objective Evaluation Metrics.Early automated evaluation metrics of text‑to‑image generation primarily focused on visual quality, such as the Inception Score (IS) [23] and the Frechet Inception Distance (FID) [8]. As research progressed, alignment between generated images and textual conditions became more important. Reference‑based metrics [17] and reference‑free embedding‑based approaches such as CLIPScore [ 7] are introduced to assess text–image consistency. Recently, self‑supervised vision representations, notably DINO [32], have been adopted for evaluation. These demonstrate improved sensitivity to object‑level and fine‑grained semantic differences in image generation and editing tasks. LLM‑based Evaluation.In recent years, research has gradually introduced large language models (LLMs) as evaluators.The representative method, VIEScore [12], uses a multimodal large language model to decompose semantic consistency and perceptual quality into fine‑grained sub‑criteria for itemized scoring. However, its evaluation remains limited to the image level. In contrast, DreamBench++ [19] focuses on personalized image generation tasks and explicitly separates evaluation into two independent dimensions: Image–Image Concept Preservation and Text–Image Prompt Following, which measure the retention of subject features and the quality of instruction execution, respectively. However, this evaluation paradigm relies on absolute scoring of individual‑generated results and still has certain limitations in capturing fine‑grained quality differences and complex human preferences. 3 Method In this section, we first present a brief analysis of the limitations of existing customization benchmarks, i.e., the insufficient variety of interaction types between textual and visual conditions, making them unable to evaluate the capability of existing customization models in handling multi‑modal conditional inputs. To address the gaps, we introduce UFO‑Bench in Section3.1, a dedicated benchmark designed to holistically evaluate the performance of existing customization models under the diverse mutual effects of textual and visual conditions. We then provide a detailed elaboration ofUFO, theUniFied Omni‑condition alignment evaluation framework in Section3.2. 4 Figure 3Overview of UFO‑Bench construction.We employ an iterative data preparation pipeline that utilizes LLMs for prompt template generation and caption adjustment to ensurehigh‑quality benchmark construction. Figure 4Distribution curves of atomic attribute complexity and corresponding case counts for different object categories in UFO‑Bench 3.1 UFO Benchmark A significant limitation of existing customization benchmarks lies in theirinsufficient variety of interaction typesbetween textual and visual conditions. Most existing benchmarks primarily focus on simple background shifts or singular attribute changes, failing to reflect the complexity offree‑form generationwhere models must navigateconflicting multi‑modal conditions. Specifically, they often overlook scenarios requiring the simultaneous execution of holistic style transformations and local interactive modifications. As illustrated in Figure2, this lack of diversity restricts the comprehensive evaluation of a model’s ability to maintain identity preservation while adhering to complex textual instructions. To address these gaps, we introduceUFO‑Bench, a dedicated benchmark designed to evaluate customization models under diverse and challenging mutual effects of textual and visual conditions. UFO‑Bench is constructed to evaluate the proficiency of generative models in balancing subject‑driven image generation with complex textual editing. As illustrated in Figure 2, the benchmark encompasses seven core subject categories—Rigid Object, Soft Object, Human, Full‑body Character, Animal, Logo, and Scene—featuring a balanced distribution of simple and complex samples. As shown in Figure3, we employ an iterative data preparation pipeline utilizing LLMs to ensure high‑quality, logically coherent prompts. The design of UFO‑Bench revolves around four primary editing paradigms:Non‑Editing, Local Editing,Global Editing, andComplex Editing, each meticulously tailored to assess distinct 5 Reference Image Atomic Decomposition Stage 1 <think> STEP 1: Reference Image Characteristics STEP 2: Edit Prompt Focus Analysis · STEP 5: Final Structural Decisions Global Attributes: gender, age_group, ...... Parts: clothing, accessories, face, grey backpack ...... </think> Spatial Atomic Aspects “A person is climbing a steel ladder while carrying a gray backpack, wearing a yellow raincoat, on the side of a red cargo ship in a busy harbor.” Prompt Stage 3 Atomic Consistency Evaluation Generated Image Global Atomic Attributes Function Call Similarity > 0.5 Similarity < 0.5 ArcFace General Evaluation σ Actomic_Score i ∗Weight i σ Weight i 0.8 <think> STEP 1: Input Analysis STEP 2: Dependency Logic Formulation STEP 3: Atomic-wise Classification Item: Face Prompt Check: Did prompt mention face/expression? Context Check ...... </think> image only text only text and image Relevance Analysis Stage 2 <think> STEP 1: Load Inputs from Stage 2 STEP 2: Select Question Templates STEP 3: Generate & Refine Questions Processing Item: Face/Identity Source: Reference. Drafting: Check identity preservation. Final Phrasing: Does the generated image match the reference image in terms of facial identity? </think> “Does the generated image match the reference image in terms of facial identity?” Atomic Scorescore=1score=0 Stage 4 Weighted Aggregation . . . . . . <think> STEP 1: Hierarchy Definition STEP 2: Intent Analysis STEP 3: Part-wise Weight Calculation Part : Gender Role: Basic Attribute. Reasoning: A fundamental check. Usually correlated with Face, but a distinct biological characteristic. Weight: 3. </think> Adaptive Important Weight T I T&I I “Does the generated image match the reference prompt in terms of bag?” 34 4 5 genderage group clothingaccessories facegrey backpack Figure 5Overall framework of UFO.UFO performs fine‑grained multi‑modal evaluation through a four‑stage pipeline. Firstly, atomic evaluation units (AEUs) are generated conditioned on the visual and textual inputs. Then each AEU is assigned a modality‑relevance class. Subsequently, UFO qualifies the alignment score of each AEU using either VQA or function calls. Finally, adaptive importance weights are applied to aggregate AEU‑level scores into a holistic score. model capabilities while ensuring the subject’s structural and semantic integrity remains intact. The distribution of atomic attribute complexity across these categories is detailed in Figure4. Non‑Editingtasks focus on recontextualizing the subject without altering its intrinsic attributes, thereby testing the model’s ability to seamlessly integrate a core subject into intricate and visually vivid environments. These scenarios prioritize natural lighting, textured backgrounds, and behaviorally consistent actions. For instance, prompts may specify a [CATEGORY] “parked under a steel bridge during steady rain at night,” with reflections on wet asphalt and atmospheric steam, or “situated above a narrow canal at dusk” amidst glowing windows and drifting boats. These tasks evaluate the robustness of identity preservation when the subject is subjected to diverse and atmospherically demanding conditions. Local Editingnecessitates the addition of specific interactive elements or props, requiring high narrative coherence and natural interaction between the subject and its surroundings. The task design emphasizes action diversity and contextual logic: for human or character subjects, this includes granular details such as “worn brown boots” or actions like “strumming a steel‑string guitar”; for other subjects, it involves adding logically consistent props like “a tin cup on the ground” or “a water bottle in a bike frame.” These prompts assess the model’s precision in performing targeted modifications without disrupting the global identity of the core subject. Global Editingfocuses on holistic style or material transformations, challenging models to apply consistent artistic or textural properties while retaining the subject’s defining features. The benchmark incorporates recognizable genres and material descriptors, such as Voxel Art (built from chunky blocks), Ukiyo‑e woodblock prints (resting on tatami floors), and Steampunk ink drawings. Each prompt provides explicit visual cues, ranging from pixelated aesthetics to traditional Japanese techniques, to evaluate how effectively models can translate a subject into diverse artistic languages without sacrificing identifiability. Complex Editingrepresents the most rigorous tier, requiring the simultaneous execution of both global stylistic transformations and local element additions. These tasks demand that models reconcile multi‑layered instructions within a single coherent frame. Examples include a “watercolor painting of [CATEGORY] carrying a folded paper map in its mouth” or a “steampunk ink drawing of [CATEGORY] operating a brass compass while wearing goggles.” By forcing models to navigate style consistency, spatial interaction, and identity preservation in parallel, these tasks provide a comprehensive evaluation of a model’s instruction‑following capabilities and visual reasoning. 6 3.2 Unified omni‑Condition Alignment (UFO) Although our framework, UFO, is generally applied to multi‑modal generation in a free‑form manner, our work in this paper focuses on a basic and well‑defined personalized image generation setting. Accurate evaluation under this fundamental setting serves as a necessary foundation for more complex multi‑image and multi‑turn generation scenarios. We propose a novel atomized chain‑of‑evaluation framework for measuring consistency in multi‑modal image generation models. Specifically, the framework comprises four stages. First, omni‑condition alignment is decomposed into AEUs. Second, UFO categorizes each AEUs into a modality‑relevance class, denoting whether visual, textual, or joint multi‑modal factors govern its alignment. Third, each AEU is quantified by VQA queries or specialized function calls. Finally, all AEU‑level scores are aggregated with adaptive weighting to obtain an interpretable holistic omni‑condition alignment score. 3.2.1 AEUs Decomposition Multi‑condition consistency refers to the visual consistency and the textual consistency of an image generation model. These include maintaining the subject’s identity, capturing key visual traits from a reference image, and meeting requirements from a text. The goal of this stage is to clearly evaluate by breaking total consistency into small, measurable, weighted semantic components, making evaluation targets clear. We design a structured semantic scheme to evaluate free‑form multi‑modal image generation. Specifically, the reference image (퐼 푟푒푓 ) and text (푇 푟푒푓 ) are put into a vision‑language model (VLM) to guide the analysis. Then, the model jointly reasons over the image and text, examining subject‑specific visual features and text editing constraints. Based on its reasoning, the omni‑condition alignment is decomposed into a sequential chain of fine‑grained, disentangled atomic evaluation units (AEUs) across two hierarchical levels, i.e., the local concrete component and global abstract attribute levels. 3.2.2 Modality‑relevance Classification Under joint multi‑modal constraints, different AEUs exhibit different modality dependencies. Therefore, it is necessary to further determine the modality relevance of each AEU after atomic decomposition. In this stage, the VLM is used to determine the modality‑relevance class of each AEU. Specifically, for each AEU, the model assigns a relevance category푟 푖 , which can be image‑only, text‑only, or text‑and‑image. As a result, each attribute is evaluated against the correct condition type, enabling fine‑grained and semantically accurate assessment of multi‑condition consistency. 3.2.3 VQA and Function Calls Scoring To assess the multi‑modal consistency of each decomposed AEU more precisely, each unit is converted into a VQA query. During evaluation, the query, reference text, reference image, and generated image are passed to the VLM. Then, the VLM decides whether the image meets the requirement and outputs Yes or No. The result is then mapped to a score: 푠 푖 = 1,if the answer is Yes, 0,if the answer is No. (1) Additionally, we introduce dedicated function calls, which integrate multiple specialized metrics to compensate for the limitations of VLMs. For example, when UFO identifies facial identity consistency as an evaluation attribute, it automatically invokes ArcFace for similarity evaluation. The resulting facial attribute score is normalized to a continuous value in[0,1], where higher scores indicate stronger identity consistency. The score is computed as: 푠 푖 = 1 − 푑 ArcFace (퐼 ref ,퐼 gen ),(2) where푑 ArcFace denotes the normalized ArcFace feature distance between the reference image퐼 ref and the generated image퐼 gen . 7 ReferenceImage BImage A Reason:Ignore “ice sculpture”. Image A failed to apply the edit. Image B applied it correctly. VIEScore 0.849 VIEScore 0.794 UFO 0.754 UFO 0.934 Prompt A carved ice sculpture of Animal placed on a black metal platform, cold blue light from the side, small cracks and trapped bubbles inside the ice. Multimodal Generation ReferenceImage BImage A Reason: Miss reference details. Image A misses the wings. Image B preserves the wings. VIEScore 0.748 VIEScore 0.520 UFO 0.246 UFO 0.836 Prompt A watercolor painting of Animal carrying a folded paper map in its mouth, wearing a red raincoat, walking past old brick buildings on a wet street. Fine-grained Decomposition ReferenceImage BImage A Reason: The ArcFace scores differ. Image A = 0.1509. Image B = 0.8146. VIEScore 0.800 VIEScore 0.794 UFO 0.865 UFO 0.919 Prompt Watercolor painting of the woman wearing a straw hat and a blue raincoat, holding a fishing rod with a metal reel, standing on wet rocks beside rough sea waves. Face Identity Figure 6Qualitative comparison with existing methods.We present three representative cases covering multimodal generation (left, blue), fine‑grained decomposition (middle, orange), and face identity evaluation (right, green). UFO produces more reliable evaluation than existing holistic metrics across diverse multi‑modal scenarios. 3.2.4 Weighted Aggregation At this stage, all AEU‑level alignment scores are aggregated. To enable the UFO to better align with human judgment, each AEU is assigned a non‑negative weight푤 푖 , where푤 푖 ∈ 1,2,3,4,5. The weight is determined through VLM‑based reasoning over the reference image and textual instruction, based on the importance of subject‑specific visual attributes and editing requirements. It reflects the contribution of each AEU to subject preservation and multi‑condition consistency. For example, if the reference text mentions a color change, color‑related units get more weight. In contrast, when a less important attribute does not meet the requirement, its inconsistency has only a limited impact on the final evaluation score. The final score is calculated for each generated image as the weighted average of its scores: Score(퐼 gen ) = ∑ 푁 푖=1 푤 푖 ⋅ 푠 푖 ∑ 푁 푖=1 푤 푖 ,(3) where푠 푖 denotes the consistency score for the푖‑th attribute or component. This aggregation strategy enables UFO to jointly consider the relative importance of different semantic conditions, resulting in a more interpretable and human‑aligned evaluation of multi‑modal consistency. 4 Experiments 4.1 Experimental Setup Judge Model.For automatic evaluation, we adopt GPT‑4o as the Judge MLLM, due to its strong multi‑modal reasoning capabilities. GPT‑4o is used to conduct structured analysis, answer targeted visual questions, and produce AEU‑level scores under predefined evaluation criteria. To ensure fair comparison and reproducibility, we use the same GPT‑4o model version (2024‑05‑13) for all evaluations. Evaluated Models.We evaluate a diverse set of both closed‑source and open‑source multi‑modal image generation models. Closed‑source models include Nano Banana [3] and Doubao (specifically utilizing the Seadream4.5 foundation) [1]. Open‑source models include Qwen‑ImageEdit‑2509, UNO [31], OmniGen2 [ 30], and Bagel [4]. All models are evaluated using their official inference pipelines and recommended hyperparameters. Baseline Metrics.We compare our method against representative automatic evaluation metrics, covering bothLLM‑basedandnon‑LLM‑basedapproaches. LLM‑based metrics include VIEScore [12] and dreambench++ [19]. Non‑LLM‑based metrics include CLIP‑I [21], CLIP‑T [21], and DINO [2]. Human Evaluation and Alignment.To assess the consistency between automatic metrics and human judgment, we conduct human evaluation on a randomly sampled subset of 100 test cases, resulting in 600 generated images from six models. For each test case, human annotators are presented 8 Table 1Quantitative results of latest customization and unified models on UFO‑Bench, including both closed‑source commercial and open‑source community variants, evaluated via the omni‑condition alignment metric UFO. Higher score denotes better performance. We highlight thebestand second‑bestvalues for each task. ModelNon‑Editing Single‑Dim‑Editing Multi‑Dim‑Editing Total Score LocalGlobal Closed‑Source Commercial Models Doubao (2025‑12)[1]0.73600.71410.78070.74750.7439 Nano‑Banana (2025‑08)[3]0.65480.71050.67390.67150.6816 Open‑Source Community Models Qwen‑Image (2025‑12)[29]0.76350.65320.78360.72570.7252 OmniGen2 (2025‑06)[30]0.51260.42140.38850.39800.4235 BAGEL (2025‑05)[ 4]0.23630.35810.44840.47190.3866 UNO (2025‑04)[31]0.40290.43810.49240.47460.4549 with the six images generated by different models and asked to rank them from 1 to 6 according to overall quality and conditional alignment. Automatic metrics follow the same protocol: for each case, the six model outputs are scored and converted into a relative ranking. We then compute the Spearman rank correlation coefficient between the human ranking and the metric‑induced ranking on a per‑case basis, and report the average Spearman correlation across all cases. Additionally, we apply Fisher Z‑transformation to estimate the average Spearman correlation∈ [−1,1]across models. A higher Spearman correlation indicates stronger agreement with human judgment in relative model comparison. 4.2 Main Results Quantitative Evaluation on UFO‑Bench.To comprehensively assess the capability of current state‑of‑the‑art models, we conducted an extensive evaluation using the UFO‑Bench dataset. The evaluation protocol involved generating outputs for every reference‑prompt pair, with each pair sampled four times to mitigate generation randomness, resulting in a total of 2,640 instances per model. As presented in Table1, the closed‑source model Doubao (2025‑12) demonstrates superior performance, achieving the highest total score of 0.7439. It exhibits remarkable stability across both local and global editing tasks, suggesting a robust understanding of fine‑grained attribute manipulation. Notably, the open‑source model Qwen‑Image (2025‑12) delivers a competitive performance (total score: 0.7252), surpassing the closed‑source Nano‑Banana (2025‑08) and significantly outperforming other community models. Interestingly, while Qwen‑Image excels in non‑editing (0.7635) and global editing (0.7836), even slightly surpassing Doubao in these categories, it shows a noticeable performance drop in local editing (0.6532). This discrepancy highlights a common limitation in current open‑source diffusion architectures: while they capture global semantic shifts effectively, precise text‑guided editing remains a challenge compared to proprietary commercial models. Correlation with Human Judgments.Following the evaluation protocol described in the Setup, we benchmark UFO against a set of existing metrics. The Spearman rank correlation coefficients are reported in Table 2. Traditional embedding metrics demonstrate a severe misalignment with human preference in the subject‑driven image generation task. Notably, CLIP‑T exhibits a negative correlation (‑0.1996). This result indicates that focusing solely on global text‑image alignment is not a reliable metric for personalized tasks. Similarly, CLIP‑I and DINO achieve only weak positive correlations (휌 ≈ 0.3 ∼ 0.4), indicating their limitations in evaluating fine‑grained conditional image editing. Recent MLLM‑based metrics (i.e., VIE‑Score, DreamBench++) show improved alignment (휌 > 0.55) by leveraging the reasoning capabilities of GPT‑4o. However, they struggle with localized precision. As evidenced in the local editing column, these methods achieve correlations around 0.58 because they lack an explicit mechanism to decouple the editing region from the preservation background. 9 Table 2Quantitative comparison of alignment with human judgments against other automatic metrics, demonstrating that our proposed UFO achieves the highest correlation with human evaluations. We highlight thebestand second‑bestvalues for each metric. MetricNon‑Editing Single‑Dim‑Editing Multi‑Dim‑EditingTotal LocalGlobal CLIP‑I[21]0.21410.30280.47950.43580.3910 CLIP‑T[21]‑0.2299‑0.1580‑0.3946‑0.2141‑0.1996 DINO[2]0.12700.12810.47910.39520.3084 VIE‑Score SC[12]0.69480.58190.45910.67880.5752 VIE‑Score PQ[12]‑0.1158‑0.03490.15380.05650.0192 VIE‑Score O[12]0.67150.58890.45380.67350.5663 DreamBench++[19]0.72620.57240.58840.57440.5977 UFO (Ours) 0.8034 (±10.63%) 0.6837 (±16.10%) 0.6295 (±6.99%) 0.7241 (±6.67%) 0.6889 (±15.26%) Table 3Ablation study on the correlation of different components in the proposed UFO with human preferences. Setting Correlation with Human Preference UFO w/o AEUs Decomposition 0.5752 UFO w/o weighted aggregation 0.6253 UFO (full setting)0.6889 Figure6further presents representative qualitative cases from three perspectives. The left case focuses on conflicting multi‑modal conditions, showing that existing metrics may favor visual similarity over correct textual instructions. The middle case shows that fine‑grained atomic decomposition better evaluates subject detail preservation, while holistic metrics often overlook subtle local inconsistencies. The right case presents identity‑sensitive face evaluation, where expert function calls enable UFO to correctly detect facial identity inconsistencies overlooked by existing metrics. 4.3 Ablation Study To investigate the individual contributions of the proposed components within the UFO framework, we conduct a series of ablation studies. The results, summarized in Table3, demonstrate that both the atomic decomposition strategy and the weighted aggregation mechanism are critical for achieving high alignment with human preferences. Effect of Weighted Aggregation.To validate the necessity of preference‑aware weighting, we replace the adaptive weighting scheme with a uniform approach,i.e., setting all reward weights to 1. As shown in the second row of Table 3, this leads to a decrease in the correlation coefficient from 0.6889 to 0.6253. This performance drop suggests that without specific preference weighting, the model treats all attributes equally. Consequently, violations in trivial attributes may overly penalize the overall score, while successful alignment with important visual characteristics or critical textual instructions is not sufficiently rewarded. Effect of Atomic Decomposition.In the setting “w/o AEUs Decomposition”, the generated image is fed directly into the VLM to produce a single holistic score. This ablation results in a significant performance degradation (0.6889 → 0.5752), indicating that holistic scoring struggles to capture complex visual details, whereas our atomic decomposition strategy enables more precise and robust evaluation. 10 5 Conclusion We introduce UFO, a unified framework for omni‑condition alignment that evaluates atomic modality‑relevance units and aggregates them into a holistic score. We further propose UFO‑Bench, a benchmark spanning diverse subjects and hierarchical editing scenarios. Extensive experiments show that UFO achieves stronger alignment with human judgment than existing metrics, particularly under complex multi‑modal conditions. Acknowledgements This research is supported by the National Natural Science Foundation of China (Grant No. 623B2094) and the Fundamental and Interdisciplinary Disciplines Breakthrough Plan of the Ministry of Education of China (Grant No. JYB2025XDXM103). 11 References [1]ByteDance. Doubao.https://w.doubao.com/, 2025. Accessed: 2025‑03‑05. [2]Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self‑supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 9650–9660, 2021. [3]Google DeepMind. Nanobanana: Gemini 2.5 flash image model.https://developers.googleblog. com/en/introducing‑gemini‑2‑5‑flash‑image/, 2025. Accessed: 2025‑10‑31. [4]Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025. [5]Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690–4699, 2019. [6]Xueqing Deng, Linjie Yang, Qihang Yu, Chenglin Yang, and Liang‑Chieh Chen. Leveraging panoptic scene graph for evaluating fine‑grained text‑to‑image generation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15107–15116, 2025. [7]Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference‑free evaluation metric for image captioning. InProceedings of the 2021 conference on empirical methods in natural language processing, pages 7514–7528, 2021. [8]Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time‑scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30, 2017. [9]Mengqi Huang, Zhendong Mao, Mingcong Liu, Qian He, and Yongdong Zhang. Realcustom: Narrowing real text word for real‑time open‑domain text‑to‑image customization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7476–7485, 2024. [10]Houcheng Jiang, Junfeng Fang, Ningyu Zhang, Guojun Ma, Mingyang Wan, Xiang Wang, Xiangnan He, and Tat‑seng Chua. Anyedit: Edit any knowledge encoded in language models.arXiv preprint arXiv:2502.05628, 2025. [11]Max Ku, Tianle Li, Kai Zhang, Yujie Lu, Xingyu Fu, Wenwen Zhuang, and Wenhu Chen. Imagenhub: Standardizing the evaluation of conditional image generation models.arXiv preprint arXiv:2310.01596, 2023. [12]Max Ku, Dongfu Jiang, Cong Wei, Xiang Yue, and Wenhu Chen. Viescore: Towards explainable metrics for conditional image synthesis evaluation. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12268–12290, 2024. [13]Black Forest Labs. Flux: Official inference repository for flux.1 models.https://github.com/ black‑forest‑labs/flux, 2024. Accessed: 2024‑11‑12. [14]Tianle Li, Max Ku, Cong Wei, and Wenhu Chen. Dreamedit: Subject‑driven image editing.arXiv preprint arXiv:2306.12624, 2023. [15]Chaojie Mao, Jingfeng Zhang, Yulin Pan, Zeyinzi Jiang, Zhen Han, Yu Liu, and Jingren Zhou. Ace++: Instruction‑based image creation and editing via context‑aware content filling.arXiv preprint arXiv:2501.02487, 2025. [16]Zhendong Mao, Mengqi Huang, Fei Ding, Mingcong Liu, Qian He, and Yongdong Zhang. Realcustom++: Representing images as real textual word for real‑time customization.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. [17]Kishore Papineni, Salim Roukos, Todd Ward, and Wei‑Jing Zhu. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 12 [18]William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023. [19]Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang, and Shu‑Tao Xia. Dreambench++: A human‑aligned benchmark for personalized image generation.arXiv preprint arXiv:2406.16855, 2024. [20]Leigang Qu, Feng Cheng, Ziyan Yang, Qi Zhao, Shanchuan Lin, Yichun Shi, Yicong Li, Wenjie Wang, Tat‑Seng Chua, and Lu Jiang. Vincie: Unlocking in‑context image editing from video.arXiv preprint arXiv:2506.10941, 2025. [21]Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. [22]Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject‑driven text‑to‑3d generation. InProceedings of the IEEE/CVF international conference on computer vision, pages 2349–2359, 2023. [23]Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans.Advances in neural information processing systems, 29, 2016. [24]Aditi Singhania, Krutik Malani, Riddhi Dhawan, Arushi Jain, Garv Tandon, Nippun Sharma, Souymodip Chakraborty, Vineet Batra, and Ankit Phogat. Beyond the pixels: Vlm‑based evaluation of identity preservation in reference‑guided synthesis.arXiv preprint arXiv:2511.08087, 2025. [25]Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. Ominicontrol: Minimal and universal control for diffusion transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 14940–14950, 2025. [26]Zhenxiong Tan, Qiaochu Xue, Xingyi Yang, Songhua Liu, and Xinchao Wang. Ominicontrol2: Efficient conditioning for diffusion transformers.arXiv preprint arXiv:2503.08280, 2025. [27]Linqing Wang, Ximing Xing, Yiji Cheng, Zhiyuan Zhao, Donghao Li, Tiankai Hang, Jiale Tao, Qixun Wang, Ruihuang Li, Comi Chen, et al. Promptenhancer: A simple approach to enhance text‑to‑image models via chain‑of‑thought prompt rewriting.arXiv preprint arXiv:2509.04545, 2025. [28]Qing Wang, Cao Te, Kuo Liao, Longfei Lu, Liqun Liu, Shuang Li, Hang Chen, Mengge Xue, Honglin Han, Jianan Li, et al. Dsh‑bench: A difficulty‑and scenario‑aware benchmark with hierarchical subject taxonomy for subject‑driven text‑to‑image generation. 2025. [29]Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng‑ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen‑image technical report.arXiv preprint arXiv:2508.02324, 2025. [30]Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, et al. Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025. [31]Shaojin Wu, Mengqi Huang, Wenxu Wu, Yufeng Cheng, Fei Ding, and Qian He. Less‑to‑more generalization: Unlocking more controllability by in‑context generation.arXiv preprint arXiv:2504.02160, 2025. [32]Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung‑Yeung Shum. Dino: Detr with improved denoising anchor boxes for end‑to‑end object detection.arXiv preprint arXiv:2203.03605, 2022. 13

Instant research discovery