Text-to-image generators are increasingly tuned and filtered by automated judges: a vision-language model looks at a generated picture and its prompt, then scores how good the picture is. This note proposes a test of a specific suspicion. When an image is beautiful, these judges may "see" requested content that is not there. They report that the red umbrella is present, or that the scene is low-light as asked, because the image looks polished. We call this the aesthetic halo. If it exists, it ties together two known failures that have so far been studied separately. Generators tuned on human preference learn to make pretty images that drop hard constraints, and multimodal judges that reason step by step invent visual evidence to support their verdicts. A judge with an aesthetic halo would reward exactly the shortcut the generator is learning.
We propose to measure the halo with synthetic image pairs whose ground truth is known by construction. Each pair varies how attractive the image is independently of whether it satisfies the prompt. We then train a judge whose reasoning must point to image regions for every prompt element before it scores. Training uses reinforcement learning on those same pairs, so the judge's verdict on whether the image matches the prompt cannot move when only aesthetics change. Finally, we test whether this judge, used as a reward, lets a generator improve compositional accuracy without giving up visual quality.
Core Claim
The claim has three parts. Off-the-shelf multimodal judges, whether they score directly or reason freely first, are more likely to hallucinate the presence of a missing prompt element when the image is high in aesthetic quality. Requiring the judge to reason from evidence, naming a bounding box or explicitly declaring absence for each prompt element, reduces this halo. It does so only if the judge is trained against counterfactual pairs, not merely prompted. A judge trained this way is a better reward for preference tuning of image generators, because it blocks the "pretty but incomplete" shortcut.
We can state the halo as a single quantity. Let $J(e \mid x)$ be the judge's probability that prompt element $e$ is satisfied in image $x$. Take an image pair $(x^{hi}, x^{lo})$ in which $e$ is absent from both and the two differ only in aesthetic quality. The halo is
$$H = \mathbb{E}\left[ J(e \mid x^{hi}) - J(e \mid x^{lo}) \right],$$
averaged over elements and pairs. An unbiased judge has $H \approx 0$. A positive $H$ means beauty is buying semantic credit.
Why This Gap Is Real
Several papers motivate the hypothesis. Sun et al. show that preference-tuned generators learn shortcuts: they produce clean, attractive images that ignore spatial arrangements and attribute bindings. Xiangru Zhu argues that gains in general visual quality can hide stagnating sensitivity to the wording of prompts. On the judging side, MLLM-as-a-Judge benchmarks multimodal models as evaluators and finds that step-by-step reasoning often lowers judging accuracy, because the judge hallucinates visual features in its rationale and then scores from them. It also finds that judges cluster scores high. GHOST generates subtle counterfactual images that trigger object hallucinations, and shows that even models trained to "think" produce reasoning chains justifying objects that do not exist. None of these works asks whether aesthetic quality is the variable that tilts a judge toward false presence. That is the measurement we add.
The closest evaluator is X-IQE. It has a vision-language model judge text-to-image outputs through ordered reasoning stages: artifacts first, then prompt alignment, then aesthetics. Its rationales are free text, not tied to image regions. It downsamples images to 224×224. It never checks whether the judge itself hallucinates. T2I-CompBench and T2I-CompBench++ decompose alignment into atomic checks using a question-answering model, a detector, and depth estimation. They also report that multimodal judges score inconsistently. Their checks, however, run over fixed vocabularies of colors, shapes, and objects, and cannot express style or negative constraints such as "a simple, low-light sketch." Chen et al. and Cao et al. show that LLM prompt rewriting injects decorative modifiers that override those same constraints. The judge we propose is meant for this harder territory: open-vocabulary, relational, and restrictive constraints.
Two papers supply the method. DeFacto masks the visual evidence a question depends on and trains a model, with group-relative reinforcement learning, both to abstain when the evidence is gone and to place its "thought" boxes on the true evidence. It shows that supervised fine-tuning alone leaves evidence-answer consistency near zero. Q-Insight shows that useful image-quality reasoning can emerge from reinforcement learning on scalar rewards, with no human rationales. We carry DeFacto's evidence-anchoring and counterfactual-abstention idea from question answering over real photos to judging generated images, and add the aesthetic axis as the controlled nuisance variable.
One paper constrains the design. Zhang and Wan show that alignment metrics such as DSGScore can swing by up to 50 points on a 100-point scale after a single-pixel change. Any judge we build therefore has to pass a perturbation-stability test. Otherwise it may simply be brittle in a new way.
Research Program
Building the counterfactual pairs. We start from prompts in T2I-CompBench's 8,000-prompt set, DrawBench, and COCO captions. We add a slice of prompts with restrictive and negative constraints, written in the style of the semantic-drift failures from Chen et al.. For each prompt we generate a base image, then make two kinds of edit.
The semantic edit removes or alters exactly one prompt element. An open-vocabulary detector produces a mask, the element is removed or recolored by inpainting, and the detector is run again to confirm the change. This follows DeFacto's evidence masking and avoids the unCLIP dependency that limits GHOST. Style and negative constraints cannot be masked by a detector, so for those we regenerate from a minimally edited prompt and verify by human spot checks.
The aesthetic edit moves the image's attractiveness while holding semantics fixed. We use three levers. The first is classifier-free guidance, where the sampler pushes harder toward the prompt. SynthAlign found that scale 11.0 over-saturates images relative to 7.0. The second is DeRaDiff's single-parameter blend between a reference model and a preference-tuned model at sampling time. The third is mild photometric degradation. Guidance changes can also change content, so every aesthetic variant is re-verified for semantic equivalence. A variant is kept only if the detector and a human spot check agree that it matches the original semantically, and if it moves HPS v2 in the intended direction.
The result is a two-by-two design for each prompt: semantics satisfied or violated, crossed with aesthetics high or low. Ground truth comes from the construction, not from a judge. We also annotate a held-out human-verified subset as the final arbiter.
Measuring the halo. We estimate $H$, along with the violated-pair ranking accuracy, for off-the-shelf judges under three prompting modes: direct scoring, free reasoning before the score, and X-IQE's staged reasoning. The judges are open models such as Qwen2.5-VL and LLaVA-OneVision, plus LLaVA-Critic. We break $H$ down by constraint type: objects, attributes, spatial relations, counts, and style or negative constraints. We also run each judge at two input resolutions, to separate the halo from the resolution bottleneck X-IQE suffers.
Training an evidence-first judge. The judge's output format requires, for each prompt element parsed from the prompt, either a box with a short description or an explicit "absent." It then gives a semantic score and a separate aesthetic score. Training uses group-relative reinforcement learning with three rewards. The first is ranking accuracy on counterfactual pairs. The second is overlap between the predicted boxes and the masks from the synthesis step, as in DeFacto. The third is an invariance term that penalizes any change in the semantic score between aesthetic variants of the same image. The aesthetic score is supervised against HPS v2, so the judge still tracks quality and is not simply blind to it.
Using the judge as a reward. We run Diffusion-DPO-style preference tuning of Stable Diffusion, with winners and losers chosen by different reward sources. We then measure whether compositional accuracy rises while HPS v2 and PickScore hold. As a second, smaller downstream check, we swap the judge into SynthAlign's image-selection step in place of ImageReward. This tests whether filtering synthetic training images for dropped elements helps the vision-language models trained on them.
Baselines
The naive reference point is HPS v2 alone, used both as the judge and as the reward. It is what a practitioner gets by optimizing human preference, and the halo story predicts it favors high-aesthetic violated images almost by definition. On the judging benchmark it shows how far pure preference scores are from semantic correctness. In downstream tuning, it is the condition expected to produce aesthetic overfitting.
The strongest existing judge-style competitor is LLaVA-Critic. It is a 7B and 72B evaluator trained on 113k examples whose reasoning comes before the score. It reports a Pearson correlation of 0.754 with GPT-4o on the MLLM-as-a-Judge benchmark, and it raised LLaVA-v1.5-7B on LLaVA-W from 63.4 to 73.5 when used as a DPO reward. We run the 7B critic with prompts adapted to text-to-image judging, at the same input resolution as our judge. We also give it a fine-tuning control: the same counterfactual data, used through supervised training on rationales distilled from a stronger model. That separates "more data" from "evidence-anchored reinforcement learning." For the T2I-specific reasoning judge, the reference is X-IQE, which reports human correlation of 0.381 for text alignment on COCO versus 0.137 for CLIPScore, and 0.418 for aesthetics. We re-implement its staged prompting on the same base model as our judge, so the comparison isolates output structure and training.
The strongest non-reasoning competitor is a factorized pipeline built from T2I-CompBench's own tools. It uses disentangled question answering and detector-based checks for semantics, HPS v2 for aesthetics, and GORS-style selection for tuning. GORS raised Stable Diffusion v2's attribute-binding score from 0.5065 to 0.6603 on T2I-CompBench. We reuse that setup with the same prompts, the same number of candidate images per prompt, and the same LoRA tuning budget. For T2I-CompBench++ we found no per-model numbers on Lacuna, so we run all models ourselves. For the halo benchmark itself there is no prior number at all, since the quantity is new.
Two ablations isolate the key ingredient. The first removes the box-or-absent requirement but keeps reinforcement learning on the counterfactual pairs. The second keeps the evidence format but drops the invariance reward. Our judge counts as beating the field if it achieves the lowest $H$ and the highest violated-pair ranking accuracy on the human-verified subset, with the gap largest on style and negative constraints. In downstream tuning, it counts as a win if it matches or exceeds the factorized pipeline's T2I-CompBench gains while holding HPS v2 within noise of the HPS-only model.
Why not just…?
Why not just score directly and skip the reasoning, given the paradox? MLLM-as-a-Judge does show that free reasoning can hurt. Direct scoring, however, has its own problems: Xie et al. document unstable ratings when a model must compress its judgment into one token, and Jia et al. describe shortcut-driven scoring. Our halo measurement runs direct scoring as its own condition. If direct scoring shows a halo no larger than free reasoning does, the paradox is about unanchored reasoning rather than reasoning as such. We expect it to show a halo, because direct scorers are exactly where global impressions dominate. Either way the result is informative, and the evidence-first judge has to beat both conditions.
Why not just use detectors and question-answering checks, which cannot be charmed by beauty? This is the most serious threat. On objects, colors, and counts, T2I-CompBench's pipeline may well have $H$ near zero and match our judge. We report the comparison broken down by constraint type for exactly this reason. Our claim stands only if the evidence-first judge wins on spatial relations, complex compositions, and style or negative constraints, where fixed-vocabulary detectors have no handle. OpenCHAIR already shows that closed checklists miss long-tail failures, which suggests the gap exists. If the detector pipeline matches us everywhere, the right conclusion is that the halo is a judge problem with a detector solution, and the benchmark becomes the main contribution.
Why not just fine-tune the judge on rationales written by a stronger model? DeFacto reports that its base and supervised variants sit near zero mAP on evidence-answer consistency, while reinforcement learning with counterfactual rewards reaches 35.5. Q-Insight reports 92.77% degradation perception under reinforcement learning versus 59.98% under supervised training. Distilled rationales also carry the teacher's own halo. Our supervised control on LLaVA-Critic tests this directly on the new benchmark.
Why not just give the judge a caption first, since judges do better from text? MLLM-as-a-Judge does find that text descriptions help. A caption written by the same model is still produced under the image's aesthetic impression, so it may launder the halo rather than remove it. We include a caption-then-judge condition and report its $H$.
Proposed Experiments
The first decisive check is existence. We look for a positive $H$, with confidence intervals obtained by resampling over prompts, for at least two off-the-shelf judges on the human-verified subset, and test whether it is larger for style and negative constraints than for objects. A null here ends the line of work, and the benchmark would be released as a negative result.
The second check is mitigation. On held-out prompts and a held-out generator, the evidence-first judge must reduce $H$ relative to every baseline and ablation, keep human correlation on X-IQE's COCO alignment protocol at or above 0.381, and keep its scores stable under the single-pixel and random-seed perturbations of Zhang and Wan.
The third check is the downstream payoff. Preference tuning with our judge should raise T2I-CompBench scores, measured by its own disentangled question-answering, detector, and judge-model metrics, while HPS v2 and PickScore stay level with the HPS-only model. A human study should confirm higher constraint satisfaction on the negative-constraint slice.
As a secondary check, swapping our judge for ImageReward in SynthAlign's selection step should be compared against SynthAlign's reported results for LLaVA-1.5-7B: an MMHal-Bench hallucination rate of 25.0%, down from 51.0%, and 87.6% POPE accuracy. This tells us whether removing dropped-element images from synthetic training data matters for the vision-language models trained on it.
References
- Diffusion-RainbowPA: Improvements Integrated Preference Alignment for Diffusion-based Text-to-Image Generation
- Evaluating Semantic Variation in Text-to-Image Synthesis: A Causal Perspective
- MLLM-as-a-Judge: Assessing Multimodal LLM-as-a-Judge with Vision-Language Benchmark
- GHOST: Hallucination-Inducing Image Generation for Multimodal LLMs
- X-IQE: eXplainable Image Quality Evaluation for Text-to-Image Generation with Visual Large Language Models
- T2I-CompBench: A Comprehensive Benchmark for Open-world Compositional Text-to-image Generation
- T2I-CompBench++: An Enhanced and Comprehensive Benchmark for Compositional Text-to-Image Generation
- Improving Text-to-Image Generation with Input-Side Inference-Time Scaling
- BeautifulPrompt: Towards Automatic Prompt Engineering for Text-to-Image Synthesis
- DeFacto: Counterfactual Thinking with Images for Enforcing Evidence-Grounded and Faithful Reasoning
- Q-Insight: Understanding Image Quality via Visual Reinforcement Learning
- Re-Thinking the Automatic Evaluation of Image-Text Alignment in Text-to-Image Models
- Multimodal Preference Data Synthetic Alignment with Reward Model
- DeRaDiff: Denoising Time Realignment of Diffusion Models
- LLaVA-Critic: Learning to Evaluate Multimodal Models
- An Empirical Analysis of Uncertainty in Large Language Model Evaluations
- Leveraging LLMs for Dialogue Quality Measurement
- Mitigating Open-Vocabulary Caption Hallucinations