acceptodds
Withdrawn from ICLR 2027

Exploring the AI Obedience: Why is Generating a Pure Color Image Harder than CyberPunk?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent advances in generative AI have shown human-level performance in complex content creation. However, we identify a "Paradox of Simplicity": models that can render complex scenes often fail at trivial, low-entropy tasks, such as generating a uniform pure color image. We argue this is a systemic failure related to uncontrollable emergent abilities. As models scale, strong priors for aesthetics and complexity override deterministic simplicity, creating an "aesthetic bias" that hinders the model's transition from data simulation to true intellectual abstraction. To better investigate this problem, we formalize the concept of AI Obedience, a hierarchical framework that grades a model's ability to transition from probabilistic approximation to pixel-level determinism (Levels 1 to 5). We introduce Violin, the first systematic benchmark designed to evaluate Level 4 Obedience through three deterministic tasks: color purity, image masking, and geometric shape generation. Using Violin, we evaluate several state-of-the-art models and reveal that closed-source models generally outperform open-source ones in deterministic precision. Interestingly, performance on our benchmark correlates with the benchmark in natural image generation. Our work provides a foundational framework and tools for achieving better alignment between human instructions and model outputs.

Primary area:
datasets and benchmarks
Keywords:
AI Obedience, Image Generation, Visual Generative Model

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_wTHcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies failures of image-generation models on visually simple tasks requiring precise execution. It proposes an L1–L5 “AI Obedience” taxonomy and introduces Violin, a benchmark of 7,745 examples covering pure-color generation, geometric shape generation, and image masking. The reported comparisons generally favor proprietary models. Additional experiments examine explicit color codes, fine-tuning, and generalization. The authors attribute the observed failures to semantic and aesthetic priors overriding exact constraints.

Critical strengths
  1. The paper addresses a useful evaluation gap: visually convincing images can still violate numerical or pixel-level requirements. This motivates complementary testing beyond semantic alignment.
  2. The tasks support programmatically constructed reference outputs and cover complementary aspects of precision (Section 3.1; Figure 4). The explicit-code extension and generalization tests in Appendix E strengthen the benchmark’s diagnostic potential.
Critical weaknesses
  1. The evaluation does not reliably measure the exact instruction following claimed by the paper. The paper defines L4 obedience as exact numerical execution (Section 2.2, p. 4), but the metrics can overlook violations of the instructions. The masking metrics do not directly measure preservation of the source content (Section 3.2; Appendix D.2, Eqs. 7–10), while geometric purity measures color variation rather than correctness of the required fill color (Appendix D.2, Eq. 14). In synthetic checks of the released evaluators, changing every retained pixel from grayscale 100 to 200 still produces zero error on all five masking metrics. Adding an extra disconnected white square similarly produces zero error on all five shape metrics. These counterexamples undermine the interpretation of the reported scores as exact obedience and warrant corrected evaluation. The main color task also uses natural-language color names but scores against fixed RGB representatives (Section 3.1; Table 5, p. 18). Appendix E, p. 21, explicitly acknowledges this ambiguity, yet the explicit-code experiments in Table 7 exclude the proprietary models used in the main comparison. Consequently, the principal color results do not cleanly separate numerical execution from differences in color-name interpretation.

  2. The proposed explanation and conclusions about fine-tuning exceed the supporting evidence. The examples in Section 2.3 and Figure 3 illustrate failures, but do not establish that aesthetic priors or emergent abilities cause them. Alternative explanations, including training-data coverage, numerical encoding, and image reconstruction, remain unresolved. The architectural and scaling claims in Appendix F, p. 25, therefore require controlled evidence or more cautious wording. The fine-tuning results also deserve a more balanced interpretation (Appendix E.3). Janus-Pro-1.5’s color error decreases from 0.191 to 0.045 (Tables 7 and 9), approximately a 76% relative reduction. Although this remains short of exact execution, it represents substantial improvement. Moreover, Qwen-Image’s post-tuning Var-1 score differs between Table 8 (0.048) and Table 9 (0.079) without a clear explanation. Figure 7 demonstrates degradation under the tested setup, rather than a general limitation of learning these tasks through additional data.

Questions for authors
  1. After correcting the metric blind spots and using explicit numerical color prompts, do the main model rankings remain stable? (Section 3.2; Appendices D.2 and E)
  2. What controlled evidence distinguishes the aesthetic/emergence explanation from data or reconstruction limitations? (Section 2.3; Appendix F)
  3. What explains the discrepancy between Tables 8 and 9, and which model and configuration produced Figure 7?
Reviewer_9bcsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates why advanced image generation models struggle with simple, deterministic tasks. The authors propose a hierarchical AI Obedience framework and introduce Violin, a benchmark evaluating pixel-level precision through pure-color generation, geometric shapes, and image masking. Experiments reveal limitations in precise instruction following and suggest that learned generative priors may contribute to these failures.

Critical strengths
  1. Important and interesting research question. The paper highlights an intriguing limitation of modern image generation models: their difficulty in following simple but precise instructions despite their strong capabilities in generating complex visual content.

  2. Meaningful conceptual framework. The proposed AI Obedience hierarchy provides a reasonable perspective for organizing different levels of instruction-following capabilities, particularly the distinction between semantic alignment and deterministic precision.

Critical weaknesses
  1. Insufficient discussion of prior work and unclear contributions. The authors claim that Violin is the first systematic benchmark for Level-4 Obedience, covering precise color generation, geometric shapes, and image masking. However, GenColorBench [1] has already systematically evaluated precise color generation using similar color representations and metrics, including ISCC-NBS colors and CIEDE2000. The paper does not adequately discuss this overlap or explain what additional insights Violin provides. Related efforts such as PaintBench [2], although potentially contemporaneous, further highlight the importance of establishing its distinct value. I encourage the authors to provide a task-by-task comparison with existing benchmarks and demonstrate which failure modes or evaluation capabilities are uniquely captured by Violin. This is important for establishing the benchmark's incremental contribution.

  2. Insufficient validation of the proposed failure mechanisms. The authors attribute deterministic generation failures to semantic gravity, aesthetic inertia, and learned generative priors. However, the main evidence comes from the illustrative cases in Figure 3, which do not systematically isolate these mechanisms. For example, the tendency to produce a 50/50 split despite asymmetric instructions could reflect difficulties in interpreting numerical proportions rather than aesthetic preferences. Similarly, the claimed "consistent" semantic prior of rusted iron is not clearly consistent with the target purple color (#9966CC). Alternative explanations, including numerical representation limitations investigated in NumColor [3] and image decoding artifacts, remain insufficiently examined. Controlled experiments across different prompts, numerical constraints, and random seeds, together with appropriate representation or decoding controls, would help establish whether the proposed mechanisms actually explain these failures.

  3. Insufficient justification and validation of evaluation metrics. The authors propose composite scores by equally averaging five normalized metrics for each task, claiming to provide a rigorous evaluation of deterministic obedience. However, the choice of equal weighting is not justified, and some metrics measure overlapping properties (e.g., RGB distance and CIEDE2000, or edge density and high-frequency energy). It is therefore unclear whether the composite scores reliably reflect obedience or are sensitive to arbitrary metric choices. I suggest adding metric ablations, examining ranking stability under different weighting schemes, and validating the scores against controlled pixel-level errors. These analyses would strengthen the reliability of the benchmark and its model comparisons.

  4. Potential inconsistency in the image masking metric. The image masking task explicitly defines mask value 1 as preserving the original pixels and value 0 as replacing them with black. However, Equation (8) defines content leakage as the average intensity within the ground-truth mask region. If the same mask convention is used, this would penalize correctly preserved non-black pixels rather than measuring leakage in regions that should be black. The paper does not clearly explain whether the mask is inverted before evaluation. I encourage the authors to clarify the implementation and provide sanity checks demonstrating that perfect ground-truth outputs achieve zero error. This is particularly important because the metric directly contributes to the reported results in Table 2.

References

[1] Butt, M. A., et al. GenColorBench: A Color Evaluation Benchmark for Text-to-Image Generation Models. CVPR, 2026. https://arxiv.org/abs/2510.20586

[2] Xu, K., et al. PaintBench: Deterministic Evaluation of Precise Visual Editing. arXiv:2606.00188, 2026. https://arxiv.org/abs/2606.00188

[3] Butt, M. A., et al. NumColor: Precise Numeric Color Control in Text-to-Image Generation. ECCV, 2026. https://arxiv.org/abs/2603.13547

Questions for authors
  1. Color ground truth: How do the authors justify assigning exact RGB values to ambiguous natural-language color names in a benchmark intended to measure deterministic obedience? Would using the absolute-color benchmark as the primary evaluation be more appropriate?

  2. Geometric evaluation: Why is the generated mask resized before computing geometric IoU? Could this obscure spatial or scale errors that the benchmark aims to measure?

  3. Statistical robustness: Could the authors report variability across random seeds, confidence intervals, or statistical significance, particularly for models with similar average scores?

  4. Obedience hierarchy: How are L4 and L5 conceptually distinguished, given that the L4 geometric tasks already involve precise coordinates and spatial constraints?

  5. Fine-tuning trade-off: Could the authors investigate alternative fine-tuning configurations or mixed-data training to determine whether the observed degradation in general generation capabilities represents an inherent trade-off?

Reviewer_MpTROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates the "Paradox of Simplicity" in generative AI, and introduces AI Obedience, a hierarchical framework with six levels (L0–L5) that characterize different degrees of instruction-following precision. To operationalize Level-4 “Instructional Obedience,” the authors introduce Violin, a benchmark with three low-entropy tasks (pure color, image masking, geometric shapes) and associated metrics to quantify color precision/purity and spatial exactness. As a result, they report that closed-source models generally outperform open-source ones and provide a coarse empirical correlation between Violin scores and standard natural-image benchmarks.

Critical strengths

The design of Violin targets low-entropy regimes (bit-level purity in colors, exact masks, and geometric constraints), which is a meaningful pivot from prevailing “plausibility” evaluations. And the benchmark is programmatically synthesized with deterministic ground truth, enabling unambiguous evaluation and reproducibility for the dataset itself.

Critical weaknesses
  1. Validity of Violin as a measure of AI Obedience is not clear. The proposed benchmark may conflate instruction-following ability with low-level image generation precision. For example, the main color generation task maps natural language color names to fixed RGB values, even though terms such as "red" or "white" do not uniquely specify exact colors. Moreover, a model may correctly understand an instruction but fail to produce pixel-perfect outputs due to limitations in its VAE decoder, sampling process, or image post-processing. Without disentangling these factors, it is unclear whether Violin genuinely measures AI Obedience or primarily evaluates the numerical precision of image generation architectures.
  2. Insufficient experimental evidence for the claimed underlying mechanisms. The paper attributes obedience failures to aesthetic bias, semantic gravity, and learned generative priors, but these claims are mainly supported by limited case studies rather than systematic experiments. Important factors, such as the effects of VAE reconstruction, sampling strategies, and prompt variations, are not adequately isolated. Furthermore, comparisons with simple deterministic rendering or existing constrained generation methods are missing. These controls would help determine whether the observed failures stem from the proposed mechanisms or from limitations of current generation pipelines. Consequently, the experiments demonstrate the existence of precision errors but do not convincingly establish their underlying causes.
Questions for authors
  1. How are the sub-metrics normalized before averaging? Are the normalization ranges fixed across all models?
  2. Could some pixel-level errors come from the image decoder rather than instruction-following failures? Could you test VAE reconstruction errors on ground-truth images to isolate this effect?
  3. How were different models evaluated on the masking task? Did they receive the same inputs and use comparable interfaces or APIs?
  4. Were multiple rounds/replications tested for each prompt? Does the models give the same/similar results every time? And how different are they between each rounds?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.