Text-to-Image & Video Generation/Editing
Systematically Benchmarking and Calibrating VLM-as-a-Judge Failure Modes in Image and Video Generation
Generated automatically from the limitations stated in 15 papers (CVPR, ICML, ICLR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Current text-to-image and video generation benchmarks overwhelmingly rely on Vision-Language Models (VLMs) like GPT-4o, Gemini, and Qwen-VL as automated judges for win rates and task success metrics. However, these VLM judges systematically exhibit low correlation with human raters on fine-grained visual details, suffer from position and same-family evaluator biases, and under-penalize visual artifacts such as boundary glitches or text corruption. Consequently, benchmark progress across generative models is structurally uninterpretable, as claimed state-of-the-art gains may reflect evaluator artifacts rather than genuine improvements in visual quality and prompt fidelity.
Why it matters
Enables reliable, reproducible benchmarking in generative text-to-image and video models without expensive full-scale human studies, ensuring reported win rates reflect true perceptual quality and semantic fidelity.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Multi-Judge Robustness and Bias Audit: Aggregate standard generated image/video evaluation datasets, run parallel evaluations across 6+ commercial and open-source VLM judges (e.g., GPT-4o, Claude 3.5, Gemini 2.5/3, Qwen2.5/3-VL), and measure rank stability, position bias, same-model bias, and Kendall tau / Spearman correlation against expert human visual error annotations.
- 2
Fine-Grained Artifact Calibration Protocols: Build a diagnostic test suite isolating specific generation failure modes (e.g., text rendering, physical interaction, spatial relations, temporal boundary artifacts) and measure the sensitivity/recall of standard prompting protocols versus decomposed rubric-based multi-crop inspection prompting.
- 3
Lightweight Task-Specific Rater Alignment: Evaluate whether small, open visual reward models fine-tuned on targeted human pairwise error distinctions can reliably de-bias frontier VLM judge scores across heterogeneous generator architectures.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
Frontier VLM providers may silently update model vision backbones to resolve fine-grained artifact blindness, or individual subfields may abandon automated VLM judging in favor of domain-specific reward models before a general audit is adopted.
Evidence
Each paper's own statement of the limitation, verbatim.
- WISE: World Knowledge-Informed Semantic Evaluation for Text-to-Image GenerationICML 2026
Scoring relies on a single VLM judge (GPT-4o-2024-05-13), so results are subject to judge bias and may not generalize to other evaluators
- VisionDirector: Vision-Language Guided Closed-Loop Refinement for Generative Image SynthesisCVPR 2026
Verifier and judge are Qwen3-VL-32B/Qwen-based VLMs themselves, so LGBench scores may reflect verifier bias toward models in the same family; responses below 0.81 confidence are filtered rather than human-audited
- VIVA: VLM-Guided Instruction-Based Video Editing with Reward OptimizationCVPR 2026
Evaluation relies primarily on a single VLM judge (Gemini-2.5-Pro), which is also used in the data pipeline, introducing potential evaluator-training circularity
- MultiBanana: A Challenging Benchmark for Multi-Reference Text-to-Image GenerationCVPR 2026
Evaluation relies on VLM judges with only moderate human correlation (Gemini 2.5 Pearson r=0.5726, Cohen's kappa=0.3908) validated on just 32 samples, so absolute scores may be unreliable
- MotiMotion: Motion-Controlled Video Generation with Visual ReasoningICML 2026
Evaluation relies on Gemini 3.1 Pro as both the reasoning agent and the automated judge, creating a potential same-model bias in VLM-based metrics
- MiVE: Multiscale Vision-language features for reference-guided video EditingICML 2026
Automated evaluation uses Gemini-3-Flash, a separate commercial VLM; potential evaluator bias is mitigated (InternVL3.5-8B replication) but the benchmark remains VLM-judged
- In-Context Generation with Regional Constraints for Instructional Video EditingICML 2026
Evaluation relies solely on a VLLM (Gemini-2.5-Flash) as judge with no human preference or automatic metrics, so scores may reflect judge bias
- ImagenWorld: Stress-Testing Image Generation Models with Explainable Human Evaluation on Open-ended Real-World TasksICLR 2026
VLM-based automated metrics systematically under-penalize artifacts (positive bias +0.06, Spearman 0.59 vs human-human 0.64), so scalable automatic evaluation fails precisely on the flaw categories (unreadable text, boundary glitches) that dominate text-heavy domains
Show all 15 papers
- FLUX-Reason-6M & PRISM-Bench: A Million-Scale Text-to-Image Reasoning Dataset and Comprehensive BenchmarkICLR 2026
Benchmark evaluation depends on proprietary VLM judges (GPT-4.1) whose scores may carry judge-specific biases, despite the reported Qwen2.5-VL agreement
- DreamOmni2: Multimodal Instruction-based Generation and EditingCVPR 2026
Evaluation relies heavily on VLM judges (Gemini 2.5, Doubao 1.6) with small test sets (apparently ~40-90 cases per cell given success rates in increments like 0.0417), so scores have high variance and VLM judges may miss subtle inconsistencies, as the authors note GPT-4o/Nano Banana issues are 'difficult for VLMs to detect accurately'
- Constantly Improving Image Models Need Constantly Improving BenchmarksICLR 2026
VLM-as-a-judge correlations with human rankings are weak (GPT tau_b=0.117, Gemini tau_b=0.083, Qwen tau_b=0.045 not significant), so the headline win-rate results rest on unreliable automatic evaluation
- BideDPO: Conditional Image Generation with Simultaneous Text and Condition AlignmentICLR 2026
Evaluation relies heavily on VLM judges (Qwen2.5-VL-72B, GPT-4o) for Success Ratio and the semantic-guided metrics; although human raters corroborate rankings, the primary metrics inherit VLM judge biases
- Benchmarking and Improving Fine-Grained Text-to-Image Alignment via Paired Reinforcement LearningICML 2026
Evaluation depends on VLM judges (Qwen3-VL-30B Pearson-r 0.77 with human eval; even Gemini3-Pro only 0.83), so benchmark scores may not fully reflect human judgment of fine-grained details
- Personalized Preference Fine-tuning of Diffusion ModelsCVPR 2025
Evaluation relies on GPT-4o as a judge (validated at 83% agreement), so win rates partly reflect VLM judge biases; only Stable Cascade Stage C is fine-tuned, leaving generalization to other architectures untested
- BlenderGym: Benchmarking Foundational Model Systems for Graphics EditingCVPR 2025
VLM verifiers misalign with human verifiers: best model (Claude 3.5 Sonnet) reaches 0.66 alignment vs 0.79 inter-human, and some verifiers (Qwen) exhibit position bias favoring the second candidate regardless of permutation
Nearest existing work
- Fooling the LVLM Judges: Visual Biases in LVLM-Based EvaluationEMNLP 2025
- Video-Bench: Human-Aligned Video Generation BenchmarkCVPR 2025
- ImagenWorld: Stress-Testing Image Generation Models with Explainable Human Evaluation on Open-ended Real-World TasksICLR 2026
- Toward Verifiable and Reproducible Human Evaluation for Text-to-Image GenerationCVPR 2023
- VF-Eval: Evaluating Multimodal LLMs for Generating Feedback on AIGC VideosACL 2025
- Gen3DEval: Using vLLMs for Automatic Evaluation of Generated 3D ObjectsCVPR 2025
- LLMScore: Unveiling the Power of Large Language Models in Text-to-Image Synthesis EvaluationNeurIPS 2023
- VideoRealBench: A Chain-of-Thought Realism Evaluation Benchmark for Generated Human-Centric VideosCVPR 2026
- Enhancing Reward Models for High-quality Image Generation: Beyond Text-Image AlignmentICCV 2025
- Automatic Evaluation for Text-to-image Generation: Task-decomposed Framework, Distilled Training, and Meta-evaluation BenchmarkACL 2025
- VideoScore: Building Automatic Metrics to Simulate Fine-grained Human Feedback for Video GenerationEMNLP 2024
- Evaluating Text-to-Visual Generation with Image-to-Text GenerationECCV 2024
- MM-JudgeBias: A Benchmark for Evaluating Compositional Biases in MLLM-as-a-JudgeACL 2026
- ADIEE: Automatic Dataset Creation and Scorer for Instruction-Guided Image Editing EvaluationICCV 2025
- Ref4D-VideoBench: Four-Dimensional Reference-Based Evaluation of Text-to-Video Generative ModelsCVPR 2026
Related open problems
Text-to-Image & Video Generation/Editing
Benchmarking and Mitigating Upstream LLM Prompting Dependencies in Text-to-Visual Generation
Modern text-to-image and text-to-video generation and editing methods rely heavily on external large language models (LLMs) for prompt rewriting, sub-prompt decomposition, and attribute filtering. Consequently, visual generation quality is bottlenecked by upstream LLM biases, template artifacts, and hallucinated attributes, frequently failing on ambiguous or in-the-wild user instructions unless manually tuned. Because no systematic study has evaluated visual generation frameworks across diverse upstream prompt generators and decomposition strategies, the sensitivity and failure modes of these pipelines remain unquantified.
Text-to-Image & Video Generation/Editing
Dense Motion Supervision Without Optical-Flow Ground Truth
Nearly every pipeline that needs dense correspondence—video generation control, camera pose estimation, motion attribution—currently anchors its supervision or its inference to optical flow, either ground-truth annotations or outputs of a pretrained model (RAFT, AllTracker). This couples training data to static-camera, high-quality, curated footage, caps dataset size (22k–630k pairs), injects flow errors directly into downstream quality (cycle-consistency errors >1000 px), and bakes in simplifying assumptions like first-frame conditioning or pinhole intrinsics. Every consumer of flow inherits the same ceiling set by flow estimators and the data they were trained on.