Text-to-Image & Video Generation/Editing
Benchmarking and Mitigating Upstream LLM Prompting Dependencies in Text-to-Visual Generation
Generated automatically from the limitations stated in 12 papers (ICCV, NeurIPS, ECCV), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
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.
Why it matters
Makes text-to-visual generation and editing pipelines robust to arbitrary, ambiguous user prompts without requiring manual prompt engineering or proprietary LLM filtering.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Multi-LLM Robustness Benchmark: Evaluate existing T2I/T2V prompt-dependent generation and editing pipelines across a standardized suite of upstream LLMs (varying in scale, training priors, and open/closed weights) on in-the-wild prompts, measuring visual fidelity, text-image alignment, and artifact frequencies.
- 2
Visual-Feedback Prompt Decomposition: Develop a closed-loop prompt refinement mechanism where downstream visual cross-attention maps and alignment scores provide direct feedback to the upstream prompt decomposition step, reducing reliance on manual template engineering.
- 3
Vision-Grounded Prompt Calibration: Train a lightweight, vision-aligned prompt adapter directly tuned on downstream generative failures rather than generic text-only LLM completions, measuring degradation rates on out-of-distribution user queries.
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
Rapid architectural transitions toward end-to-end native multimodal foundation models that directly process raw, uncurated user text without auxiliary LLM prompt rewriting stages could dissolve the problem.
Evidence
Each paper's own statement of the limitation, verbatim.
- Multimodal Semantic Bias Mitigation for Diverse Text-To-3D GenerationCVPR 2026
Requires ~25 hours of fine-tuning plus enriched sample generation, and relies on GPT-4 prompt generation with manual filtering of duplicates and inappropriate prompts
- Exposing Hidden Biases in Text-to-Image Models via Automated Prompt SearchICML 2026
Method depends on external classifiers trained on a manually curated dataset and on the LLM prior, both of which can inject their own biases into the discovered prompts
- Stable Cinemetrics : Structured Taxonomy and Evaluation for Professional Video GenerationNeurIPS 2025
Prompt generation depends on LLMs whose proprietary nature and biases can influence prompt language and structure
- SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image EditingNeurIPS 2025
Dependence on LLM prompt decomposition: output quality varies with the choice and reliability of the LLM used to generate sub-prompts
- PixTalk: Controlling Photorealistic Image Processing and Editing with LanguageICCV 2025
Sensitivity to prompt quality: ambiguous or naive prompts (e.g. 'enhance it') and out-of-distribution instructions degrade performance
- MatchDiffusion: Training-free Generation of Match-CutsICCV 2025
Effective output depends on carefully crafted prompts; the authors note automated prompt engineering would be needed for broader accessibility and there is no fine-grained user control over which structural aspects align
- DIMCIM: A Quantitative Evaluation Framework for Default-mode Diversity and Generalization in Text-to-Image Generative ModelsICCV 2025
Relies on external models (Llama3.1 for prompt/attribute construction, InstructBLIP-Flan-T5-XXL VQAScore for scoring), so results are subject to LLM hallucination and VQA bias, only partially mitigated by attribute filtering and truncated prompts
- Unveiling and Mitigating Memorization in Text-to-image Diffusion Models through Cross AttentionECCV 2024
Memorized prompts come only from Webster's extracted set and non-memorized prompts from ChatGPT-4 generations, so the entropy separation may not hold for arbitrary in-the-wild prompts
Show all 12 papers
- Factorized Diffusion: Perceptual Illusions by Noise DecompositionECCV 2024
One prompt's component may dominate the generated image; success requires carefully choosing prompt pairs and manually tuning decomposition parameters
- Advancing Fine-Grained Classification by Structure and Subject Preserving AugmentationNeurIPS 2024
Relies on LLM-generated prompts; fails to produce optimal prompts if the LLM lacks knowledge of the meta-class
- Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image SynthesisICCV 2023
Potential for artifacts if local descriptions and global prompts are not semantically consistent or well-integrated
- HRS-Bench: Holistic, Reliable and Scalable Benchmark for Text-to-Image ModelsICCV 2023
Prompt generation depends on GPT-3.5 and manual filtering, introducing potential template and model-specific biases in the prompt distribution
Nearest existing work
- The Devil is in the Prompts: Retrieval-Augmented Prompt Optimization for Text-to-Video GenerationCVPR 2025
- Prompt-A-Video: Prompt Your Video Diffusion Model via Preference-Aligned LLMICCV 2025
- T2I-Copilot: A Training-Free Multi-Agent Text-to-Image System for Enhanced Prompt Interpretation and Interactive GenerationICCV 2025
- VisualPrompter: Semantic-Aware Prompt Optimization with Visual Feedback for Text-to-Image SynthesisICLR 2026
- Optimizing Prompts for Text-to-Image GenerationNeurIPS 2023
- RePrompt: Reasoning-Augmented Reprompting for Text-to-Image Generation via Reinforcement LearningICLR 2026
- VPO: Aligning Text-to-Video Generation Models with Prompt OptimizationICCV 2025
- Dynamic Prompt Optimizing for Text-to-Image GenerationCVPR 2024
- PromptRL: Prompt Matters in RL for Flow-Based Image GenerationICML 2026
- PromptCoT: Align Prompt Distribution via Adapted Chain-of-ThoughtCVPR 2024
- TIPO: Text to Image with Text Pre-sampling for Prompt OptimizationICLR 2026
- PROMPTMINER: Black-Box Prompt Stealing against Text-to-Image Generative Models via Reinforcement Learning and VLM-Guided OptimizationCVPR 2026
- PromptEnhancer: Taming Your Rewriter for Text-to-Image Generation via Fine-Grained RewardCVPR 2026
- Rethinking Prompt Design for Inference-time Scaling in Text-to-Visual GenerationCVPR 2026
- Compositional 3D-aware Video Generation with LLM DirectorNeurIPS 2024
Related open problems
Text-to-Image & Video Generation/Editing
Systematically Benchmarking and Calibrating VLM-as-a-Judge Failure Modes in Image and Video Generation
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.
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.