Critic-10K: 10,000 Image Triplets for Correcting Generative Inconsistencies

Critic-10K provides approximately 10,000 image triplets designed to train models to rectify inconsistencies in AI-generated visual content. Created by ziheng1234 and associated with the 2025 research paper 'The Consistency Critic', the data uses VLM-based selection to pair reference images with degraded and target versions.

Use Cases

• Training image-to-image diffusion models using the degraded and target pairs to learn error correction
• Fine-tuning Vision Language Models (VLMs) to identify specific visual inconsistencies between the reference and degraded images
• Benchmarking alignment algorithms using the reference-guided triplet structure to measure restoration accuracy

Strengths

• 10,000 curated image triplets
• VLM-based selection process for high-quality inconsistency modeling
• Structured reference-degraded-target format for supervised alignment

Limitations

• Small sample size of 10,000 records compared to general-purpose generative datasets
• Synthetic nature of explicit degradation may not capture all real-world model failure modes

Provenance

Source

The Consistency Critic: Correcting Inconsistencies in Generated Images via Reference-Guided Attentive Alignment (Arxiv 2511.20614)

Collection Method

VLM-based selection and explicit degradation of generated images

Freshness

Last updated December 2025; reflects current state-of-the-art methods in VLM-based image correction.