shawn-codedev/awesome-consistency-diffusion-visual-generation

This repository is a research reference list accompanying an academic survey paper about a specific problem in AI image and video generation: consistency. In this context, consistency means whether the AI-generated output actually matches what was asked for or what was shown to it. Researchers from several universities in China and the UK, along with contributors from Li Auto and ByteDance, compiled the survey and this resource list. The collection is organized around three types of consistency problems. External consistency is about whether the output follows instructions: if you tell an AI to generate an image of a red ball on a table, does it actually include a red ball on a table, in the right place, in the right color. Internal consistency is about whether things stay the same across multiple generated outputs: if you generate several views of the same person's face, or multiple frames of a video, does the face look like the same person in all of them. Normative consistency is broader, covering whether outputs are safe, fair, physically plausible, and match common sense expectations. For each of these three categories, the repository lists relevant research papers, benchmarks used to measure the problem, and datasets researchers have used for testing. Each entry has a short note explaining what specific consistency issue it was designed to address. Papers that have not yet gone through full academic publishing are labeled clearly rather than guessing their venue. The repository is primarily useful to researchers or developers working on AI image generation who want to understand what has been studied, what evaluation tools exist, and where the known failure points are. It is not software you run, there is no code to install. It is a curated reading list and reference collection in the format common to academic "awesome list" repositories on GitHub. The full README is longer than what was shown.