Technical
Consistency Critic: An Academic Fix for the Reference-Drift Problem
ยท RenderBob team
ImageCritic, from a November 2025 paper, corrects a generated image against its reference after the fact. Public weights and a ComfyUI node exist. Commercial video APIs do not run it for you.

The previous article is the user-facing version of reference drift. The research response attacks it after generation, instead of waiting for every base model to preserve a reference on the first pass.
"The Consistency Critic: Correcting Inconsistencies in Generated Images via Reference-Guided Attentive Alignment" (Ouyang, Song, Liu, Zhu, Hou, Cheng, and Shou, arXiv 2511.20614) introduces ImageCritic. The method takes an image that has already been generated and corrects the regions that drifted from a reference.
Train on the kind of damage models already do
The authors build reference-degraded-target triplets. A vision-language model picks degradation that looks like the inconsistencies current generators produce, and they apply that degradation on purpose. Training then uses two mechanisms. An attention alignment loss pushes the correction model to attend to the relevant reference detail. A detail encoder targets the fine information that VAE encode and decode, and decoder-only stacks, tend to lose.
An agent can run the correction
The paper describes an inference chain that, given a reference and an image to fix, finds the patches and writes the prompt for a local edit. The stated prompt form is to use the object in the reference to correct, replace, or enhance the object in the generated image. They also say ImageCritic can sit inside an agent loop that detects inconsistencies and corrects them over multiple local rounds. The intended use is a pass that runs before a person sees the frame.
Code is public. The video APIs are not running it.
Weights are on Hugging Face, and the authors publish a ComfyUI node. That is further along than a PDF. It is still a research correction for images, evaluated on the inconsistency patterns the authors built. It is not a default pass inside Firefly, Kling, or Seedance. Watch for this class of post-generation check in production APIs. Until a vendor you already pay ships it, the habits in the next article are what get a sequence finished.
More from the blog
- The Same Video Models Keep Appearing Everywhere: Kling 4.0 and Seedance 2.5 Now in Adobe and Runway, Within Weeks
Kling 4.0 and Dreamina Seedance 2.5 are listed on Runway's API changelog and on Adobe's Firefly partner-models page. Shared names do not share clip limits or credit rates.
- Mistral's Voxtral 3: Speech-to-Text, Text-to-Speech, and a Voice-Agent SDK in One Release
Mistral released Voxtral 3 on 20 October, covering transcription, text-to-speech, and an SDK that chains listening, reasoning, and speaking. Test the transcripts first.