Using closed-decision Jev-like models to kill tag hallucination in dataset captioning

Iory1998 · reddit · 2026-09-21

A Reddit user proposes replacing free-form LLM captioning with closed-decision 'Jev-like' models for training dataset tagging.

Problem: LLMs hallucinate out-of-vocabulary Danbooru tags and drift across batches, which breaks models like Illustrious and Anima that only respond to tags in their learned vocabulary.

Core idea: Invert open questions into closed ones (photo or illustration? hair color from a list? character count 0/1/2/3+?). Each model pass picks from a fixed state list with probabilities; only above-threshold tags enter the caption — every dimension gets answered and zero invented tags. To save compute, a cheap first pass shortlists candidates, then a Jev-like pass confirms each against the image. Finally, the frozen tag list is handed to an LLM purely as a translator to produce a natural-language description, so one captioning run yields both Booru tag strings and prose for either training architecture.

Caveat: it can still affirm a real but absent tag — bounded failure, with the threshold as your recall/precision dial. Runs in ComfyUI, no fine-tuned captioner or API. The author hasn't built it yet and invites feedback.

Related event: Reddit Proposal: Closed Decision Models to Fix Dataset Captioning Hallucinations(2 posts)→

Original post →

More from coding & agent

coding & agent channel →