Description Generation One Year Later. What We Managed to Understand Over the Year

Sergey Klyakhandler24 May 2025

Problem

Avito's first description generator took one photo and learned from 250 thousand human-written listings. It ignored the other photos, and its texts came out dry and short, because the model copied how people wrote before any generation existed. The same gap hurts search: when a seller leaves an attribute out of the title and description, full-text search cannot find the listing by that word.

Example from the article: a khaki jacket whose parameters name the colour differently, and whose title and description never say "khaki", cannot be found by the query "khaki jacket".

Solution in brief

Pass all listing photos to the LLM as image embeddings, and rewrite the training descriptions to be livelier and richer in photo details. Reuse the same multimodal model to generate attribute keywords from a photo for the search index, validate the labels so generic words drop out, and score only one image to keep the load manageable.

Questions this case answers
  • How to generate text from several images and a short author text with a multimodal language model so that the output is livelier and still factual
  • How to add image-derived attribute keywords to a search index for attributes that the author left out of the text
  • How to generate listing descriptions with a multimodal LLM so that sellers use them and buyers still get the facts
  • How to fix a generation model that writes dry, short texts because it was trained on human-written descriptions

Tools and methods

  • ViT image encoder (0.4B) + LLM decoder (1.3B) — image embeddings go into the LLM together with the prompt.
  • Relabeled training set — seller texts rewritten to add photo details.
  • LLM labeling with post-validation — about 200 thousand validated "image to characteristics" pairs.
  • Search index enrichment — generated keywords stored in Redis and added to the index.
  • Offline recall test and A/B tests — total found, buyer-seller pairs, empty results.
Continue reading — it's free

Google or a one-time email code. No password.