DigestAI news desk

Cut through the AI noise.

Research6 min read

Developer shares Python script using LLM logprobs for vision tasks

A developer has published a standalone Python script that leverages Large Language Model (LLM) token probabilities to perform simple vision tasks. The tool, inspired by the Jev framework and projects like OpenJev, uses the logprobs parameter in API requests to force models to answer single-token questions about images. By analyzing the probability distribution of the output tokens, the script…

1 source HN 43

Key points

  • Script uses LLM logprobs to answer single-token vision questions like brightness or presence.
  • Gemma 4 12B on RTX 3090 achieved 1 FPS; OpenAI gpt-6-luna achieved 0.2 FPS.
  • Method requires no specialized CV libraries, using only standard Python and OpenCV for capture.

The author tested the approach using Gemma 4 12B running locally via llama.cpp on an RTX 3090, achieving approximately 1 frame per second. They also tested it against OpenAI's gpt-6-luna, which ran at about 0.2 frames per second due to connection overhead. The script handles API differences between local servers and OpenAI, normalizing the returned log probabilities into usable scores. While specialized computer vision models are more efficient, this method offers flexibility by allowing users to define conditions in plain text rather than training new models.

Model page: GPT-6 Luna →

Full story from allanrbo.blogspot.com · via Hacker NewsOpen source ↗

A single function Jev-like wrapper for LLMs, including vision models

allanrbo.blogspot.com · 26 September 2026

Loading the full article…

This text was published by allanrbo.blogspot.com. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗

Coverage and discussion

1source
Topics · follow one to build your own front page

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.

Comments

via GitHub Discussions

More in Research

All →

Related stories