← sources

Jev mode for images!

source: reddit (reddit:LocalLLaMA)published 2026-09-29 10:21ingested 2026-09-30 02:44

retained text

So Codacus created Jev mode for Lllama.cpp , and I thought Why not extend this concept further and ask questions about images and have the constrained answer be an image selection? So i spun up an agent and added image support and a harness. Now you can use images as your prompt without the decode step, no caption pause, just a decision based on an image or group of images. Ask the same question for a batch of images, like clasification. OR hand 1 context a whole group of images and ask it to pick on. like which of these 20 images has a ruber duck?   
[https://github.com/thecodacus/llama.cpp/pull/17](https://github.com/thecodacus/llama.cpp/pull/17)

Youtube explainer using Codacus own RenderDiv framework to create the video.   
[https://youtu.be/Xuw3la2zVpg?si=rtSAydhuF9n3SYWV](https://youtu.be/Xuw3la2zVpg?si=rtSAydhuF9n3SYWV)

document fetched 2026-09-30 00:46 · canonical https://reddit.com/r/LocalLLaMA/comments/1wt6f3d/jev_mode_for_images