1. Classify/
2. Directory/
3. Multimodal/
4. Structured questions over images

Result

Typed image answers instead of generated captions

Jev multimodal use cases look like a form, not a caption. The model answers scene, subject, lighting, and visibility questions one field at a time.

The published demo compares a local 0.8B model against an OpenAI call on the same question list.

This directory lists it because the output is code-ready: every answer is a label the rest of the app can branch on.

Pipeline

1. Load an image
2. Ask a fixed list of scene and object questions
3. Return typed fields such as scene=indoors and person=yes
4. Compare a small local model with a hosted API on the same schema

Builder: Shin Megami Boson. Stack: jev, qwen, openai.