FROM CLIP TO VISION-LANGUAGE MODELS
VLM Lab
Compare two models on the same image and question. Then change the evidence.
256M: about 575 MB · 500M: about 812 MB on first live load. Downloads are cached. Live answers run locally with WebGPU; recorded answers work offline. Larger does not guarantee a better answer.
Same evidence. Same question. Two models.
Choose a question above, then compare both models.
Reveal reference answer
Both runs use greedy decoding, up to 64 new tokens, repetition penalty 1.1 and no image splitting. Models load and run sequentially. Compare answer support first; timings depend on the device and warm-up.
What animal is shown?
Exactly the same prompt and decoding settings; only the image changes. Blank means a white image through the same processor.
Compare p(yt | y<t, I, P). Changes in answers show image sensitivity; they do not establish reliable grounding.
Raw model output
No run selectedNo model output yet.
Reveal reference answer
Show what happened · computation steps
These are observable runtime events and documented system stages, not hidden reasoning or attention maps. Image encoding, context assembly and prefill occur inside the first model forward pass; the library does not time them separately.
Technical metadata
Model not loaded.