image + context
HTTPS URL or base64 image. Optional caller-supplied OCR and layout metadata.
●Image decision API // independently evaluated
Every image.
One decision.
Laya Vision turns an image and a bounded question into a typed answer with a probability. We ran the complete public POPE benchmark to show exactly where that promise holds—and where it does not.
IMAGE / 224²
LAYAVISIONENCODER
YES / NO CHOICE SCORE
SHARED VISUAL PREFIX // TYPED DECISION BATCH
POPE QUESTIONS9,000complete public suite
OVERALL ACCURACY81.38%local verified run
MEDIAN LATENCY112.9 mssingle question / RTX 3090
GROUPED THROUGHPUT112.9 q/s18 questions / image
PEAK VRAM493 MBallocated
●What it does
Visual decisions,
not visual essays.
Laya Vision is most useful when the output space is known: a choice, a score, or a yes/no probability. It does not generate an open-ended answer and the benchmark does not prove OCR, fine-grained localization, or safety-critical reliability.
●Decision console // precomputed
One image. Typed answers.
CITY / 01
X 51.4 Y 49.8
YES_NO
Is there a bicycle in the image?
YESP = 0.91
0.25.50.751.0
Precomputed illustrative UI example. No live inference is represented.●How it works
One visual state. Many questions.
Encode the image once, fan out bounded questions, and return probabilities your application can threshold. The contract stays visible; uncertainty stays machine-readable.
one shared prefix
Questions about the same image are grouped to amortize vision work.
typed results
Yes/no, choice, or bounded score with a probability for every decision.