Problem Statement
[ClawX webchat] Uploaded images over-compressed, degrading vision model accuracy
Images uploaded through ClawX webchat get compressed pretty hard before reaching the model. The result is that vision-capable models struggle with text in screenshots, miss UI details, and sometimes hallucinate content that isn't there — presumably because the image is too blurry to work with, so the model fills in the gaps.
Screenshots with small text or dense UI are hit the hardest. Reliable OCR on them is basically impossible.
What I'd expect: Something like max 1568px on the longest side with JPEG quality 85+ would be a reasonable tradeoff. Right now it feels like quality is being sacrificed way too much for file size.
Environment: ClawX on Windows, webchat channel
Proposed Solution
Alternatives Considered
Success Criteria
Additional Context
No response
Problem Statement
[ClawX webchat] Uploaded images over-compressed, degrading vision model accuracy
Images uploaded through ClawX webchat get compressed pretty hard before reaching the model. The result is that vision-capable models struggle with text in screenshots, miss UI details, and sometimes hallucinate content that isn't there — presumably because the image is too blurry to work with, so the model fills in the gaps.
Screenshots with small text or dense UI are hit the hardest. Reliable OCR on them is basically impossible.
What I'd expect: Something like max 1568px on the longest side with JPEG quality 85+ would be a reasonable tradeoff. Right now it feels like quality is being sacrificed way too much for file size.
Environment: ClawX on Windows, webchat channel
Proposed Solution
Alternatives Considered
Success Criteria
Additional Context
No response