DeepSeek Vision Test: Auto-fills forms but struggles with box alignment
teortaxesTex · x · 2026-08-22
A tester evaluated DeepSeek's newly released vision model using a form-filling benchmark. The model was required to locate and fill fields based solely on a screenshot, without access to coordinates or the DOM.
Methodology & Results:
- Task: Identify form fields and simulate clicks/inputs using only visual data.
- Performance: The model probed positions, took screenshots of itself, placed elements, and nudged them into place. It filled every field and performed self-verification before submission.
- Stats: Took 5 minutes 48 seconds; consumed 288k input tokens and 38k output tokens.
Issues Identified:
- Checkmark ticks landed next to the box rather than inside it.
- Character-box drift occurred with longer text strings.
Technical Insight: The reviewer notes the model uses a "foveated vision" approach (max 387 tokens per glance), designed to "look" rather than "see." The non-Exp version is expected to be more video-oriented, extracting sequences for reasoning.
More from Multimodal
- fal launches Worlds for real-time creation and interaction with AI-generated worlds — Scobleizer · 2026-10-16
- Gummy worms as a physics test: Gemini 4 Argon vs. Claude Opus 5.5 elasticity sims — DynamicWebPaige · 2026-10-16
- 260M-parameter looped diffusion transformer beats 6x larger baselines with 5x less compute — burkov · 2026-10-16
- Seedance 2.5 demo: stretching one second of an explosion into a 30-second shot — LudovicCreator · 2026-10-16
- Octopus Short Film Made with Grok Imagine, Upscaled via Topaz — creatoroff · 2026-10-16
- Open-sourced 61-second computing history film generated entirely by Muse Spark — sophiamyang · 2026-10-16