Built my side project with DeepSeek V4 Pro. Where it struggled was the UI, because it couldn't see what it was making

Some feedback from actually shipping something with V4 Pro , in case it's useful to anyone here. When V4 Pro launched I used it to build FauxTalks ( fauxtalks.com ) , a fake chat screenshot generator. It recreates the chat screens of WhatsApp, iMessage, Discord and about 17 other apps . Plain JavaScript , no framework , everything runs in the browser . Where it fell short was the UI. The whole point of the site is that each chat looks exactly like the real app, and V4 Pro can't take images. S o I couldn't show it a screenshot of the real WhatsApp next to its version and say "fix what's off". It had no way of knowing how close it was. For that part I switched to Gemini, the 3 Flash models and 3.1 Pro, to generate the mockup UIs. Because they could look at screenshots, they iterated on the UI much more aggressively: compare, spot the difference, fix, repeat. So my feedback: for anything visual, a model that can't see images is basically working blind . I'd like to try DeepSeek's new vision model on exactly this. Give it the real app screenshot and my render, and see if it closes the gap. Has anyone here tried it on UI work yet?