The recent decline in quality you’ve noticed in Google Stitch (Google’s AI UI design tool) after the update is certainly not an illusion. Since the major update in mid-March 2026—which primarily introduced the Gemini 3 engine and Infinite Canvas mode—there has been a flood of similar complaints within the global community.
While this update brought flashy features like “Instant Prototyping” and “Voice-Guided Design,” for many power users, “design stability” and “control over details” have indeed seen a significant decline.
1 Like
Hey @hong_guo, thank you for sharing your thoughts. We are actively tracking the issues users are encountering while preparing several exciting updates for the community. Our team is working hard on numerous new features designed to enhance your design generation journey, alongside many critical fixes. We truly value your continued support and collaboration.
I’m seeing the same behavior. While prototyping a mobile app, I briefly prompted it to switch to Material Design for comparison. After reverting to my original mockups, the tool kept reapplying Material Design patterns during further iterations.
It feels like the system is overfitting to recent prompts instead of respecting the current context/state. Before the mid-March update, iteration control was much more predictable, now it’s noticeably harder to further prototyping and needs to create multiple projects for one final design.
Hey @Rajat_Jain, I have a few tips that might help fix that styling issue! You can use the design.md feature to set hard rules that the system must follow, like telling it to avoid Material Design or to use specific hex codes and square corners. It checks this file before every new prompt, which should clean up the behavior. You can also use the Annotate tool to mark the parts you don’t like and tell the tool to switch those back to your original style. Trying a different mode might also help refresh the designs. Hope this helps!
1 Like
Hmm, I get what you’re saying, but during prototyping we typically don’t have well-defined design rules to document in a design.md. At that stage, the goal is more about exploration, we may have a general sense of preferences, but we’re intentionally trying different directions to see what works best.
Previously, the behavior felt more intuitive, as it would iterate based on the currently selected design. This made it much easier to explore multiple design directions within the same canvas.
At the moment, it seems to rely more on previously generated styles, which makes the process a bit less predictable and harder to steer during iteration.