The topic Open Design is replacing ComfyUI for my local AI workflows, and the difference is… is currently the subject of lively discussion — readers and analysts are keeping a close eye on developments.
This is taking place in a dynamic environment: companies’ decisions and competitors’ reactions can quickly change the picture.
Self-hosting AI started as a side curiosity for me and ended up being a real part of my stack I reach for now. Most of what I’ve used it for is text, so chatting, brainstorming, parsing documents, and analyzing images when I’m running something with vision like Gemma 4. Real design work was a different story though…
Local AI couldn’t touch what Claude Design or Google Stitch or Replit pulls off, and on the gen art side it was nowhere near Midjourney or Firefly either. So when I came across ComfyUI a while back, I was pretty stoked to finally have a local option that could actually generate visuals on my own hardware. Then Open Design dropped and changed what local visual AI even meant for me. It’s the first tool that lets me vibe design locally, which is a thing I didn’t know I’d been waiting for.
For anyone who hasn’t come across it, ComfyUI is an open-source, node-based interface for running open-weight image models like Stable Diffusion and Flux. You build the actual generation pipeline yourself by wiring blocks together on a canvas, and each one handles a specific job like loading the model, encoding prompts, sampling, decoding, and so on. It runs locally and the whole thing is community-driven, so it gets new model support pretty quickly compared to the closed alternatives.
For the longest time it was the only realistic way for me to generate visuals locally, design-adjacent or not. And I actually genuinely enjoy the node setup, in a way it’s good practice for UI/UX thinking – you’re building a flow with inputs, outputs, dependencies, and a clear order of operations, which is closer to wireframing logic than people give it credit for. So even when I wasn’t really using the outputs, I’d run it now and then just to mess with the graphs.
The catch is what came out the other end. ComfyUI can produce a static visual that you could maybe drop into a mockup as a placeholder, but it can’t give you something interactive or editable. There’s no version of a Stable Diffusion output that you can click into and rearrange or hand to a developer. And my 8GB GPU was already capping the quality of what I could generate, so the placeholder visuals weren’t even particularly strong placeholders. It’s a good tool, but it was solving a problem I didn’t really have.
ComfyUI did pick up an App Builder feature recently that lets you wrap a workflow into a simplified app interface, but the output is still whatever the underlying graph generates – an image or a video, not an editable prototype.
Claude Design has been my go-to for vibe designing for a while now, sitting alongside Figma Make and Google Stitch in the cloud rotation. Having design baked into Claude is honestly really convenient when I’m already in there for other stuff and I don’t have many complaints about the product itself. But a fair bit of my workflow is offline these days, and I like to keep an eye on what’s happening in the open-source space too, so Open Design landed at a pretty good time for that, dropping more or less right after Anthropic launched Claude Design as a direct alternative.
It runs as a native desktop app on macOS, Windows, and Linux, and unlike Claude Design it isn’t cloud-only. The basic premise is the same though – prompt in, real HTML/CSS/JS artifact out, with export options for HTML, PDF, PPTX, and even MP4. It doesn’t ship its own model and instead shells out to whatever coding agent CLI you’ve got on your system, or you can BYOK to any OpenAI-compatible endpoint, which is the route I took. It also ships with 250-plus skills and 140-plus design systems out of the box.
I’ve had a decent run with it locally by pointing it at Qwen 3.5 9B through LM Studio’s API, and the results were honestly fine for something running entirely on my modest machine. Gemma 4 didn’t really work though – it kept echoing the skill template back as broken XML, which I think is partly a vision-vs-agentic-work mismatch and partly my limited VRAM, but I’m not entirely sure. A cloud model is still going to outperform local on hardware like mine and that’s where the Anthropic API earned its place for me, with generations finishing in well under a minute and a higher-quality visual output.
ComfyUI isn’t really a designer’s tool. It has its uses if you specifically need a generated image and nothing more, and the appeal for me was always more about the setup than the output anyway. But just because something gives you a visual locally doesn’t mean it actually slots into a working design pipeline.
Open Design is a much better fit for me, plus the setup is a fraction of what ComfyUI demands – hooking up a model takes a couple of minutes and the actual generation side moves at a faster pace because it uses natural language instead of nodes. Nobody mid-deadline has time to be wiring up a node graph from scratch.
ComfyUI is still a solid tool and I’d recommend it to anyone who genuinely wants to generate images locally. It just isn’t always the right tool depending on what you’re actually working on. Open Design is doing something I didn’t realise I’d been wanting a local version of, and the fact that it exists at all is pretty cool. Both still live on my machine though.