How Rooze runs Stable Diffusion 1.5 in the browser (and makes 2GB models usable)

Rooze.ai runs Stable Diffusion 1.5 locally in your browser — including large community checkpoints — and still feels like a normal web app.

This post explains the practical problems you hit when you try to ship “browser SD1.5” to real users, and what Rooze does to make it workable.

The hard part isn’t “can it run?” — it’s “can people use it?”

Lots of things can “run” in a demo. Shipping to end users means solving boring-but-critical problems that decide whether the product feels trustworthy.

• Model size (often 1–2GB)
• Storage limits + clearing storage
• Download UX (progress, failure recovery, user control)
• Performance (fast enough to feel magical)
• Reliability (not crashing tabs, not leaking memory, consistent output)

High-level architecture

At a high level, Rooze follows a simple loop:

1. User picks a model
2. If the model isn’t local yet, Rooze downloads it
3. The model is stored locally in the browser (so it doesn’t re-download next time)
4. When you generate, inference runs locally using your machine’s GPU capabilities (where available)
5. Output is returned as an image in the UI (and batch mode repeats this loop)

Making multi-GB model downloads not terrible

When your “asset” is 2GB, normal web assumptions break down. Users need clarity and control.

What users need

• Clear progress (and a sense of time/size)
• The ability to cancel
• A way to reclaim storage later
• Confidence that “download once” really means download once

What Rooze provides

• A deliberate model download flow (not hidden behind the scenes)
• A dedicated Model Manager where users can see what’s installed and delete it

While models download locally and run on-device, the UX should still feel like a “normal web app” — smooth, predictable, and user-controlled.

Why the Model Manager matters

If a user tries one photorealistic model, then another, then another… you can eat multiple gigabytes quickly.

This is one of the biggest differences between a “cool demo” and something you can actually keep using:

• Users should be able to try models freely
• But also clean up easily when they’re done
• And understand what’s taking space

Performance notes (what makes it feel real)

In my testing, Rooze can generate:

• <30 seconds per image on a MacBook Pro M4
• <15 seconds per image on an M3 Pro

That’s not “cloud GPU fast,” but it is fast enough to be legitimately useful — and the tradeoff is huge: no install, no server cost, and on-device privacy.

Batch mode: why local inference changes the workflow

Batch mode is a first-class feature because the economics are different when generation is local:

• You can iterate prompts without thinking about cost
• You can generate lots of variations while you do something else
• It’s great for “explore first, curate later” workflows

As a rough mental model: in ~8 hours you can create ~1,000 high-quality photorealistic images on a standard MacBook Pro — completely free — because generation is running locally.

Limitations (being honest helps)

A few realities of browser-based SD:

• Performance varies a lot by device/browser
• Storage limits vary and can be surprising
• Very large models can be slow to download on poor connections
• Some older machines just won’t have a great time

Rooze is designed to make the best-case experience feel magical, while still giving users control (like deleting models) when the platform imposes limits.

What’s next

The near-term focus is:

• More model compatibility and smoother model management
• UX polish around generation + batching
• Better performance consistency and clearer “recommended settings” per device

If you have a device Rooze struggles on, I genuinely want to know — that’s how it improves.

Try it here: https://www.rooze.ai