DiffUI

When a former Figma employee leaves to build a design tool, the default assumption is another canvas with better collaboration. DiffUI makes a different bet: that the canvas should be a staging ground for generated options, and that the underlying model should be a diffusion model, not an LLM. The tagline on the landing page is blunt: "Prompt to design, without the slop." The site's own comparison claims DiffUI generates a page in about 80 seconds, while "existing LLM-based competitors" take 3–5 minutes for a single output. That speed gap is the product thesis.

The case for diffusion over LLMs in UI design

Most AI design tools today are built on large language models that output code or structured UI descriptions. DiffUI instead leans on diffusion—the same family of models behind image generators—to produce pixel-level designs directly. The distinction matters: an LLM describes a button, but a diffusion model renders it, with all the texture, spacing, and visual noise that makes a design feel finished.

The landing page leans into this with a slider: "Diffusion-powered UI Design" sits next to a mock timer showing 80 seconds, while "Existing LLM-based competitors" shows 3–5 minutes. It's a stark comparison, though the site doesn't disclose the exact models or benchmarks behind it. Still, the architectural choice is clear. Diffusion models are inherently parallelizable—you can sample multiple candidates from the same prompt in one pass. LLMs, by contrast, generate tokens sequentially, making it harder to produce eight distinct design directions at once.

That parallelism is DiffUI's core workflow advantage. The product doesn't just generate one screen; it generates a grid of options, letting the designer pick, edit, and branch without leaving the canvas.

Eight directions at once: the parallel generation workflow

The canvas is the center of DiffUI's interface. A prompt node sits on the left, and image slots fan out to the right—up to eight per generation. The API documentation confirms this: the generate endpoint accepts slotNodeIds ("Image slot node IDs to fill (up to 8)"). This isn't a marketing gimmick; it's baked into the product's architecture.

From a single prompt like "A modern coffee shop landing page," DiffUI returns multiple options in parallel. The designer can then click "Ask for edits" on any option, or "Add to brand" to feed it back into the brand library. The canvas supports branching: you can generate new screens that reference existing ones, creating a tree of design decisions rather than a linear file.

The interface also exposes advanced controls that hint at the diffusion backend: sliders for SPEED, FUZZ, GRAIN, RIM, and entropy-like terms such as SYNTHESIS, DIMENSION, EXPANSION, WEIGHT, VARIANCE, BIAS, FREQUENCY, CONVERGENCE, and OUTPUT. These aren't typical design-tool controls; they're direct levers on the generation process, giving designers a level of control that a simple text prompt can't offer. It's a power-user feature that could intimidate newcomers, but it signals a tool built for people who want to tune the model, not just type a prompt.

Your brand, learned from a few screenshots

DiffUI's most distinctive feature is brand learning. The landing page shows screenshots of Reddit, Netflix, and Slack, and the tool generates a "landing page for this brand's new satellite connectivity feature" that matches each brand's visual language. The copy: "Upload existing screenshots and Diffui learns your design. Generate new features that look like they've always been there."

The mechanics are straightforward: you add screenshots to a brand library, and the model biases generation toward that style. The API includes a brand_id field in the generation request, and the canvas lets you "Add to brand" any generated image. This turns the tool from a generic generator into a brand-specific one—a huge practical advantage for teams that need to maintain consistency across features.

The site lists three promises: Consistent Components ("Buttons, tables, inputs—always on brand"), Unified Interactions ("Toggles, switches, and forms stay cohesive"), and Matching Data Visuals ("Charts and data displays, perfectly aligned"). These are the pain points of any design system, and DiffUI is positioning itself as a way to generate new screens that don't require a designer to manually match existing styles.

A pricing page that reads like an API bill

DiffUI's pricing is unusual: "13¢ per design," priced "at API cost, rounded up to the nearest cent." There are no tiers, no seats, no subscriptions. You top up a wallet and pay per generation. The fee breakdown is equally specific: 13¢ for a design, 1¢ for a standard image asset, 13¢ for high quality (which uses gpt-image-2), 3¢ for a pattern, 2¢ for an SVG, 1¢ for background removal.

This is a bold move. Most SaaS tools hide their infrastructure costs and charge a premium for convenience. DiffUI is essentially saying: we don't make money on the margin; we pass through the raw model cost and only round up a fraction of a cent. The site claims "No hidden fees" and "Unused credits stay in your wallet."

For a solo founder, this is a smart trust-building play. It lowers the barrier to trying the tool—no credit card required, pay as you go—and aligns DiffUI's incentives with the user's. But it also raises questions about sustainability: if the API costs rise, the price rises with them, and there's no obvious path to a profitable margin. The site doesn't disclose whether the 13¢ figure includes all model costs or just the inference, but the transparency is refreshing.

The open question: can a solo founder outrun the incumbents?

DiffUI's founder, known as jjcm, left Figma to build this. The blog post title is "Show HN: I left Figma to build a diffusion-based UI design tool." That's a strong signal of conviction, but it also means DiffUI is a solo operation competing against teams with massive resources. The site lists a Discord, an X account, and a changelog, but there's no team page, no funding announcement, and no public roadmap.

The product is real—the API docs are detailed, the canvas is functional, and the pricing is live. But the long-term question is whether a solo founder can maintain the pace of innovation needed to keep up with both incumbents like Figma and well-funded AI design startups. DiffUI's bet is that the diffusion-based approach is fundamentally different enough to carve out a niche, and that the parallel generation workflow plus brand learning will be sticky enough to retain users.

For now, DiffUI is a compelling proof of concept: a design tool that treats generation as the primary interaction, not an add-on. Whether it becomes a category-defining product or a footnote in the AI design story depends on execution. But the thesis is clear: diffusion, not LLMs, is the right engine for UI design, and the tool that gets there first might just win.