SPIF is real-time AI image generation that runs at 15 FPS at 2K resolution in a browser. You speak, touch, or gesture — and a Stable Diffusion XL model renders live on screen. No app install. No GPU on the client. Just a WebRTC stream to any device with a browser.
We invented style-locked structural pruning. We take the full 12-billion-parameter SDXL model and prune it down to ~2.5 billion parameters while preserving a specific artist's style with near-zero degradation. This isn't quantization or distillation — it's architectural surgery. We remove the parts of the network that don't contribute to a given style, then fine-tune what remains. The result is a model that's 5× smaller, runs 5× faster, and looks indistinguishable from the original within its style domain.
Combination of three things: (1) style-locked pruning reduced compute per frame dramatically, (2) inference pipeline optimization on RTX 4090s with custom CUDA kernels, (3) WebRTC streaming architecture that decouples rendering from display. The bottleneck was never the network — it was the model. Once we solved that, everything else fell into place.
Training on A100s (RunPod). Inference on RTX 4090s (RunPod). A single 4090 serves one concurrent user at 15 FPS / 2K. Targeting 4–6 concurrent streams per GPU as we optimize further.
They can try. Style-locked pruning — optimizing architecture for a specific aesthetic rather than general capability — is our invention. The open-source community is focused on making models smaller in general. We make models smaller and style-specific, which turns out to be a much more tractable problem with much better results.
Our pruning pipeline is model-agnostic. When SDXL's successor drops, we apply the same methodology. The technique transfers. The artist style data transfers. We're not married to any single base architecture.
Zero friction. No downloads, no GPU requirements on the client, works on phones, tablets, laptops, smart TVs. A grandmother on a Chromebook gets the same experience as a designer on a workstation. Accessibility is a core value, not a feature.
Those are prompt-and-wait tools. You type words, you get a static image in 10–60 seconds. SPIF is a real-time creative instrument. You talk to it, touch it, sculpt it live. It's the difference between a typewriter and a piano. Different category entirely.
Tools for technical users who own GPUs. They require setup, troubleshooting, and expertise. Our audience is everyone else — the 99% who will never install Python. We also offer something they can't: a curated, licensed style marketplace where artists are compensated.
AI image generation market projected at $15–20B by 2028. Broader creative tools market ~$30B. Licensing/royalties for visual IP adds another layer. The intersection is where we live.
Adobe. They have Firefly, the creative suite, the enterprise relationships. But they're slow, they're corporate, and they can't do real-time. They also have a massive legal exposure problem with training data. We're clean.
Different problem. They're doing text-to-video (seconds of generated footage). We're doing real-time interactive generation (continuous, responsive, live). We're closer to a game engine than a video editor.
At current RunPod 4090 pricing (~$0.35–0.55/hr per GPU), serving one concurrent user at 15 FPS costs roughly $0.40–0.55/hour. A pro user doing 2 hours/day, 20 days/month = ~$18/month in compute. At $20/month subscription, that's tight. But: (a) utilization won't be 100%, (b) we're targeting 4–6 concurrent streams per GPU, dropping cost to $0.08–0.14/hr per user, (c) enterprise pricing is 5–10× consumer.
At 4 concurrent users per GPU, compute cost per user-hour drops to ~$0.10. A $25/mo pro user averaging 30 hours/month costs us $3 in compute. That's 88% gross margin on compute alone. We expect to hit this efficiency within 6 months of launch.
Artists submit their portfolio. We train a style-locked pruned model (the Stylus). When users create with that Stylus, the artist earns a royalty — percentage of subscription revenue allocated by usage time. Estate licensing is supported: a deceased artist's estate can authorize a Stylus and collect royalties.
Passive income from their artistic style without giving up their copyright. They're not selling their art — they're licensing the feel of their art. And unlike every other AI platform, they opted in and they get paid.
Patents are offensive weapons. We don't want to exclude — we want to prevent others from excluding us. The defensive publication on tdcommons.org establishes prior art for style-locked structural pruning. Nobody can patent this technique now. We keep the specific implementation as trade secret. Best of both worlds.
Current US law says AI-generated images without meaningful human creative input aren't copyrightable. But SPIF is different — the user is actively directing the output in real-time through voice and touch. There's a strong argument for human authorship when the tool is this interactive. Our interaction model positions us well.
Every Stylus is created with explicit artist consent. We don't scrape. We don't use LAION. We don't use anyone's art without permission. The base SDXL model has its own training data provenance (Stability AI's responsibility), but our style-specific fine-tuning and pruning uses only licensed, consented material.
They pull it. Immediately. No questions. The pruned model is deleted. Artist control is absolute.
Our style-locked models are trained only on opted-in artist work. We don't generate "in the style of [famous artist]" from the base model. If an artist isn't in our system, their style isn't available. Period.
Doug has been building interactive creative tools since before most AI researchers were born. ChipWits taught kids programming through AI in 1984. The King of Chicago was procedurally generated interactive storytelling in 1987. This isn't someone jumping on the AI hype — this is someone who's been doing this for 40 years and finally has the technology to realize the vision. Michal brings business discipline and enterprise credibility from three of the biggest tech companies on earth.
We need to hire: a senior ML engineer focused on inference optimization, a frontend/UX lead for The Pad interface, and a head of artist relations. Seed funding covers these first three hires.
Lean. RunPod infrastructure ~$2–4K/month during development. No salaries yet — founders are bootstrapping. Total monthly burn: ~$3–5K.
Seed round. Targeting $1.5–2.5M.
Beta with paying users within 6 months of funding. First enterprise contracts within 12 months. Target: $50K MRR by month 18.
18–24 months at planned burn. We expect meaningful revenue before runway runs out. We'll raise Series A at month 12–15 if traction supports it.
GPU costs. If inference costs don't come down (they will — every trend points down), margins get squeezed. Mitigation: our pruning technique inherently reduces compute needs. Second risk: slower-than-expected artist adoption. Mitigation: start with estates and established artists who have clear economic incentive.
Year 1: Launch The Pad with 50–100 Styluses. Prove the model. Hit $50K MRR. Build the community.
Year 2: Scale to 500+ Styluses. Launch enterprise product. API for third-party integration. Expand to video. Target $500K MRR.
Year 3: The Pad becomes the default creative platform for style-licensed generation. 2,000+ Styluses. Major brand enterprise accounts. International expansion. $2–5M MRR. Series B.
We become the platform layer between AI models and creative expression. Every artist has a Stylus. Every brand has a Stylus. Every person has access to real-time, style-authentic creative tools in their browser. We're building the Spotify of visual creativity — except artists actually get paid properly.
Yes. Style-locked pruning applies to video diffusion models. Real-time style-locked video generation is the logical next step. Prototyping in Year 2.
Roadmap for Year 3. As 3D generation models mature, our pruning technique applies. Style-locked 3D asset generation for game devs and designers.
We reject the "creator economy" framing where platforms extract value from creators. Artists aren't "creators" performing for an algorithm. They're artists. We pay them. They control their work. They can leave anytime. No engagement metrics, no algorithmic feeds, no dark patterns. Just tools that respect people.
They'll try. But: (1) they can't move this fast — we went from 2 FPS to 15 FPS in a week, (2) they have legacy products to protect and can't cannibalize, (3) they have massive legal exposure on training data that we don't, (4) artist trust is earned, not bought. Nobody's going to license their style to the company that scraped it without permission.
Because real-time AI creativity is inevitable, and we're the only team that's actually built it. Not demoed it. Not promised it. Built it. 15 FPS, 2K, in a browser, today. The technology works. The business model is clear. The team has 40+ years of shipping things that didn't exist before. And we're early enough that your money actually matters.