Guide
The AI Suite, its tiers, and “will it run on my device?”
The AI Suite runs entirely in your browser — nothing you upload ever leaves your device. But “runs in your browser” covers everything from an instant filter that needs no download to a 1 GB diffusion model that wants a real GPU. This guide explains the three tiers, what each downloads, how to work offline, and exactly what hardware you need.
The three tiers at a glance
| Tier | Download | Runs on | What you get |
|---|---|---|---|
| Lite | None (0 MB) | Any device, even phones | Classical filters, adjustments, smart-crop, quality score, A/B variations, classical denoise / deblur / colorize / face-cleanup, the procedural text-to-image composer, and classical scene / highlight / reframe heuristics. |
| Standard | ~400 MB (per model, on demand) | Most laptops; GPU helps, WASM fallback | Real ML models that run offline after one download: background removal, CLIP tags / categorize / similarity, captioning, OCR, document Q&A, depth, Whisper subtitles, pose / segmentation. |
| Pro | ~2 GB+ | A WebGPU-capable GPU | GB-scale generative + restoration: SD-Turbo text-to-image, GFPGAN face restoration, plus bring-your-own ONNX model URLs. Falls back to the classical path where a GPU isn't available. |
Nothing downloads until you opt in. The Lite tier is the default and never costs you a byte of bandwidth. You only fetch a model the first time you actually run a tool that needs one, and it's cached afterwards.
Will it run on my device?
- Any phone or low-end laptop — the entire Lite tier works. Background removal's Fast model and most Standard tools also run, just slower (they fall back to WebAssembly on the CPU).
- A typical modern laptop (8 GB+ RAM) — the full Standard tier is comfortable. Captioning, OCR, CLIP, depth, and Whisper all run; a discrete or integrated GPU shortens each run.
- A WebGPU GPU (recent Chrome/Edge, Apple Silicon, RTX-class, or modern integrated graphics) — unlocks the Pro tier: SD-Turbo and GFPGAN run locally. Without WebGPU these tools cleanly fall back to the classical path instead of erroring.
You don't have to guess. Open the Model Manager (the database icon, top-right of any page). It probes your browser for WebGPU support and available memory and tells you, in plain English, the highest tier your device can comfortably run — and why.
How model downloads & caching work
Models are fetched from the Hugging Face CDN (or, for bring-your-own, the URL you supply) and stored in your browser's Cache API — under transformers-cache for Transformers.js models and nss-managed-onnx for managed ONNX models like SD-Turbo and GFPGAN. Only model weights are downloaded; your images, video, and audio are never sent anywhere. After a model is cached, re-running a tool is instant and works with no internet connection.
You can review what's downloaded, see how much space it uses, and delete any or all of it from the Model Manager at any time. Deleting a model just means it re-downloads next time you opt to use that tool.
Working offline
Once the app and any models you use are cached, the AI Suite works offline. Install it as a PWA (the install prompt, or your browser's “Install app”), opt into the tier you need while you're online so the models download once, and from then on you can caption, remove backgrounds, transcribe, and more on a plane or with the Wi-Fi off. The Lite tier is offline from the very first load because it downloads nothing.
Bring your own model (Pro)
On the Pro tier you can point the generative and restoration tools at your own hosted ONNX model URL. Your browser downloads and runs that model directly via onnxruntime-web — we never see it or your data. You're responsible for the model's licence and trustworthiness; see the Terms for the bring-your-own clause.
Explore the AI Suite
90+ tools across five clusters — all client-side, all private.