AI Suite

AI Enhance cluster

Face restoration, denoise, deblur, colorize, relight, HDR tone-map, auto-grade — classical baselines with 0 MB download, plus GFPGAN face restoration on the Pro tier. Non-destructive, lands as new layers.

The Enhance cluster handles the photo restoration work that used to require desktop apps and skill: recovering small faces, killing noise, fixing motion blur, colourizing B&W, re-lighting compositing subjects to match a scene, lifting shadow detail, and translating a vibe word into a concrete adjustment recipe.

Most tools here live in the Lite or Standard tier; nothing exceeds 350 MB on its own.

The seven tools

AI Face Restore. Classical CLAHE + bilateral + unsharp pipeline (Lite tier, 0 MB download) recovers detail in small or blurry faces — group photos shot from across the room, screenshots of older video calls, scanned family portraits. Apply to a face bbox for local enhancement or to the whole image for a global pass. Each restored face lands as a non-destructive layer so you can dial the blend down if a result is too aggressive. For learned face restoration, opt into the Pro tier: on a WebGPU device, Face Restore runs GFPGAN as a managed ONNX model (downloaded once, cached), falling back to the classical pipeline when no GPU / Pro opt-in is present. You can also bring your own ONNX URL on the Pro tier.

AI Denoise. Edge-preserving bilateral filter (Lite tier, 0 MB download) strips ISO grain, sensor noise, JPEG artefacts, and scanner speckle without smearing edges. Profiles for low-light, iso-grain, jpeg-artefacts, and scanner-speckle tune the filter's range sigma to the noise type, or auto runs a quick statistical analysis to pick.

AI Deblur. Unsharp mask + directional sharpening (Lite tier, 0 MB download). Estimates motion-blur angle via directional gradient analysis, then applies a second-pass sharpen perpendicular to the motion direction. Recovers modest handheld shake and focus-miss; cannot save catastrophically blurred shots — those need a true learned deconvolution model.

AI Colorize. LAB-space sepia + palette-bias colouriser (Lite tier, 0 MB download). Five style presets (realistic, vintage warm, sepia tint, modern saturated, pastel) plus optional reference palette for guided results. For learned colourisation (e.g. DDColor — not yet served as a clean web ONNX), bring your own ONNX URL via the Pro tier, which is live; until a hosted model is wired in, the classical path is the default.

AI Relight. This is the single biggest fix for the "pasted-on" look in lifestyle composites. Estimates scene light direction via the existing depth-anything-small model (already cached if you've used portrait blur), then applies a shading-shader pass to re-light the subject to match a target preset (golden-hour, studio-softbox, dramatic-side, overcast-even) or the actual scene the subject will composite onto.

AI HDR. A small CNN (~40 MB, Lite) tone-mapping pass that pulls detail back into blown highlights and crushed shadows from a single-exposure source. Not true multi-exposure HDR (that needs a bracketed source), but visually similar to the result of one. Four looks: natural, punchy, subtle-lift, print-hdr.

AI Auto-Grade. Type a vibe word — "cinematic", "Apple keynote", "moody noir", "Wes Anderson", "bright airy" — and the existing WebLLM tier translates it into a GradeRecipe: brightness / contrast / saturation / temperature / tint / highlights / shadows / optional filter + intensity. The recipe applies to the editor's adjustment store so it lives in the normal undo/redo history.

How they integrate with the editor

The Enhance toolbar dropdown (✨ in the floating canvas toolbar) gives one-click access to Face Restore, Denoise, Colorize, and Auto-Grade. Results land as adjustment-store changes or new non-destructive layers so nothing is destructive.

In the editor sidebar, the AI Suite tab → Enhance section surfaces all seven tools with preset chips for the common cases. Clicking any chip opens the AI panel with the prompt pre-filled.

The Enhance cluster is also wired into common workflows: the restore-old-photo recipe chains Denoise → Face Restore → Colorize with auto-tuned strengths based on what the vision model detects (heavy noise → stronger denoise, faded sepia → enable colorize, sharp scan → skip colorize). One click runs the whole pipeline.

Performance expectations

ToolTypicalMax
Face Restore1.8 s/face6 s
Denoise0.8 s3.5 s
Deblur1.0 s4 s
Colorize1.5 s5 s
Relight0.9 s3 s
HDR0.7 s2.5 s
Auto-Grade0.8 s4 s (LLM-dependent)

All numbers are for a mid-range WebGPU device (M2 / RTX 3060). Lower-end hardware falls back to WASM and runs ~3–5× slower.