Generative Plugins
A Lens is an effect you play with in the Snapchat camera, like dog ears or a new background.
Expert creators build Lenses in Lens Studio Desktop, a professional IDE. I designed six ML and generative tools inside it, most of them on one shared design template. Easy Lens on desktop wasn't mine.
Problem
Lens Studio Desktop is a full creation IDE, with a scene hierarchy, inspector and timeline. Expert creators live there. But icons, 3D models and face effects meant leaving for external pipelines, or waiting on someone who could build them.
Where it started
Both started on the web, in the Easy Lens internal ML tool and beta. But expert creators build their Lenses in Lens Studio Desktop. So both tools moved there as plugins, and creators could use them without switching between tools.
ML Face Effects and Immersive ML work in the same four steps. The screens show ML Face Effects.
The creator describes the effect in a text prompt, or uses an image as the reference.
They can start from a preset. Presets set how strong the effect is, how much the person still looks like themselves — overall and feature by feature — and whether their skin tone stays the same.
The AI generates a preview first, on sample faces, so the creator sees the effect before committing to training, the slow step. To adjust it, they copy the settings into a new version, tweak them, and preview again.
When the preview looks right, the creator sends it to training. Training is the slow step, which is why the preview comes first. Shown at the end: an effect trained earlier, ready to import.
What held it back
They were designed one at a time, so each new tool would have been another flow to learn.
The shared template
Then four generative tools were on the way. I took what ML Face Effects and Immersive ML had in common and turned it into a template. Every tool on it works the same way. The creator writes a prompt, adds a reference if they want, gets previews, picks one and applies it to the Lens. 3D Asset Generation, Head Morph and Body Morph all run on it, so creators learn the flow once.
The creator describes the asset they want in a prompt.
The AI generates four 3D previews to compare.
Not quite right? The creator refines the prompt and regenerates, then picks the preview they want.
The finished model opens in a 3D viewer. The creator sets the polycount and texture size, then imports it into the Lens project. Shown here with an asset generated earlier.
Creators reshape an asset from a text or image reference. They change its geometry and texture, tune the intensity and preview it live before importing. Both tools follow the same template, with one preview instead of four.
Icon Generation
The Lens icon is one of the first things people see, so it got its own design outside the shared template. Creators generate it right in Project Settings: a prompt and a style give four options to pick from, then auto-crop and background controls finish it. A generated icon is a starting point they can keep refining, or they can import their own instead.
Outcome
| Feature | Before | After |
|---|---|---|
| Asset generation | Leave for external pipelines, or wait on someone who could build them | Generate inside Lens Studio Desktop |
| Face effects (as of spring 2024, at launch) | Internal ML engineers spent a couple of days creating and refining each effect, then shared it with creators as a template | Creators make their own effects |
| 3D assets (as of spring 2024, at launch) | A 3D engineer spent an estimated 16–20 hours building one asset for a Lens by hand | First internal designers and later any Lens creator, in about 15 minutes of creating and iterating on average |
Together with Easy Lens on the web, the plugins cover both ends: no-code creation for anyone, and AI generation for experts.
Took Snap's web Lens creation from an engineer-only ML tool to describing a Lens in plain words.
Creation on mobile was a bar people didn't notice. I gave it its own place, then designed how creators edit there.
