Google introduced an on-device interactive segmentation feature for Snapseed on iOS on October 1, 2025. The update adds Object Brush for selective editing with real-time feedback. Details on the rollout and technology are in a Google Research announcement by Ben Hahn and Florian Kübler.
Key details
- Object Brush appears in Snapseed’s Adjust tool on iOS. Draw a stroke over an object to isolate and edit it. Google says the feature will roll out over the coming week.
- The feature runs entirely on device using the Interactive Segmenter model, built on MediaPipe with LiteRT’s GPU acceleration for inference.
- Performance: object selection completes in under 20 ms after a user prompt. The decoder runs in 7.4 ms on an iPhone 16 Pro.
- Models: student models are quantized to 8 bits and execute on the GPU. A heavyweight image encoder runs once per image, and a lightweight decoder responds to prompts.
- Prompts: supports foreground and background strokes, points, and boxes. Users can add or subtract areas to refine selections.
- Masks: the model predicts a 768×768 mask, then upsamples to image resolution up to 4K. An edge-preserving joint-bilateral upsampling method runs after a gesture completes to reduce latency.
- Training: a teacher-student approach using knowledge distillation. Annotators created about 30,000 masks across more than 350 categories for teacher fine-tuning. Distillation used a weakly labeled dataset with over 2 million images and generated realistic prompts. Google reports similar IoU performance between Edge and Teacher for fixed prompts.
- Availability: the model also ships on the Chromebook Plus 14 for Gallery app editing. Google plans broader integration across Google products and more Snapseed tools.
Background
Snapseed is Google’s mobile photo editor on iOS. Object Brush adds object-level selections to existing tools, enabling fast, precise edits directly on device.
The pipeline relies on MediaPipe for on-device ML and LiteRT’s GPU acceleration for real-time inference while preserving mask quality with joint-bilateral upsampling.
Sources
- Google Research announcement
- Snapseed on the App Store
- MediaPipe documentation
- LiteRT’s GPU acceleration documentation
- Chromebook Plus 14 product post
.svg)
.svg)
.svg)