Painting with EmbeddingGemma 2: ~30 images/sec on Apple Neural Engine after 12h A100 training
GlennCameronjr · x · 2026-10-28
An early experiment uses Google's EmbeddingGemma 2 shared text-image embedding space to paint, generating an image at the end of every word. It runs at 30ms per image (30 fps) on the Apple Neural Engine of a base M1 Pro Mac, after just 12 hours of training on one A100 — still undertrained, but promising with more data.
More from Multimodal
- Nous Research ships ElevenLabs SFX plugin for Hermes Agent with looping and variation tools — NousResearch · 2026-10-28
- MiniMax H3 generates a surprisingly faithful Persona 5 anime scene for Halloween — Certain_Potato_4509 · 2026-10-28
- Google reportedly launches Genie 4, its strongest world model yet — oran_ge · 2026-10-28
- Free macOS App Concept Studio Streamlines Video-Model Dataset Captioning and Adapter Training — Ok-Republic764 · 2026-10-28
- AI-Generated Fake Dreamcast Cover 'Canyon Overdrive' Is Indistinguishable From Real Box Art — AIandDesign · 2026-10-28
- Open-source ComfyUI music system Plenio 1.3.2 adds phone player and party mode — Vivid_Promise1700 · 2026-10-28