CO2Jump sampler improves text-image generation consistency

A NeurIPS 2026 paper from Google, Google DeepMind and Stony Brook proposes the CO2Jump sampler, which uses text confidence and cross-modal attention to guide image updates and can regenerate low-confidence tokens; on editing, maze and nonogram tasks it improved both editing quality and grounding as sampling steps increased.

• Google Releases Open Multimodal Embedding Model EmbeddingGemma 22026-10-06
• Google launches Gemini 3.8 Live with Live Avatar2026-09-24
• Google DeepMind says Gemini 4 is nearly ready2026-09-24
• Google DeepMind adds secure server-side memory to Private AI Compute2026-09-23
• Google launches interactive AI & Economy ATLAS2026-09-15