DreamOmni2: Unifying Image Generation and Editing Through Multimodal AI (www.dreamomni.net)

DreamOmni2 is a unified multimodal image generation and editing model that accepts both text and image prompts to create or modify visuals with a single system. It supports in-context multimodal instructions, multi-reference compositional synthesis (three- and four-reference generation), and a suite of industry-style edits — object/background replacement, pose imitation, lighting render, hair and font imitation, and more — while prioritizing identity and pose consistency across iterations. The interface workflow is simple: upload reference images, add natural-language instructions, and generate high-resolution outputs; creators report robust style and identity preservation and seamless mixing of visual and semantic guidance. For the AI/ML community this signals progress in controllable, multimodal conditioning and disentangled attribute control: one model handling both generation and fine-grained edits reduces pipeline complexity and suggests stronger latent representations for pose, lighting, and identity. Practical implications include faster creative iteration for design, advertising, product visualization, and typography while enabling research into evaluation metrics for cross-edit consistency, multi-reference fusion, and multimodal instruction alignment. If broadly adopted (and—per some user quotes—open-sourced), DreamOmni2 could become a useful baseline for future work on unified multimodal editing/generation systems and real-world creative tooling.