OmniGen 2, an advanced multimodal AI generation model, has been released on Hugging Face with open weights under the Apache 2.0 license. The model features decoupled transformers for separate understanding and generation tasks, enhanced consistency through Omni-RoPE and isolated VAE processing, and self-correcting generation with editing capabilities. It supports a range of functions including image editing, in-context generation, text-to-image conversion, and visual understanding. Early demonstrations highlight its strong performance in image editing. The release has garnered attention for its potential to automate content creation, exemplified by a Reddit user who used a similar AI workflow to fully automate videos on Instagram, achieving approximately 5 million views in three weeks.
Wow! OmniGen 2 is quite amazing - State of the Art in Image edits - Apache 2.0 licensed 🔥 Bonus: can also do in context generation, text to image, visual understanding and image edits Play with directly on the demo below and models on the hub 🤗 https://t.co/AwRwQ5UtTU
Sorry for the dumb prompt but OmniGen2 seems super strong at image editing 🚀 ⬇️ Demo available on Hugging Face https://t.co/jLEwKxDknb
OmniGen 2 is out! Open weights with 🔀 Decoupled Transformers: one for understanding, one for generation 🎯 Enhanced Consistency: Omni-RoPE + isolated VAE processing allow for both editing & in-context generation 🤔 Reflection: self correcting generation + editing https://t.co/xtUUoB7USU