Imagen 1 (2022) Should Be Open Sourced
First of all, huge W on Gemma 4 — the 31B release is genuinely impressive and the open source community appreciates it massively.
With that said, I want to use this space to bring up something that the research community has been wanting for a long time: could Google DeepMind consider open sourcing the original 2022 Imagen model weights?
The reasoning is simple — Nano Banana 2 (Gemini 3.1 Flash Image) is now your active image generation flagship, deployed across Gemini, Search, Flow, Ads and AI Studio. The 2022 Imagen model is multiple generations behind and is not competitive with anything in your current lineup. Releasing it carries essentially zero commercial risk at this point.
Stable Diffusion launched around the same time and the open source community ran with it — producing FLUX, ControlNet, and thousands of fine-tunes. Imagen had comparable or superior architecture but was never accessible for study. Students and researchers are still trying to replicate results from the paper when they could just have the real weights.
You just proved with Gemma 4 that Google is committed to open source. Dropping the OG 2022 Imagen weights on HuggingFace would be the same energy — low cost, massive community goodwill, huge research value.
Please make it happen 🙏