Developer Turns DiffusionGemma into an Image Generation Pipeline
A community developer built a text-to-image and image-manipulation pipeline for DiffusionGemma using LoRA, a trained image output head, and a vision decoder. The author says the adapted model produced recognizable images and improved visual manipulation when its generated visual latents were fed back into the next forward pass, and released the adapter, decoder weights, and code. The project is a self-reported experiment heavily assisted by LLMs, and the author describes the resulting system as rough and largely impractical.
The project demonstrates a source-supported method for adding native visual generation and manipulation to a text diffusion model, while also requiring an upstream Transformers patch and offering only preliminary, community-reported results.