DiffusionGemma, a text diffusion model built on a Gemma 4 backbone, can be adapted into a shitty image generation model. Maybe the ship sank? You may be asking yourself, "WTF, why would you want to do that?" And you would be right to ask. This is honestly kinda useless, but I still think it's pretty neat. Also, fair warning, being on a team of one means the project has been heavily — I'll be nice to myself and say — LLM assisted. I invite you to sidle on up to the trough and dig in to the slop! Still with me? Good, because I've been poking at this for a while, on-and-off since DiffusionGemma was released back in June. I was immediately hooked on this model as soon as I read its release blog and model card. 4x faster text generation they said — hell yeah! But here's the thing: this model sucks. I mean it's a great chat bot. It can do simple NLP tasks, and it's really quick for text generation as they say, especially for long-form output. But when it comes to agentic tasks and reasoning it's a total loser. 0% on ARC-3 (but that's not really a fair dig, this is a tough one) 0% on ARC-2 0% on The Gemma 4 Developer Agent Competition (this competition targets the base model DiffusionGemma was trained from) 11% on HLE on the model card (9.2 measured in deployment) 18.7% on SWE-bench Verified So not great, and particularly bad when you consider that on the hardware you'd need to run a good quant of DG (so basically bf16), you could easily run qwen-3.8-27B or a respectable FN quant. But please don't get me wrong. I absolutely love this model. Or at least I love monkeying around with it, and I doubt I'm the only one. Huge shoutout to the Unsloth team BTW! Reading their Colab notebook training DG to play sudoku-Sudoku.ipynb) is what made me start asking the questions that led to this project. Alright, enough fluff. So what did I actually do. Well the thing I'd headline is that I created a text to image and image manipulation training and inference pipeline for DiffusionGemma and then used that pipeline to adapt DG (like   submitted by