Turning text-diffusion DiffusionGemma into an image generator: <300M params, weirdly decent
bumblebeer · reddit · 2026-11-13
A solo developer (heavy LLM assistance, by his own admission) adapted Google's text diffusion model DiffusionGemma (Gemma 4 backbone) into a crude image generation model under 300M total parameters, open-sourcing the full text-to-image and image-manipulation training/inference pipeline. The spark was Unsloth's Colab notebook training DG to play sudoku.
He first roasts DG itself: 4x faster text generation, but 0% on ARC-3 and ARC-2, 0% on the Gemma 4 developer competition, 9.2% HLE measured, 18.7% SWE-bench Verified — on the same hardware you could run far better models.
Key trick: DG refused to make images until he prefilled prompts up to the input string of an "image generation tool" described in the system prompt — with DG itself doing the generation, not any external tool. It also required patching an upstream Transformers image-attention bug (PR #50684). Because DG emits in its native visual latent space (256 tokens at a time), it can feed its in-progress visual representation into the next forward pass — on-domain block autoregressive — which measurably improved quality and generalization on image manipulation tasks.
More from Multimodal
- Community shrinks open-source MiniMax-H3-Image from 68GB to 6.4GB, 110k downloads — alexcovo_eth · 2026-11-13
- Custom motion-director skill turns any voiceover into 1080x1920 motion-graphics animations with Gemini Spark — mijozakinho_2 · 2026-11-13
- MiniMax's H3 Video Model Lands on Envato, Free and Unlimited for 24 Hours from Nov 19 — cneuralnetwork · 2026-11-13
- Open Replication of DeepMind's Vision Banana Beats Original on Cityscapes — y_m_asano · 2026-11-13
- Envato offers free unlimited MiniMax H3 access for 24 hours starting Nov 19 — EXM7777 · 2026-11-13
- Synthesia launches interactive AI clones: record a short video, get a talking you in 2 minutes — synthesiaIO · 2026-11-13