Google's Diffusion Controller: the 90% win rate and gray-box access refer to different setups
Crescitaly · reddit · 2026-10-02
A Reddit analysis points out that Google Research's Sept 29 Diffusion Controller write-up reports its 90% win rate on the fully unlocked white-box variant (Stable Diffusion v1.4 backbone), while the gray-box controller is a different setup: frozen backbone but taking the intermediate reverse mean as input — more access than a typical prompt-in/image-out API provides.
- 'No weight editing' should not be read as 'no internal access needed'
- For developers, the access contract matters as much as the headline score; reports should pair each result with its exact backbone, controller variant, accessible signals, reward and inference cost
Author notes the analysis is AI-assisted and not reproduced.
More from Research
- DyRAD introduces first off-path evaluation protocol for radar novel view synthesis — orlitany · 2026-10-02
- EMBL-EBI's saezlab open-sources Karenina, a framework for multi-dimensional biomedical AI evaluation — anshulkundaje · 2026-10-02
- Nature Biotech: calibrated metrics show deep learning perturbation models can beat baselines — VectorInst · 2026-10-02
- arXiv papers grew 17% longer over the past decade, and AI may stretch them further — luismbat · 2026-10-02
- Suffix Cache Reuse: fixing KV cache for agents that edit context in place — RulinShao · 2026-10-02
- Satirical dialogue skewers RL environments: nobody actually reads the thousands of tasks and rubrics — samiramanabi · 2026-10-02