In partnership with
Google Research published an explanation of Diffusion Controller on September 29. The method seeks better control over generated images. Its linked paper was submitted on March 7, so the current development is a research explanation rather than a newly completed experiment.
The problem is familiar in image generation. A picture can look plausible while omitting a requested attribute. Google uses a lizard wearing sunglasses as its example, describing the tension between satisfying a prompt and preserving visual quality.
200+ Proven Ways to Make Money With AI in 2026
The next wave of millionaires will be people who figured out how to make AI work for them.
The window to get ahead is still open. But not for long.
Here are 200+ proven ways to make money with AI in 2026.
Sign up for Superhuman AI, the free daily newsletter read by 1M+ professionals, and get instant access to all 200+ ways to profit from AI this year.
Its proposed controller adds a small trainable network beside an existing generator. In one configuration, the original model remains frozen. The correction changes the generation process without retraining the full underlying model.
Google presents this as a way to connect image guidance with control theory. It describes gradual corrections and a penalty for departing from the original behavior. The intended balance is closer prompt matching with less disruption to the baseline image.
The company’s explanation also discusses supervised training and two approaches using reward scores. Those approaches differ in how they learn corrections. A reward is an evaluation target, so its meaning matters when interpreting any reported improvement.
The original paper places a specific condition on restricted model access. Its controller needs intermediate denoising outputs, such as predicted noise or a reverse mean. Access to a finished image alone does not meet that described interface.
That condition narrows the broad idea of controlling closed models. A service exposing only final images is different from one exposing intermediate generation signals. The implementation question is therefore what the provider permits the controller to observe.
The paper experiments use Stable Diffusion version 1.4. Its methods include supervised training and reward based training. The reported results apply to that experimental backbone and do not directly validate every current commercial image model.
One evaluation compares controller outputs with pretrained images using Human Preference Score version two. The authors also use other automated metrics and human evaluations. These tests examine preference and image quality under defined prompts and training conditions.
Become an email marketing GURU.
Is your email strategy due for a tune-up? GURU Conference (powered by Constant Contact) is back Nov 12–13, 100% free and virtual. You'll get email tactics you can steal the same day from top B2B and B2C marketers. Last year, 29,000+ marketers showed up. Save your free spot.
The benchmark’s own documentation explains how to evaluate generated images with its score model. Its prompt collection includes photographs, paintings, concept art, and anime. This range helps describe the test material, but it does not cover every visual task.
A preference score ranks outputs according to a learned measure. It is useful to ask which prompts were used and which model version produced the score. The repository also distinguishes HPS version two from its later version 2.1.
Those distinctions make a single reported win rate insufficient for broad judgments. A reproducible comparison should identify the scoring version, comparison images, generation settings, and prompt set. Otherwise, two results can appear comparable while measuring different procedures.
The Google explanation discusses additional directions involving personalization, harmful content controls, and video generation. These are future research directions in the post. They should be evaluated as proposals, with separate evidence required for their eventual performance.
The practical interest lies in control with limited access. A separate correction network could let researchers study adaptation without modifying every generator parameter. Whether that works in a particular service depends on the exposed interface and the intended evaluation target.
This development adds an implementation question to the discussion of image quality. The September post makes earlier research easier to inspect, while its paper supplies the experimental boundaries. Future updates should be judged against those boundaries, helping readers assess what improved and what still requires testing.
The agentic era needs a different CRM. That’s Attio.
Teams like Parallel, Turbopuffer, and Wordsmith are already setting the pace on Attio. Get an always-on revenue engine, with agents and workflows that build pipeline, chase every buying signal, and move deals forward with your team. Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.
