Skip to content

SRM

Spatial reasoning with denoising generative models

Visit official site

What SRM is

The framework is described as a spatial reasoning approach built on denoising generative models. It assigns an independent noise level to each unobserved variable and infers continuous representations step by step.

What you can do with it

  • Solve visual tasks such as visual sudoku through denoising
  • Set custom noise levels to control the generation order
  • Choose among ordering strategies including a greedy heuristic

Application scenarios

  • Step by step reasoning on the MNIST sudoku dataset
  • Handling complex image distributions on Even Pixels
  • Testing visual reasoning on Counting Polygons FFHQ

Pros & cons

✓ What we like

  • Several ordering strategies are provided
  • Benchmark datasets are included for evaluation

! What to watch out for

  • The material describes it as a research framework
  • Ordering strategy and noise level need to be tuned

FAQ

What are the listed steps?

Visit the project page, get the code and pretrained models, use the benchmarks, then tune the settings.

What strategies are named?

An uncertainty based greedy heuristic and two stage noise sampling are listed.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review