OpenAI has introduced a new method for AI image generation that focuses on a practical bottleneck: speed. The approach, called sCM, short for simplified, stabilized and scaled Consistency Models, is designed to make image models faster to train and faster to use while keeping image quality close to leading diffusion models.
The central claim is direct. OpenAI reports that its largest sCM model, with 1.5 billion parameters, can generate an image in 0.11 seconds on an A100 GPU without special optimizations. The company says that represents a 50-fold speed increase compared with conventional diffusion models.
What OpenAI changed with sCM
The new method builds on Consistency Models, or CMs, a class of diffusion-based generative models that OpenAI has been researching to improve fast image sampling. The source article describes sCM as a way to simplify, stabilize, and scale that earlier line of work.
The key improvement is in how the models are trained. Previous Consistency Models worked with discrete time steps. According to OpenAI, that setup required additional parameters and was error-prone. The researchers developed a simplified theoretical framework that brings several approaches together, helping them identify and address the main sources of training instability.
That matters because image generation systems often depend on many repeated computation steps to move from noise toward a finished image. OpenAI says the new sCM models can produce high-quality images in just two computation steps, while previous methods required significantly more steps.
Why two computation steps matter
Two computation steps is the most important practical detail in the report. A model that needs fewer steps can finish a generation task faster, and OpenAI’s reported timing shows how large that difference can become.
The largest sCM model cited in the source has 1.5 billion parameters and generates one image in 0.11 seconds per image on an A100 GPU without special optimizations. OpenAI compares that result with conventional diffusion models and describes it as a 50-fold speed increase.
The result does not mean every image model instantly becomes faster. It describes a specific method, a specific largest reported model, and a specific hardware setting. But it does show that OpenAI has found a way to sharply reduce the number of computation steps while still reporting strong image quality.
How quality compares with diffusion models
Speed alone would not be enough if image quality collapsed. The reported test results suggest that sCM keeps quality within a narrow gap of the strongest diffusion systems named in the source.
OpenAI reports FID scores of 2.06 on the CIFAR-10 dataset and 1.88 on ImageNet with 512x512 pixel images, using just two computation steps. By those metrics, the quality of the generated images is only about ten percent behind the best existing diffusion models.
That comparison is important because it frames sCM as more than a shortcut. The method appears to trade a small quality gap, measured by the cited metrics, for a much larger gain in generation speed. The source does not claim that sCM fully surpasses the best diffusion models on quality. It claims that it comes close while using far fewer computation steps.
Scaling is part of the breakthrough
The source article also emphasizes scaling. OpenAI successfully trained sCM models with up to 1.5 billion parameters on the ImageNet dataset. It describes that as an unprecedented size for this type of model.
The researchers also observed that image quality consistently improves as model size increases. That finding matters because it suggests the method is not only fast at a small scale. It can be trained at a much larger scale while continuing to improve.
The logical implication is that sCM may be relevant beyond the specific model sizes reported so far. The source says the method could work for even larger models. It also notes that the development may be important for the future of AI image generation, and potentially for video, audio, and 3D models as well.
What to watch next
The sCM method is significant because it addresses three problems at once: training complexity, model stability, and generation speed. OpenAI’s reported results show high-quality image generation in two computation steps, a 0.11-second generation time on an A100 GPU for its largest reported model, and scaling up to 1.5 billion parameters.
For AI image generation, that combination is the main story. The method does not remove the quality race against diffusion models, but it changes the speed side of the equation. If the same approach continues to scale, it could make fast image sampling a more central part of future generative systems.