PixArt-δ is positioning itself as a faster and more controllable open-source image generator, with performance claims that put it in direct comparison with the Stable Diffusion family. The new framework improves on PixArt-α by targeting high-resolution output, shorter inference, and better prompt following without moving away from an accessible open-source direction.
What PixArt-δ changes
Researchers from Huawei Noah's Ark Lab, Dalian University of Technology, Tsinghua University, and Hugging Face presented PixArt-δ as an advanced text-to-image synthesis framework. It is described as a significant improvement over PixArt-α, which was already able to generate 1024 x 1024 pixel images quickly.
The central change is the integration of the Latent Consistency Model (LCM) and ControlNet into PixArt-α. That combination is meant to accelerate inference while also giving users more control over what the model produces.
In practical terms, PixArt-δ can generate high-quality images at a resolution of 1,024 x 1,024 pixels in just two to four steps. The reported generation time is as little as 0.5 seconds, which the source describes as seven times faster than PixArt-α.
That matters because speed and resolution are often competing priorities in text-to-image systems. A model that can keep a relatively high resolution while reducing the number of steps can make image generation feel more immediate, especially for workflows where users iterate repeatedly on prompts and reference material.
How it compares with SDXL Turbo
The comparison point in the source is SDXL Turbo, introduced by Stability AI in November 2023. SDXL Turbo can generate 512 x 512 pixel images in one step, or about 0.2 seconds.
PixArt-δ is not presented as faster than SDXL Turbo on raw time alone. Instead, the distinction is resolution and apparent consistency. PixArt-δ produces 1,024 x 1,024 pixel images, while SDXL Turbo is described in the source at 512 x 512 pixels.
The source also says PixArt-δ's results seem more consistent compared to SDXL Turbo and a four-step variant of SDXL with LCM. The images appear to have fewer errors, and the model follows instructions more accurately.
Those differences are important for users who care about usable output rather than only benchmark speed. A slightly longer generation time can still be valuable if the resulting image is larger, cleaner, and closer to the written request.
Why LCM and ControlNet matter
The Latent Consistency Model is the component credited with helping PixArt-δ reduce the number of steps needed for image generation. Fewer steps can translate into faster inference, which is the process of producing an image from a prompt.
ControlNet adds another layer: controllability. In PixArt-δ, the ControlNet module allows finer control of text-to-image diffusion models by using reference images.
The researchers introduced a novel ControlNet architecture for transformer-based models. According to the source, the goal is explicit controllability while maintaining high-quality image generation.
For image generation systems, this kind of control can be just as meaningful as speed. Text prompts alone can be imprecise, while reference images can help guide structure, composition, or other visual constraints. PixArt-δ's ControlNet integration is aimed at making that guided generation more practical within the framework.
Training and hardware access
The new PixArt model is also designed around efficient training. The source says it can train on V100 GPUs with 32 GB of VRAM in less than a day.
It also supports 8-bit inference. That capability allows it to synthesize 1024-pixel images even on 8-GB GPUs, which improves usability and accessibility for users with more limited hardware.
The hardware details are a key part of the open-source story. A model may be technically available, but if it demands unusually large hardware resources, the number of people who can actually run it becomes much smaller. PixArt-δ appears designed to reduce that barrier for both training and image synthesis.
The source highlights three practical advantages together:
- 1,024 x 1,024 pixel image generation in two to four steps.
- Reported generation times as low as 0.5 seconds.
- 8-bit inference that works for 1024-pixel images on 8-GB GPUs.
Taken together, those claims make PixArt-δ relevant not only as a research model, but also as a potential tool for developers and creators who want faster open-source image generation on more accessible machines.
What is available now
The researchers have published the weights for the ControlNet variant of PixArt-δ on Hugging Face. That gives users access to the controllable version of the model described in the source.
However, the online demo situation appears more limited. The source says an online demo seems to be available only for PixArt-α with and without LCM.
That means PixArt-δ is not presented as a fully polished public demo experience in the source. The more concrete availability point is the publication of the ControlNet variant weights on Hugging Face.
The broader significance is clear: PixArt-δ combines speed, high resolution, open-source availability, and stronger control mechanisms in one framework. If the reported consistency and prompt-following improvements hold up in wider use, it could become a meaningful competitor to established open-source image generators in the Stable Diffusion family.