PixArt Checkpoint Loader (auto)

PixArt had a real reddit moment in 2024. Threads called it "the first model with complete prompt adherence that can be used locally," people seriously argued the whole community should pivot to it instead of waiting on SD3, and someone summed the whole situation up in a title that still holds up: "SD should move to Pixart sigma." It's a T5-conditioned diffusion transformer that hit SD3-class prompt following at a fraction of the parameter count - genuinely impressive for the size. Then the PixArt team got acquired by Nvidia, Flux launched a few weeks later, and the ecosystem investment that would've made PixArt a lasting alternative never fully materialized. It's still a good model. It's just one you're running mostly for yourself now, not joining a thriving LoRA/finetune scene around.

This loader is where a PixArt workflow starts.

What "auto" buys you

PixArt shipped in several variants over its life - 512px and 1024px alpha checkpoints, the Sigma revisions (512/1024/2K), and an LCM-distilled version - and each variant technically needs a different internal model config to load correctly. Earlier loaders in this pack made you pick that config by hand. This one is "simple" for a reason: it reads the checkpoint file itself and figures out which variant it is, so you just point it at a file and go.

• ckpt_name - a dropdown of files sitting in ComfyUI/models/checkpoints.
• Output: model - a MODEL object, ready to wire into a KSampler alongside T5-based conditioning (PixArt doesn't use CLIP at all) and the correct VAE for your variant.

Installing it

Part of the whole ComfyUI_ExtraModels pack:

• ComfyUI Manager - search "Extra Models for ComfyUI", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/city96/ComfyUI_ExtraModels, then pip install -r requirements.txt inside your venv. Restart.

Then grab a checkpoint: the alpha 1024px model (PixArt-XL-2-1024-MS.pth) from the PixArt-alpha HF repo is the README's recommended starting point, or one of the Sigma checkpoints (512/1024/2K) if you want the improved revision. Drop it in ComfyUI/models/checkpoints. You'll also need a T5 text encoder set up separately - see the T5v1.1 Loader or PixArt T5 from SD3 CLIP nodes.

Common issues

VAE mismatch is the trap most people hit first: alpha and the 512/1024 checkpoints share SD1.5's latent space and want an SD1.5 VAE, while Sigma explicitly needs an SDXL VAE - load the wrong one and you get visibly wrong colors and structure, not an error message telling you what's wrong.

xformers is technically optional but strongly recommended by the pack's own README - PyTorch's built-in SDP attention is only partially implemented for this architecture, and without xformers images look noticeably worse, not just slower to generate. If you're loading an LCM checkpoint, this node alone won't make it behave - you additionally need a ModelSamplingDiscrete node set to "lcm," the KSampler's sampler set to lcm, CFG dropped to roughly 1.1–1.5, and steps down around 5; skip any of that and an LCM checkpoint will look muddy or oversaturated even though it loaded fine.