🖼️PixArtAlpha ModelLoader

If you've ever stared at ComfyUI's models folder wondering where to drop a PixArt checkpoint, here's the answer: you don't. This node skips ComfyUI's model management entirely and pulls the whole model from the Hugging Face hub the first time you run it. The "fromhub" in the name doesn't mean an API call - there's no key, no cloud, no account. It just downloads weights to disk once and loads them locally. That's the entire job, and it's done well.

What it actually is

This is the entry point of the ZHO-ZHO-ZHO PixArt Alpha pack, and it exists because PixArt Alpha never shipped the way you're used to. It's not a .safetensors checkpoint you drop in a folder; it's a full diffusers pipeline - a T5 text encoder, a DiT transformer, and a VAE, bundled together as one directory on Hugging Face. The README was written in March 2024 specifically because the older ComfyUI_ExtraModels path for running PixArt had gone stale against newer ComfyUI, so the author built a diffusers-native alternative. PixArt is a T5-conditioned diffusion transformer - the architecture that made its prompt adherence famous before Flux showed up and took the crown - and this node loads that whole stack at once.

Mechanically it's one line: PixArtAlphaPipeline.from_pretrained(base_model_path, torch_dtype=torch.float16).to("cuda"). It downloads into your Hugging Face cache, caches it, and every run after the first is instant and offline. The output is a single pipe of custom type PAMODEL, and the only node that accepts it is the pack's PA_Generation_Zho. So the wiring is fixed: ModelLoader → (Styler in the middle for text) → Generation → your preview node.

The inputs and outputs

• base_model_path (STRING) - a Hugging Face repo ID. Default is PixArt-alpha/PixArt-XL-2-1024-MS (1024 = native resolution, MS = multi-scale), which is the right one to leave alone. You can point it at another PixArt Alpha repo, but don't point it at Sigma - Sigma uses a different pipeline class and won't load through this node.
• pipe (PAMODEL) - the only output. Wire it to Generation.

The gotchas are real

First run downloads roughly 22 GB of fp32 weights into your HF cache. That's the "C: drive space warning" in the README, and it's not a joke - on Windows the cache lives on C: by default, so check your disk before you start. If the loader looks frozen on first use, it's downloading; watch the console and the cache folder.

Second, this "0.6B model" is heavier than it sounds. The DiT itself is small, but the T5-XXL text encoder bundled in is the fat part, and the whole pipeline loads onto the GPU at fp16 in one go. Realistically you want a 12 GB card; an 8 GB card will struggle. There's no low-VRAM offload option here - it loads everything at once.

Because it bypasses ComfyUI's model management, there's no "is it downloaded yet" indicator in the UI - just the pipe appearing and the image eventually coming out. For the pack's shared install steps (ComfyUI Manager or git clone plus pip install -r requirements.txt for diffusers, transformers, and sentencepiece), see the pack README; this node is the one that pays for all of it on first use.