Replicate cjwbw/pixart-sigma

PixArt-Σ (Sigma) is a text-to-image model built on a Diffusion Transformer instead of the UNet that SD 1.5 and SDXL use. It was one of the early "DiT" models to show you could get strong prompt-following and 4K-capable output while training far more cheaply than the big labs. It's also the kind of model that never got the deep, first-class ComfyUI node treatment that SDXL did - which is exactly where this node earns its place. Through ComfyUI-Replicate it runs PixArt-Σ on Replicate's cloud, so you can drop it into a workflow without wrangling the weights and its text encoder yourself.

How it works

PixArt-Σ is a transformer that denoises image latents while cross-attending to text embeddings from a large T5 encoder - that big text encoder is a chunk of why its prompt adherence is good. The "Σ" version improved on PixArt-α with weak-to-strong training and higher-resolution output. You don't manage any of that here: the node sends your prompt and settings to Replicate, runs the model, and returns an IMAGE.

Inputs and outputs that matter

Output is an IMAGE.

• prompt / negative_prompt - your text. PixArt rewards descriptive natural-language prompts.
• width / height (1024) - output size.
• num_inference_steps (20) - a normal, non-distilled step count. 20 is a fine default; raise toward 30-40 if you want more polish.
• guidance_scale (4.5) - this is a proper full model, so it uses normal CFG (unlike the distilled Lightning/LCM nodes in this pack, which need CFG near 0). 4.5 is the recommended starting point.
• seed - fix it to reproduce a result.

force_rerun forces a fresh generation on identical inputs.

How to install it

ComfyUI Manager: search ComfyUI-Replicate, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/replicate/comfyui-replicate
cd comfyui-replicate
pip install -r requirements.txt

Restart, and set your Replicate token first:

export REPLICATE_API_TOKEN="r8_************"; python main.py

Token: replicate.com/account/api-tokens.

Common issues

Clean image output, fully supported by the pack - no audio/multi-output caveats.

The setting that trips people coming from the distilled nodes: don't zero out the guidance_scale here. PixArt-Σ is a full model and wants normal CFG (~4.5); set it to 0 like you would for a Lightning model and the output goes flat and washed out. Likewise 20 steps is a real 20 steps, not a distilled 4.

Beyond that, the standing pack facts: a missing REPLICATE_API_TOKEN is the first-run error, and every generation costs money on Replicate. This is one of the more defensible uses of the pack, honestly - PixArt-Σ isn't a one-click local node the way SDXL is, so calling it over the API to try it out or fold it into a graph is a reasonable trade rather than paying cloud rates for something you could trivially run at home.