CLIP Text Encode (PixArt Alpha)

CLIPTextEncodePixArtAlpha looks like a plain text encoder, but it's smuggling one extra feature: it also tells PixArt Alpha what resolution to draw at. That's the "sets the resolution conditioning" part that the official description warns about. PixArt Alpha - the 2023-era PixArt-α image model, not the newer Sigma - encodes its target resolution into the conditioning itself, so a 1024×1024 generation and a 512×768 generation genuinely are different conditioning, not just a bigger canvas.

If you've used the standard CLIP Text Encode node with PixArt Alpha and wondered why results were off, this is why. The generic node doesn't set the resolution conditioning, so the model is left guessing. This node exists precisely to close that gap.

How it works

Mechanically it's a single tokenize + encode, with one addition: width and height are passed into the encode as extra conditioning values (add_dict), and PixArt Alpha's model code reads them to set its internal resolution embeddings. Everything else is a standard text encode - same text and clip you'd feed any CLIP encoder.

Inputs and output

• clip - the PixArt Alpha CLIP (from your checkpoint or a CLIP loader).
• text - your prompt, multiline.
• width / height - the resolution to condition on. Defaults to 1024×1024. Set them to match the latent you're generating into - if you're using a different aspect ratio, update both.

Output is one CONDITIONING, straight into the sampler. There's a commented-out aspect_ratio input in the source, but it's not exposed - you get the two dimensions.

Where it fits

Ships with ComfyUI core, and it's an old node - it's been in the built-ins since PixArt Alpha first got ComfyUI support. No model files to download beyond the PixArt Alpha checkpoint itself. The main gotcha is in the name, honestly: it does not apply to PixArt Sigma. Sigma works fine with the plain CLIP Text Encode node and doesn't want this resolution conditioning. If you've loaded a Sigma checkpoint and are confused why this node feels off, you're using the wrong encoder for the model.

Common issues

The classic mistake is leaving width/height at the 1024 defaults while generating into a differently-shaped latent - the conditioning and the canvas disagree and you get composition or aspect distortion. Set them to match. Also, because it's PixArt-specific, don't reuse this node as a general text encoder for other models; it stamps values other models will ignore or misread. And if the node errors on a mismatched CLIP, you've likely plugged in an SD/SDXL clip - it needs the PixArt Alpha text encoder.