PixArt T5 from SD3 CLIP

PixArt conditions on T5-XXL - the same class of large Google language model that's the whole reason Flux understands natural-language prompts as well as it does; it's an actual language model, not just an image-text alignment network the way CLIP is. The catch: T5-XXL is a multi-gigabyte download on its own, and if you already run Flux or SD3, you've almost certainly already downloaded it once as part of an SD3-style CLIP bundle. Downloading it again just for PixArt is a waste of bandwidth and disk. This node exists specifically to avoid that: point it at a CLIP you've already loaded for SD3, and it pulls the T5 component straight out for PixArt to use.

How it works

SD3's CLIP loader isn't a single encoder - it's a bundle of up to three (CLIP-L, CLIP-G, and T5-XXL) loaded together, commonly through a Triple or Quad CLIP Loader node. This node reaches into that already-loaded bundle, extracts the T5 half specifically, and repackages it as the T5 object PixArt's text-encode nodes expect - no separate download, no separate loader.

Inputs and output

• sd3_clip - a CLIP, from an SD3-style CLIP loader that was set up with the t5xxl file included. If your SD3 workflow skipped loading T5 to save VRAM (some do, using CLIP-L/G only), there's nothing here for this node to extract.
• padding - INT, default 1, range 1–300. Controls how many padding tokens get appended to the encoded sequence. PixArt's T5 conditioning was trained expecting a specific fixed-length padded input, so this isn't a free-form tuning knob - leave it at the default unless you're deliberately matching a known-good workflow's setting.

Output: t5 (typed as CLIP, matching what PixArt's text-encode nodes expect as input) - feed it in wherever you'd otherwise plug a T5v1.1 Loader's output.

Installing it

No extra model download beyond what your SD3/Flux setup already has. It's 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. Restart.

Common issues

The most common failure is the CLIP bundle simply not containing a T5 - if your SD3 loader was configured CLIP-L/G only (a common VRAM-saving move), this node has nothing to work with and will error out. Go back and check your SD3 CLIP loader actually has the t5xxl file wired in.

Padding mismatches usually don't crash anything outright - they just quietly degrade how well PixArt follows your prompt. If output looks noticeably worse than PixArt's native T5 Text Encode node with the same prompt, padding is the first thing worth adjusting before you blame the model or the prompt itself.