Clip Text Encode SD3 (Shinsplat)

If you're running an SD3-family model and just want the Shinsplat BREAK/END behavior without the extra diagnostic outputs, this is the version to grab. It's the SD3 sibling of the T5 encoder - same three text areas (clip_l, clip_g, t5xxl), same empty_padding enum, same prompt_before/prompt_after inputs - but it only returns CONDITIONING plus a single prompt_out string. Clean, and about as close to a drop-in replacement for stock CLIP Text Encode (SD3) as this pack makes.

How it works

Mechanically it's the T5 encoder with the debugging surface removed: each text stream gets the END chop applied independently, BREAK segments the context blocks, and the three streams are tokenized and encoded into one combined conditioning. prompt_out gives you the assembled, directive-free text the encoder actually saw, which is handy for mirroring the prompt into a second stage or a save node. The author's own README is lukewarm about the whole SD3 situation - his changes notes say "SD3 is probably not viable and I won't need to do anything more than testing" - so temper expectations accordingly; this node is more useful as a clean SD3 encoder than as a magic fix.

The inputs that matter

• clip - CLIP model from an SD3-style checkpoint.
• clip_l, clip_g, t5xxl - the three text areas; t5xxl carries most of the meaning on SD3-family models.
• empty_padding - none or empty_prompt, controlling padding behavior on the empty streams.
• prompt_before / prompt_after - optional prepend/append text, wired the same way as the other pack encoders (so a LoRA loader's prompt_out drops in cleanly).

How to install it

Same pack, same one-time install:

cd ComfyUI/custom_nodes
git clone https://github.com/Shinsplat/ComfyUI-Shinsplat

or via ComfyUI Manager, searching "ComfyUI-Shinsplat", then restart. No extra dependencies; the pack carries its own tokenizer files.

Common issues

• Only one prompt output - if you need tokens_raw/tokens_out for inspection or token reweighting, you want Clip Text Encode T5 (Shinsplat) instead. This node is deliberately minimal.
• Model is actually a Flux-style clip - Flux wants the separate CLIP Text Encode Flux (Shinsplat) node, which has a dedicated guidance input. Using the SD3 variant on Flux won't give you guidance handling.
• Same-seed images differ from stock - pack-wide behavior; the tokenizer restructures tokens before encoding. Expected, not a bug.