Text Encoder 3 in 1

This one's built for a very specific and very common annoyance if you generate on Pony-family checkpoints: every prompt has to open with score_9, score_8_up, score_7_up, score_6_up... for the model to behave, and retyping (or re-copy-pasting) that boilerplate every time you tweak your actual subject prompt gets old fast. This node splits your positive prompt into two fields so the score-tag string can sit there permanently while you iterate on the character or scene in the other box, then concatenates the two into a single conditioning under the hood.

Why the score tags matter this much

Pony Diffusion's quality system is a genuine quirk of how it was trained, not just a style suggestion. AstraliteHeart's team hand-ranked roughly 20,000 images on a 1–5 quality scale and captioned the training set with the resulting score labels - the idea being that low-quality images could still teach concepts without teaching ugly. A training bug meant the model learned the whole concatenated string as the quality signal rather than the individual tags, so score_9, score_8_up, score_7_up (and onward) has to be written out in full to do anything. It's specific to the Pony family - it does nothing on Illustrious, SDXL base, or Flux, the same way generic quality tags like masterpiece, best quality do comparatively little on Pony.

Inputs and outputs that matter

• clip - the CLIP model to encode with.
• positive_Text_A - defaults to the full Pony score-tag opener plus some generic quality boilerplate (masterpiece, 4k, 8k, high quality...). Worth knowing: on Pony, the score tags are doing essentially all the real work here - the masterpiece/4k/8k filler is along for the ride and carries comparatively little weight on this model family. It's harmless, just not pulling its weight.
• positive_Text_B - your actual subject/scene description. This is the field you're meant to actually rewrite between generations, while A stays fixed.
• stop_at_clip_layer - CLIP skip, range −24 to −1, default −2. This matters more than it sounds: Pony V6, Illustrious and NoobAI all expect CLIP skip 2 (i.e. −2), and pushing it to −1 is a documented way to break Pony output into noise. Leave it at the default unless you specifically know your checkpoint wants otherwise.
• negative_Text - a single field, defaults to the standard negative boilerplate plus Pony's low-score negative tags (score_5, score_4, worst quality...).

Two outputs: Positive Conditioning (A and B concatenated and encoded) and Negative Conditioning, both CONDITIONING - wire them into your sampler exactly like any other text-encode node's output.

Installing it

ComfyUI Manager: search mzmaxam in Custom Nodes Manager and install. Or manually:

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

Restart ComfyUI. No models to download and no extra dependencies - it's a CLIP-encoding node built on types ComfyUI already has.

Where people get tripped up

The biggest trap is stop_at_clip_layer at −1 on a Pony-family model - that's the setting that turns clean output into garbage, so don't nudge it "for better quality" without checking your checkpoint's own recommendation first. Second, this node's defaults are Pony-specific by design: drop it into a workflow running Illustrious, base SDXL, or Flux and the score tags in positive_Text_A just sit there doing nothing useful - not harmful, just dead weight eating into your prompt's token budget. Third, and worth knowing if you're on anything newer: models with LLM-based text encoders (the current wave that replaced CLIP-and-T5 setups) don't have a CLIP layer to skip and don't understand score tags at all - this node's whole design assumes a classic CLIP-conditioned checkpoint, which is exactly the world Pony, Illustrious and NoobAI still live in.