Text Preview + Token Count

The reason you want this node is the 75-token wall. An LLM writing your prompts will happily hand you a 300-token masterpiece - but if you're on SDXL or any CLIP-encoded model, the encoder splits your prompt into chunks and everything past token 75 lands in a new chunk, which quietly weakens the whole thing. This node is a text preview that counts, so the problem shows up on screen in red instead of as a mystery quality drop. It's the Text Preview + Token Count node, and it's exactly as useful as that sounds.

How it works

Feed it text (or type into the box) and it shows the text on the node plus a count. Four counting methods, picked with the tokenizer dropdown, plus auto:

• estimate - pure arithmetic, needs nothing installed, lands within ~15%. It takes the larger of characters/4 and words×1.33, so comma-heavy tag soup doesn't get undercounted - the direction that actually costs you.
• CLIP (connected) - connect the clip of your image model and you get the exact count the encoder will see. The padding and start/end markers are measured from the encoder itself by tokenizing an empty string, so it works on CLIP-L, CLIP-G, T5 and whatever ships next, no hard-coding per flavour.
• tiktoken - OpenAI's BPE, for budgeting (o200k for GPT-4o/GPT-5, cl100k for GPT-4/3.5). Requires pip install tiktoken.
• LLM server /tokenize - posts to a vLLM server's /tokenize endpoint for the exact count for the model that will actually read the prompt. LM Studio doesn't serve /tokenize, so on LM Studio it falls back to the estimate and says so.

auto picks the best of what's available: connected CLIP, then tiktoken if installed, then the estimate. Anything that fails just falls back to the estimate with a note in the report - it never hard-errors the graph.

token_limit is the budgeting knob. Set it to 75 (the CLIP chunk) and the report tells you how many tokens are left, or how far over you are - in red. When counting through CLIP, going past 75 also prints how many chunks the encoder will split the prompt into. That number is the whole reason to install this node: it makes a silent quality killer visible.

Outputs

• text - the text, passed straight through. Wire this downstream if you want the preview node sitting in your prompt path without touching the text.
• tokens - the count as an INT, which you can feed into logic or a display node.
• report - the full multi-line summary: count, method, chars/words/lines, budget status, chunk warnings.

Install

Part of the comfyui-llm-prompt-studio pack. ComfyUI Manager: search comfyui-llm-prompt-studio, or:

cd ComfyUI/custom_nodes
git clone https://github.com/grandju51/comfyui-llm-prompt-studio

Restart and refresh. Nothing else to install for the estimate or CLIP modes; pip install tiktoken only if you want the OpenAI counts. ComfyUI Manager will offer to handle that dependency automatically.

Notes

The node deliberately has no IS_CHANGED hook. Forcing a re-run on every queue would invalidate everything wired to the passthrough text output and re-sample the whole graph just to redraw a number - a bad trade. Unchanged inputs mean an unchanged count anyway, so it only refreshes when its inputs actually change. Preview nodes like this are the cheap insurance of a ComfyUI graph: they don't touch pixels, they just stop you from shipping a prompt you'd regret.