Prompt Token Counter 🔢

The silent killer of long prompts is truncation. CLIP encoders on the SD1.5/SDXL lineage cap your text at 77 tokens (75 usable plus the special tokens), and everything past that boundary is silently dropped - you've been writing prompts that get chopped mid-sentence and wondering why the tail never shows up in the image. Prompt Token Counter 🔢 counts your prompt's tokens and tells you, before you spend GPU time, whether it's going to be truncated.

It's a diagnostic node with a bit of meat on it: it uses the real transformers CLIP tokenizer when available (exact count) and falls back to a whitespace/punctuation heuristic when not (roughly right for typical English, enough to flag truncation risk).

How it works

Token counting on CLIP is a BPE wordpiece operation, which is why a naive character count lies to you. This node hands the prompt to CLIPTokenizer if transformers is installed - and it usually is, since ComfyUI bundles it - and strips the BOS/EOS the tokenizer adds, giving you the usable token count. Then it compares against a budget of 75 tokens per chunk, multiplied by chunk_size.

The chunk_size input matters more than beginners expect: ComfyUI's default CLIP Text Encode concatenates multiple 75-token chunks automatically, so a chunk_size of 2 means a 150-token budget. Set it to match the encoder your workflow actually uses, or you'll get false "will truncate" alarms on legitimately long prompts.

The inputs and outputs

Inputs: prompt and chunk_size (1–10, default 1). Outputs: token_count (INT), will_truncate (BOOLEAN), truncated_preview (STRING - the prompt as it would survive truncation), and summary. truncated_preview is the genuinely useful one: it shows you the exact cut point, so you can see what's dying and fix the prompt instead of guessing.

Wire will_truncate into a conditional gate if you want a workflow that refuses to queue over-budget prompts - the honest version of a token guard.

Installing it

Part of OmniNodes:

cd ComfyUI/custom_nodes/
git clone https://github.com/TensorVizion/OmniNodes

Restart ComfyUI (or search "OmniNodes" in ComfyUI Manager). No extra pip installs - it uses transformers if ComfyUI already has it and degrades gracefully otherwise. Registers under TensorVizion/Prompt. If it doesn't appear, restart completely and check the terminal for [OmniNodes] ✅ Loaded lines.

Two honest caveats. First, this is about CLIP-style encoders - the LLM-encoded 2026 models have no 77-token boundary at all, and this node's budget math doesn't apply to them. Second, an exact count still depends on the tokenizer being available; the heuristic fallback is close but not exact. For the SDXL-lineage world where truncation is a real daily problem, it's exactly what you want between the prompt builder and the encoder.