Deprecated · Sampling Trace CLIP · Connect Both Prompts
The CLIP Snoop That Records What Your Prompt Actually Said
- clip
- ② 긍정·부정 Text Encode
- ③ CLIP 프롬프트 추적 보내기
The name is a slight lie - Sampling Trace CLIP doesn't trace anything by itself. It's a pass-through CLIP proxy that listens: every tokenize() call that flows through it gets recorded, so the Sampling Trace Inspector panel can show you real per-word attention instead of guesses. It's also deprecated in the current 0.4.0 release and hidden from normal node search, because the one-node setup now does this automatically. You only meet this node if you're reading an older saved workflow or splitting the trace manually.
What it actually does
You connect your checkpoint CLIP in, it hands the same CLIP straight back out - unchanged, pixel-for-pixel, token-for-token - plus a second output of type PROMPT_TRACE. That second wire is the payload: it carries the real prompt text, CLIP-L/G token IDs, input weights, and readable word groups from the actual Text Encode calls that went through the proxy. The README's key warning, in its own words: connect prompt_trace to the Sampling Trace Model node's prompt_trace input - not to the Inspector panel, and not to a trace_session wire.
That distinction is the whole trap, and it's a classic case of the node-plumbing layer's two-different-wires confusion. PROMPT_TRACE is a data payload meant for the model node that creates the run. TRACE_SESSION is the run itself. Mix them up and your prompt evidence silently goes nowhere. The 0.4.0 one-node setup removed the need to care by binding both in one node.
Inputs and outputs
- clip (required, CLIP) - checkpoint CLIP, fanned out to both positive and negative Text Encode nodes so the proxy sees every call.
- Outputs: CLIP (the unchanged pass-through) and PROMPT_TRACE.
That's it. Two wires, one input. There's no label, no mode, no session to attach.
Install
Same pack, same story:
git clone https://github.com/ILcrowe/ComfyUI-SamplingTrace-Inspector.git ComfyUI/custom_nodes/ComfyUI-SamplingTrace-Inspector
or ComfyUI Manager → search "Sampling Trace". Restart, and remember: in current versions you won't find this node in search - it loads only from saved workflows that still reference it.
Where people get burned
- "Why can't I find this node?" It's deprecated and hidden from normal search by design. Saved workflows still load it; new graphs should use Sampling Trace · One Node Setup.
- Wiring
prompt_traceinto the panel ortrace_session. The README calls this out explicitly - it goes to theprompt_traceinput on Sampling Trace Model. The one-node setup spares you this footgun entirely. - Only one Text Encode path connected. Route the CLIP through the proxy to both encoders, or the panel shows a connection guide instead of word measurements rather than assigning attention to the wrong words.
Honest take: this node is a perfectly good piece of plumbing, but it exists because the pack used to be split across several nodes. The one-node setup is strictly less work for the same result. Treat this one as "how the machine works under the hood" and move on.
Inputs (1)
| Name | Type | Default | Description |
|---|---|---|---|
| clip | CLIP | — |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| ② 긍정·부정 Text Encode | CLIP | — |
| ③ CLIP 프롬프트 추적 보내기 | PROMPT_TRACE | — |