Sampling Trace · One Node Setup

You've hit the classic wall: the image looks wrong, you don't know if it's CFG, the LoRA, the ControlNet, or just your prompt, so you change a setting and pray. This node is the part of Sampling Trace Inspector that breaks that loop. Sampling Trace · One Node Setup is the only node in the pack that still shows up in normal node search - the author deprecated the other nine - and it's the one you actually want.

It records what happens inside your sampler, step by step, without you swapping KSampler for some custom replacement. After a run, the bottom Sampling Trace Inspector panel gives you a draggable denoise slider with an approximate x0 preview, sigma, CFG delta, ControlNet residual summaries, and per-word prompt attention. The hook: find the step where the image first breaks, see which signal moved, and change only the setting that matters. It's a debugger, not a claim that "this word caused 30% of the image" - the author is refreshingly explicit about that.

How it works

Internally this node is a one-stop wrapper around the pack's two deprecated workhorses. It clones the MODEL you feed it, installs sampling hooks on the clone, and wraps your checkpoint CLIP in a non-mutating proxy that records the actual tokenize() calls your Text Encode nodes make. It returns float("nan") from its change check so a fresh trace session is minted every single queue run - the always-rerun trick documented in the ComfyUI node-plumbing layer. Your existing KSampler stays exactly where it is; any sampler that follows ComfyUI's standard CFGGuider → ModelPatcher path gets traced for free.

Wiring that matters

Only three inputs exist, and two of them are the whole point:

• final_model (MODEL) - feed it the last MODEL in your chain, after every LoRA and IPAdapter patch. Snapshot before the patches and you'll get an incomplete picture.
• checkpoint_clip (CLIP) - the raw CLIP from your checkpoint loader. This is what makes prompt-word attention possible.
• trace_preset (enum, default basic) - Basic is lightweight flow + previews. Advanced adds the expensive stuff: CFG delta, Control residuals, and step-level word attention.

The node's MODEL output goes to your sampler. Its CLIP output fans out to both the positive and negative Text Encode nodes. There's also a TRACE_SESSION output - leave it alone unless you're wiring deprecated nodes into the same run.

Install

This is a fresh public preview (0.4.0-beta.4) from ILcrowe, GPL-3.0, with English and Korean docs. It has no third-party runtime dependencies beyond ComfyUI's own Python environment, and the previews use ComfyUI's built-in TAESD / Latent2RGB machinery - no model downloads. Install via ComfyUI Manager (search "Sampling Trace") or:

git clone https://github.com/ILcrowe/ComfyUI-SamplingTrace-Inspector.git ComfyUI/custom_nodes/ComfyUI-SamplingTrace-Inspector

Then restart ComfyUI and search for Sampling Trace · One Node Setup.

Where people get burned

• Word attention shows a "connection guide" instead of words. You didn't route the checkpoint CLIP through this node. If only one Text Encode path came through, it won't guess.
• It sits in the wrong place. After the last MODEL patch, not before it - otherwise runtime patch keys don't match the sampling call.
• Batch previews look half-dead. A four-image batch captures per-item previews, but runs from before batch capture show only the first image with the other 1/4 buttons disabled until you re-run.
• Flux or Qwen? The CLIP-tokenizer path is what's validated (SDXL, Illustrious, ControlNet, IPAdapter, Impact Detailer all verified). Video models and Flux are listed as still-unverified - don't blame this node when they don't trace.

One last thing worth knowing: runs land in ComfyUI/user/trace_inspector/runs/ and can contain your prompt text, model filenames, and previews. The author strips local paths, but review a run folder before you share it.