KSampler Bridge to Text (SEB)

Let's get the misleading name out of the way: KSampler Bridge to Text (SEB) - the class is KSamplerControl - does not sample anything. It doesn't touch your denoising, it isn't a better KSampler, and you still need a real KSampler node downstream. What it does is sit at the front of your sampling block, capture the parameters you set, stamp a start time, and hand you a formatted text string describing the generation. It's the "bridge" between your sampler and a text overlay, and it's the first node in a three-node chain that burns your seed, steps, CFG, and sampler into the final image.

What it actually does

Every time it runs, KSamplerControl does three things:

• Records time.time() as start_time.
• Peeks at the latent you fed it, reads the tensor shape, and multiplies by 8 to work out the real image resolution (the VAE downsamples 8×, so the node corrects for it).
• Builds a multi-line string like Seed: 123\nSteps: 9 | CFG: 1.0\neuler | simple\nResolution: 1024x1024.

Then it passes everything through untouched. All your inputs flow straight out again, which is the point - you wire those outputs into the actual KSampler, and the string goes to the rest of the overlay chain.

That's a clean design detail: the sampler and scheduler dropdowns come directly from comfy.samplers.KSampler, so the values it emits are guaranteed to be valid for the KSampler you're feeding. No out-of-range enum, no "unknown sampler" validation errors.

The inputs that matter

• latent - from an Empty Latent Image or a previous sampler. Needed mainly so the node can report resolution.
• seed, steps, cfg - these are just integers/floats; set them here and they become your overlay text and your KSampler's inputs.
• sampler_name / scheduler - 44 sampler choices and 9 scheduler choices, identical to what the stock KSampler offers.

One thing to notice: the defaults are steps=9 and cfg=1.0. That's not a classic SDXL setup (which usually wants 20–30 steps, CFG 4–7) - it's distilled-model territory. The author clearly built this for the modern Turbo/Lightning/flow-matching era where 8 steps and CFG 1 are normal. If you're on a conventional SD 1.5 or SDXL checkpoint, override those defaults, or your overlay will lie about what a sensible generation looks like.

How to install it

This pack has zero Python dependencies of its own and downloads no models - the code only needs torch, PIL, and comfy.samplers, all of which ship with ComfyUI. So install is the easy kind:

cd ComfyUI/custom_nodes/
git clone https://github.com/Seb-Lis/ComfyUI_Node_Pack

Restart ComfyUI. Or, lazier and recommended: open ComfyUI Manager → Install Custom Nodes, search for "ComfyUI_Node_Pack" (the display name you'll see in the node list is "KSampler Bridge to Text (SEB)"), and click install.

Gotchas

• Don't skip the real KSampler. The node only records and passes through. Feed its latent, seed, etc. outputs into a normal KSampler or you get no image at all.
• start_time is when this node executes, not when sampling finishes. Because the KSampler depends on this node's outputs, that's effectively the moment before sampling - which is exactly what the pack's Generation Time node expects.
• No model name, no prompt. The overlay only contains what you feed it. If you want the checkpoint baked in too, you're wiring that yourself with a text node.