SamplerCustom (Sampler Params)

Every ComfyUI user eventually hits the moment where a clean text-to-image graph dissolves into a spiderweb of five thin wires just to do one sampling pass. SamplerCustom (Sampler Params) is this pack's answer to that: it's a full SamplerCustom you drive from a single sampler_params bundle instead of separate sampler, scheduler, steps, seed, and cfg sockets.

The honest headline, though, is that this node is a thin, deliberate wrapper. Per the README (and true in the code), it does no custom sampling logic - it takes your bundle, feeds it through the official KSamplerSelect, BasicScheduler, and SamplerCustom nodes, and returns exactly what those would give you. That's a feature, not a cop-out. It means results are bit-identical to hand-wiring the core nodes, and it means there's no new sampling behavior to learn or trust.

Why you'd reach for it

The point is what the bundle buys you. Because the pack's Sampler Params node collapses six settings into one value, that bundle can be stored, swapped, and persisted. Wire a Sampler Params → this node and your graph reads as one unit instead of six dangling inputs. That becomes genuinely valuable when you're building the pack's intended use case: metadata round-tripping. Save an image with the pack's Image Saver, and the sampler settings ride along in the A1111 infotext; load it later with Image Reader and the whole sampling configuration comes back intact, ready to drive this node again.

There's also a tiled variant (SamplerCustom (Sampler Params, Tiled)) that reuses the same bundle and splits inference into spatial tiles - useful for VRAM-constrained rigs and for output resolutions beyond a model's practical size. Verified operation there is currently SDXL (with ControlNet Tile) and Anima; other models may work but aren't guaranteed, so treat the tiled version as model-dependent rather than universal.

Inputs and outputs

Five required inputs, all forced-as-sockets (you wire them, you don't type them):

• model - your loaded runtime model.
• positive / negative - the CLIP-encoded conditionings.
• sampler_params - the IPT-SamplerParams bundle. This is the whole trick: sampler, scheduler, steps, denoise, seed, and cfg all arrive through one wire.
• latent_image - the latent you're sampling from.

Outputs match the core node you're replacing: output (the denoised latent) and denoised_output (the result before final noise is added - usually what you VAE-decode). output is what you'll feed into a VAE decode in most workflows.

Install

Standard pack install, same as its siblings:

cd ComfyUI/custom_nodes
git clone https://github.com/kinorax/comfyui-info-prompt-toolkit.git
cd comfyui-info-prompt-toolkit
pip install -r requirements.txt

Or via ComfyUI Manager - search the pack title. Restart and you're done; this node has zero extra model or dependency needs beyond the pack's base install.

Where people get burned

Because it delegates to the core BasicScheduler, the usual sampler/scheduler advice applies verbatim - and it's not the same advice for every model family. If you're on SD 1.5 or SDXL, DPM++ 2M Karras remains the safe default. If you're on a flow-matching model like Anima or Flux, Karras is actively wrong - euler plus a conservative scheduler is the move, and Anima specifically expects the family's own scheduler (the pack's Anima workflows assume Beta57/RES4LYF-style schedulers, so a stock simple might not match what the model card wants). And don't expect the bundle to fight for you on VRAM: the standard (non-tiled) version samples at full resolution like any normal sampler - that's what the tiled sibling is for.