pipeKSampler

pipeKSampler is where a tinyterra workflow actually does the generating. It takes the pipe from pipeLoader, runs the sampler, and gives you back the results plus an updated pipe - and it folds in three extra jobs you'd normally need separate nodes for: an optional LoRA apply, an optional upscale before sampling, and the image preview/save. It's the KSampler that thinks it's a Swiss Army knife, and mostly it's right.

The design goal, borrowed from the Efficiency Nodes lineage: keep the graph small enough to read. Instead of Load Checkpoint → LoRA → Text Encode ×2 → Empty Latent → KSampler → Preview, you get pipeLoader → pipeKSampler → done. The README even shows the old multi-node layout next to the pipe version to make the point.

How it works

Feed it a PIPE_LINE (from pipeLoader_v2 or the legacy loader). The sampler pulls model, conditioning, latent, VAE, CLIP and seed out of the pipe, but every one of those has an optional_* input to override it - that's how you hook in a ControlNet-applied conditioning or a custom latent without unwrapping the pipe. The seed input on the node overrides the pipe's seed if you provide it.

Three built-ins you'll actually use:

• lora_name / lora_strength - one quick LoRA apply right before sampling, without touching the loader.
• upscale_method - the hires-fix block. Choose from [latent] (latent-space upscale: nearest-exact, bilinear, area…) or [hiresFix] (pixel upscale via an upscale_model_name), then set factor, or rescale to a percentage / width-height / longer side. In practice: latent upscale for a cheap second pass, hiresFix with a real upscaler model for quality.
• image_output / save_prefix / file_type / embed_workflow - preview or save right here, with the same prefix-parsing as the standalone imageOutput node.

Sampler controls are the standard steps, cfg, sampler_name (44 samplers), scheduler (including the AYS presets the pack adds), and denoise.

Inputs and outputs

The inputs that matter day-to-day: pipe, steps, cfg, sampler_name, scheduler, denoise, and the upscale trio. Outputs: the updated pipe, plus model, positive, negative, latent, vae, clip, images, seed, and plot_image (the last one exists so an advanced xyPlot can inject a grid image into the flow).

Installing it

Part of tinyterraNodes:

cd ComfyUI/custom_nodes
git clone https://github.com/TinyTerra/ComfyUI_tinyterraNodes.git

or ComfyUI Manager → tinyterraNodes. No extra dependencies.

Gotchas

The [hiresFix] upscale path needs an upscale model in ComfyUI/models/upscale_models - pick None there and the hiresFix options silently can't work, so set upscale_model_name first. And because image_output here saves every run when set to Save, iterate in Preview and flip to Hide/Save when you're done. One more: the optional overrides are a real power feature but also the classic "why is my result ignoring the loader?" culprit - if a wired optional_model or optional_latent is connected, it wins over the pipe.