pipeKSamplerAdvanced

pipeKSamplerAdvanced_v2 (ttN pipeKSamplerAdvanced_v2, displayed as pipeKSamplerAdvanced) is the workhorse of the tinyterra pipe system: the one node that samples, upscales, saves, and even writes its own prompts, all inside a ttN pipe. It's the v2 of the legacy advanced sampler, and the upgrade list is exactly what you'd want - embedded Advanced CLIP text encoding, a proper hires-fix path, and full XY-plot support.

If you're doing multi-stage SD1.5/SDXL work with tinyterra, this is the node you'll reach for most. It's the pipe counterpart of ComfyUI's KSamplerAdvanced, but instead of raw wires it eats and emits a PIPE_LINE, which keeps the graph small when you're chaining two or three passes.

How it works

The node pulls model, conditioning, latent and VAE from the incoming pipe (overridable via optional_model, optional_positive, optional_negative, optional_latent, optional_vae, optional_clip), applies an optional LoRA, runs the step window you specify, decodes through the VAE, and pushes the result through preview/save and back into a fresh pipe.

The two features that separate it from the v1:

• Embedded prompt boxes. positive and negative text fields with token_normalization and weight_interpretation dropdowns (comfy / A1111 / compel / comfy++ / down_weight) - the vendored ADV_CLIP_emb encoding, so (keyword:1.3) syntax works if you flip interpretation to A1111.
• A real hires path. upscale_method offers [latent] modes and [hiresFix] modes (the latter with an upscale_model_name), plus a rescale block to push the decoded result to a target size. The classic two-stage recipe - sample at low res, [latent]-upscale with a factor, then run a second pass over just the top of the schedule with add_noise disabled - all lives in one node.

The inputs that matter

• steps, start_at_step, end_at_step - the sampling window. First pass 0→20, then a hires continuation 16→20 with add_noise off.
• add_noise + noise - noise is a 0–1 float; drop it below 1 for a partially-denoised continuation.
• return_with_leftover_noise - enable when a later pass continues from this latent.
• lora_name + lora_strength - single strength here (the v1 split model/clip strengths into two).
• cfg, sampler_name, scheduler (includes AYS SD1/SDXL/SVD presets), denoise-adjacent controls.
• image_output, save_prefix, file_type, embed_workflow - preview/save with full format control and the pack's %date%/%widget% prefix parsing.

Outputs: pipe, model, positive, negative, latent, vae, clip, images, seed, and plot_image (activated when you connect an adv_xyPlot - the pack's grid engine, which the community leans on heavily for LoRA weight testing). There's also input_image_override for img2img-style runs.

Install and gotchas

One-time pack install - ComfyUI Manager → search ComfyUI_tinyterraNodes, or clone into ComfyUI/custom_nodes, restart. No deps, no model downloads.

Gotchas: the default positive/negative boxes contain the literal strings "Positive"/"Negative" - clear them or you'll bake the word into every render. And remember this is an SD1.5/SDXL-era node; the [latent]/[hiresFix] split doesn't map cleanly onto the newest architectures, the pack's known weak spot. For old v1 workflows, the node will version-flag on load - that's by design, and migrating just means moving your prompts into the node.