Base (full-pipe)

Every mudknight-utils workflow funnels toward this node. Base (full-pipe) is the sampler: a full_pipe with model, conditioning, and (optionally) an image or latent goes in, and a generated IMAGE comes out. It's the pack's replacement for the KSampler / EmptyLatentImage / VAE decode stack, collapsed into one node with sensible resolution presets.

How it decides what to generate: if you feed it an image, it encodes it to latent and runs img2img at whatever denoise you set. If the pipe already carries a latent, it samples that (which is how the pack's Custom Resolution and inpaint nodes hand it a starting point). If neither is present, it creates an empty latent at the selected resolution and forces denoise to 1.0 - a clean text-to-image. Same node, all three jobs, decided by what's in the pipe. It's marked as an output node, so it shows its previews like a save/preview node and returns both the updated full_pipe and the decoded image.

The inputs that matter

• sampler_name / scheduler / steps / cfg - the usual sampler panel, defaulting to euler_ancestral_cfg_pp and karras. These defaults are the author's taste and a fine starting point for SDXL-class models.
• resolution - seven presets, all multiples of 8 near 1MP (1024×1024 down to 640×1536). This matters: SDXL-class models were trained around 1MP, and the presets exist to keep you there.
• portrait - flips the chosen resolution to landscape.
• scale - a multiplier on the latent's width and height. This is the pack's answer to "I want a bigger image": go above 1.0 and you're past training territory, so pair it with a high-res fix or detail pass rather than expecting a clean 2MP render.
• variations - generate N images in one run, each using seed + index. Batch generation without wiring a seed source.
• Optional image - switches the node into img2img/inpaint mode.

Installing it

In mudknight-utils - ComfyUI Manager (search comfyui-mudknight-utils) or:

cd ComfyUI/custom_nodes
git clone https://github.com/mudknight/comfyui-mudknight-utils

No Impact Pack, no extra model files. It uses the pack's align_your_steps scheduler support and beta scheduler, both built in.

Where people get burned

The classic Base Node surprise is img2img-by-accident: wire an image into the optional input and forget the denoise, and you've quietly turned t2i into i2i. Set denoise to 1.0 explicitly when you want the image ignored... actually, you can't ignore it that way - disconnect the image wire. The other trap is scale past 1.0 without a plan; the community's well-worn "1MP curse" applies, and the pack's own docs recommend going past 1.5× only with a high-res LoRA. Keep the presets for clean t2i, use scale deliberately, and let the detailers fix what the resolution can't.