SDXL Settings (JPS)

This is the biggest single settings node in the pack - nine SDXL generation parameters bundled into one output, from resolution down to the filename prefix on your saves. If you've been dragging a KSampler's sampler/scheduler/steps/cfg widgets plus a separate resolution node plus a clip skip node around your graph, this is the "put it all in one place" version. Pair it with a matching Pipe node downstream to pull the individual values back out wherever they're actually consumed.

The inputs that matter

• resolution - a dropdown of ten options: "Use Image Resolution" plus nine fixed choices. Those nine aren't arbitrary - they're SDXL's own officially trained multi-aspect resolution buckets (1024×1024 at 1:1, then landscape and portrait pairs out to 1536×640 / 640×1536 at roughly 21:9). Generating off one of these ratios is the single most common cause of stretched or doubled anatomy on SDXL - stick to a bucket on this list and you're generating at a resolution the model actually saw during training.
• res_factor (1–8, default 4) - a multiplier the rest of your pipeline can use to compute an upscale target off the base resolution (the pack's README describes a matching standalone "Resolution Multiply" node for the same purpose) - default 4 lines up with a common 4x upscale pass.
• sampler_name / scheduler - the full stock ComfyUI sampler and scheduler lists (34 samplers, 9 schedulers). For SDXL photorealism specifically, the community's settled preference is a DPM++ 2M or 3M SDE sampler with the Karras scheduler; Euler ancestral is common but tends toward the over-smooth "SDXL skin" look people complain about, which is partly a sampler artifact rather than a model limitation.
• steps (20–250, default 60) - worth flagging: 60 is noticeably higher than the community-standard 20–35 step range for SDXL. That's not wrong, just tuned for slower, higher-quality final renders - drop it to 25–35 for faster iteration while you're still dialing in composition and prompt, and save the higher counts for the render you're keeping.
• cfg (1–20, default 6.5) - sits right in the community's standard 5–9 band for SDXL, a sensible default.
• cfg_rescale (0–1, default 0) - CFG rescale, which counters the washed-out, overexposed look high CFG values can produce by rescaling the noise prediction back toward the model's natural range. Leave at 0 unless you're pushing CFG high and seeing blown highlights, then try 0.5–0.7.
• clip_skip (-24 to -1, default -2) - SDXL's near-universal convention is clip skip -2, which is what most SDXL and SDXL-derived (Illustrious, Pony, etc.) checkpoints expect; check a checkpoint's own card before changing this, since some finetunes specify a different value.
• filename (default "JPS") - a string, presumably the prefix used by a downstream save node if you're using this settings pipe end to end through the author's own workflow templates.

Output: sdxl_settings, typed BASIC_PIPE - a JPS-specific bundle (a plain tuple, not Impact Pack's object of the same type label), meant to feed a matching SDXL Settings Pipe node to unpack these nine values into individual wires.

Installing it

ComfyUI Manager: search JPS Custom Nodes for ComfyUI. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/JPS-GER/ComfyUI_JPS-Nodes.git

Restart ComfyUI, deleting any earlier install of the pack first. No models - pure settings logic; the actual checkpoint, sampler, and VAE still load through their own dedicated nodes elsewhere in your graph.

Where people get tripped up

This node changes nothing on its own - it's a bundle, and every value in it is inert until you unpack it with the matching Pipe node and wire each output into the node that actually reads it (resolution into your latent, sampler/scheduler/steps/cfg into your KSampler, clip skip into your CLIP Set Last Layer node, and so on). If your generation doesn't reflect a change you made here, check that the corresponding Pipe output is actually connected downstream rather than assuming the settings menu applies itself.