Checkpoint Automatic Config

The most reliable way to ruin a Lightning or Turbo checkpoint is to feed it your normal settings. CFG 7, 20 steps, DPM++ 2M Karras - that's how you get the oversaturated, artifacted mess that ends up in half the "why does my output look burnt" threads. This node is a small fix for exactly that failure mode: it's the standard CheckpointLoaderSimple with a table of "right settings" bolted on, so the loader and the sampler recipe travel as one unit.

What it does

Pick a checkpoint and it loads it exactly like the built-in loader - same MODEL, CLIP, VAE outputs - but it also hands you STEPS, CFG, SAMPLER, and SCHEDULER on separate ports. Wire those straight into a KSampler and you never set steps, CFG, sampler, and scheduler by hand again. When automatic_config is on (it is by default), the values don't come from the widget you typed into - they come from a YAML file the pack ships, keyed by checkpoint filename.

How it works

The node is a subclass of ComfyUI's built-in CheckpointLoaderSimple, so it inherits the loader and just adds plumbing around it. At startup it reads models_config.yaml from its own folder. When you run, it looks your ckpt_name up in that table; if found, it validates the entry and overrides whatever you typed in steps_total, cfg, sampler_name, and scheduler_name, printing the applied values to the console. If the checkpoint isn't in the table and auto-config is on, it refuses to run with an "unknown checkpoint" error.

The shipped table covers 13 SDXL checkpoints, most of them Lightning or distilled merges: Juggernaut XL Lightning, DreamShaper XL Lightning, RealVisXL v4 Lightning, TurboVisionXL, and friends. They're set to 5-8 steps, CFG 1-2, on dpmpp_sde or euler/euler_ancestral with karras or sgm_uniform - the low-CFG, few-steps playbook that distilled models require. The specific sampler pairs are the author's reading of each model card, not universal law; the sane rule is still to take the pair your model card recommends first.

The inputs and outputs that matter

• ckpt_name - the checkpoint to load.
• automatic_config - the whole feature. Flip it off and the node degrades to a plain loader that uses your typed values.
• steps_total, cfg, sampler_name, scheduler_name - your fallbacks, and what auto-config overrides.

Outputs: MODEL, CLIP, VAE (identical to stock), plus STEPS, CFG, SAMPLER, SCHEDULER. The last four are what you connect to a KSampler's steps, cfg, sampler_name, and scheduler inputs.

Installing it

It's published to the ComfyUI registry, so the easy route is ComfyUI Manager → search "Checkpoint Automatic Config" and install. Or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/mech-tools/comfyui-checkpoint-automatic-config

Restart ComfyUI after either. No pip dependencies beyond PyYAML (ComfyUI already has it) and it downloads no models - it only references checkpoints you're expected to have in models/checkpoints already.

Common issues

The "unknown checkpoint" error is the one you'll hit first: auto-config is on and your model isn't in the table. Fix it by editing models_config.yaml (in custom_nodes/comfyui-checkpoint-automatic-config/) or by flipping automatic_config off. Two footguns when you edit that file by hand: cfg must be written with a decimal point - cfg: 2.0, not cfg: 2, or validation throws "invalid cfg format" - and steps_total must be a whole number. The file is read at import, so restart ComfyUI after editing. And yes, the config silently overrides your widgets: type 30 steps, get the table's 6. That's the point, not a bug.

One honest warning: this is a small, GPL-licensed pack, last touched in September 2025, with a README that's one sentence ("noded" typo included). If you live on one or two Lightning SDXL models, it genuinely removes a recurring footgun. If you use none of the 13 listed checkpoints, you're really just adopting a YAML template - which is fine, but know that going in.