Checkpoint Params Loader v074

Every model has its favorite settings. One wants euler_ancestral at 28 steps, another wants dpmpp_2m with a karras scheduler, and you always forget which is which the moment you load a new checkpoint. Checkpoint Params Loader v074 solves the remembering: point it at a checkpoint and it returns the sampler, scheduler, steps, cfg, clip_skip, resolution, and CLIP/VAE preferences that were saved for that model - ready to wire straight into your KSampler and loader.

It's part of the AAA Metadata System pack's checkpoint tester family, and it's the payoff node for the whole "set info, then use it" loop. The info setter records, this node recalls. If you've used CheckpointInfoSetter_v074 to give your models per-model defaults, this is how you actually collect on that work: swap the checkpoint path, and your sampler settings follow automatically.

How it works

The node hashes the checkpoint file and looks up its record in the pack's JSON database. It reads the stored generation params (sampler, scheduler, steps, cfg, clip_skip, width, height) and the stored CLIP/VAE preferences (preferred VAE, clip type, preferred CLIP 1–4, device, T5 bit depth), and returns them. If the file has no record, it falls back to sensible defaults rather than erroring - so wiring this node into a workflow before you've catalogued everything won't break it, it'll just give you generic values until you set real ones.

The practical pattern: put this node next to your checkpoint loader and KSampler. Wire sampler into the sampler field, steps into steps, cfg into cfg, clip_skip into your skip input, width/height into your latent sizing, and the preferred CLIP/VAE outputs into your loader's choices. Swap the checkpoint and everything reconfigures itself.

Inputs and outputs that matter

One required input: checkpoint_path - the absolute path to the model file.

Outputs, in the order you'll actually use them: sampler, scheduler, steps, cfg, clip_skip, width, height (the generation settings), then preferred_vae, clip_type_pref, preferred_clip1–clip4, clip_device_pref, t5_bit_depth_pref (the model-loading preferences).

Note there's no MODEL/CLIP/VAE tensor output here - this is a parameters loader, not a model loader. Pair it with a normal CheckpointLoader or the pack's CheckpointTester_v074 for the actual weights.

Installing it

Part of AAA Metadata System by Eric Hiss (GitHub: EricRollei). Via ComfyUI Manager (search "AAA Metadata System") or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/AAA_Metadata_System.git
cd AAA_Metadata_System
pip install -r requirements.txt

Restart ComfyUI. No extra dependencies.

Common issues

If you get default-looking values for a model you know you configured, the record isn't where the loader is looking - the same JSON DB must have been written by the setter in the same install; wipe or move the pack folder and the records go with it. If a preferred_clipN output is "None", that's stored data, not an error - the setter didn't have a choice saved. And the subtle one: this node loads what was stored, not what's correct - if you batch-tagged a mixed folder with the wrong architecture, the loader will faithfully hand back the wrong sampler preferences. Garbage in, faithfully remembered out.