LoadCheckpointLite

Most checkpoint loaders are dead simple: pick a file, get a model. LoadCheckpointLite (shown as "LoadCheckpointLite" in the mittimiTools menu) is that, with one twist - its settings arrive as wires instead of widgets, so another node can drive it.

The three inputs - checkpoint_name, clip_num, vae_name - are exactly what LoadSetParametersLite outputs. In the pack's sample workflow the preset hub feeds this node directly: you pick a preset, and this loader obediently loads the checkpoint that preset names. Standalone it behaves like a slightly pickier CheckpointLoaderSimple, with the same three outputs (model, clip, vae) and the same underlying loading path - load_checkpoint_guess_config - under the hood.

Two inputs deserve a second look, because they're the reason this node exists.

clip_num is CLIP skip, as a negative number. The default of -1 means "use the second-to-last text-encoder layer"; -2 skips further back. SDXL finetunes commonly want a specific skip, and a preset can carry that along with the model - which is the whole point of a preset hub. The range is -24 to -1, so you're always skipping from the end, never from the start.

vae_name has a special "Use_merged_vae" option that isn't a real file. It's this pack's sentinel for "use the VAE baked into the checkpoint." Pick any other entry in the list and the node loads a separate .safetensors VAE from your vae folder instead. If you've been burned by mismatched VAEs before, the sentinel default is a small mercy.

The outputs are the standard trio - model, clip, vae - and they plug into the same places core loaders do: model into the sampler, clip into the text-encoder path, vae into VAEDecode.

Honest framing: this node is only worth installing for the preset workflow. If you're not using LoadSetParametersLite, the stock CheckpointLoaderSimple does the same job with one less moving part - no CLIP-skip widget you have to remember to set, no "Use_merged_vae" sentinel to explain to yourself later. But inside the pack it's what lets the preset hub stay pure data: the hub says what to load, this node loads it, and the graph stays readable. It's the cleanest split of "configuration" from "compute" in the whole pack, and it only makes sense in that context.

Install via ComfyUI Manager (search "ComfyUI_mittimiLoadPresetLite") or git clone https://github.com/mittimi/ComfyUI_mittimiLoadPresetLite into custom_nodes, then restart. No extra dependencies beyond the pack's single toml requirement, and no model files to download - it reads whatever checkpoints and VAEs you already have.