简易Checkpoint加载器 (Simple Checkpoint Loader)

Normally, loading a checkpoint with a non-default VAE and a specific CLIP skip value takes three separate nodes: Load Checkpoint, a VAE Loader, and a CLIP Set Last Layer node, all chained together. This node - the pack's own description calls it "lightweight and efficient" - folds all three into one, and throws in something none of ComfyUI's stock loaders give you: the model, VAE, and CLIP-skip values also come back out as plain strings and an int, ready to feed straight into logging or metadata nodes elsewhere in this same pack.

The inputs and outputs that matter

Required inputs:

• ckpt_name - the standard checkpoint dropdown, populated from your models folder exactly like ComfyUI's built-in Load Checkpoint.
• vae_name (default Baked VAE) - pick a specific VAE file, or leave it on the default to use whatever VAE is baked into the checkpoint itself.
• clip_skip (default -2, range -24 to -1) - how many layers early to stop the CLIP text encoder. ComfyUI counts this negatively (-1 = use the very last layer, -2 = stop one layer short), which is the opposite convention from A1111's 1-to-12 numbering - if you're translating a setting from an A1111-style model card, their "CLIP skip 2" is this node's -2.

Outputs - six of them, more than a stock loader gives you:

• MODEL, CLIP, VAE - the usual three, wire them exactly like any checkpoint loader's outputs.
• model_name (STRING) - its own tooltip says it can be used to save the model name via an Image Save node. Wire it straight into this pack's Save Image Plus checkpoint_name field, or use this pack's separate Model Name Extractor node if you'd rather derive it from the MODEL object elsewhere in the graph.
• vae_name (STRING) - shows whether you're running the baked-in VAE or a custom file.
• clip_skip (INT) - echoes the value you set, for anything downstream that wants it as data rather than just as this node's own setting.

Is -2 actually doing anything?

Worth knowing before you fiddle with it: the default of -2 is the standard convention for the Danbooru-lineage anime models this pack is built around - Illustrious, NoobAI, and Pony V6 all document CLIP skip 2 (ComfyUI's -2) as their expected setting, a convention that traces back to NovelAI's original penultimate-layer choice in 2022. But SDXL - which underlies all three of those model families - already conditions on the penultimate layer by default at the architecture level, so leaving this at -2 is mostly confirming the model's own default rather than changing anything. Where it actually matters is the failure case: setting it to -1 (or anything other than -2) reliably produces garbage on Pony V6 specifically, so the practical rule is less "-2 improves quality" and more "don't set it to -1." On a genuine SD 1.5 checkpoint, by contrast, the value does meaningfully change output, and -2 vs -1 there is a real stylistic choice rather than a landmine to avoid.

How to install it

ComfyUI Manager: search "ComfyUI-Danbooru-Gallery", install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Aaalice233/ComfyUI-Danbooru-Gallery.git
cd ComfyUI-Danbooru-Gallery
pip install -r requirements.txt

Restart ComfyUI. No extra downloads for this node itself - it loads whatever checkpoints and VAEs you already have on disk.

Common issues & troubleshooting

ckpt_name dropdown is empty. This is populated from your ComfyUI checkpoints folder exactly like the stock loader - if it's empty, no checkpoints are present where ComfyUI expects them, not a bug in this node specifically.

Output looks wrong / noisy on an SDXL-family model. Check clip_skip isn't sitting at -1 - that's the value that reliably breaks Pony-family models in particular, and doesn't help anything on Illustrious or NoobAI either. -2 is the safe default for this whole model family.

A LoRA trained for this checkpoint "doesn't work." Mismatched CLIP skip between training and inference is a standard entry on the LoRA-not-working checklist - if a LoRA card specifies CLIP skip 2 and you're running this node at anything else, that mismatch is worth ruling out before suspecting the LoRA itself.

Node doesn't appear in the menu. Standard pack-load check: confirm the repo path, confirm dependencies installed, check the ComfyUI console at startup for an import error.