USK Smart Checkpoint Loader

The hardest part of switching anime checkpoints isn't downloading them - it's remembering which one needs what. Pony wants its score_9, score_8_up incantation, Illustrious and Animagine want Danbooru quality tags, and the resolution you should generate at changes between SD1.5 (512-ish) and SDXL (1024-ish). USK Smart Checkpoint Loader is a one-node answer to all three problems: it reads the checkpoint file itself, works out what family it belongs to, and hands you the model, the right latent, and the right prompt prefix in one go.

The "smart" part is worth being precise about. There's no API call and no key - the name means heuristic, not AI. The node opens the safetensors file in CPU mode and inspects the header keys directly. Spot conditioner.embedders.* keys and it's SDXL; spot first_stage_model.* keys and the VAE is baked in; then it checks the filename and the file's metadata for pony, illustrious, animagine, or noobai to pin the lineage, with a vocabulary-size check (the token embedding beyond 49408) to catch extended-vocab Illustrious-style models. Then it builds an empty latent at the right resolution and calls ComfyUI's own load_checkpoint_guess_config to do the actual loading. Nothing exotic under the hood - which is why it has zero extra dependencies and runs anywhere ComfyUI does.

Three inputs, and only two you'll really touch. ckpt_name lists your checkpoints folder, aspect_ratio gives you the five usuals (1:1, 3:4, 4:3, 9:16, 16:9), and prompt_profile has the interesting choice: Auto-Detect, or force a family manually (Pony, Illustrious, Animagine, NoobAI/NAI, Real/Photo, or None). Auto-Detect is the point of the node - pick it and stop thinking about prefixes.

The seven outputs wire up like a normal loader plus extras: MODEL, CLIP, and VAE go to your KSampler chain exactly as from any checkpoint loader. LATENT replaces the Empty Latent node - its size is already chosen for the detected architecture. BASE_PROMPT is a STRING holding the quality prefix, meant to be concatenated onto your real prompt. MODEL_TYPE tells you what it detected, and INFO_TEXT is a full readout you can pipe into a ShowText node to see the verdict and any VAE warning.

Where people get burned: the VAE warning fires constantly on SD1.5 checkpoints, because most of those shipped without a baked VAE - that's not a bug, you just need a VAE Loader on the side, and the INFO_TEXT output tells you so. The bigger gotcha is detection being filename-and-metadata heuristics. Rename a Pony file to my_cool_model.safetensors and Auto-Detect falls back to "standard structure," the Pony prefix silently goes empty, and you're generating mediocre anime wondering why. That's exactly when you set prompt_profile manually. Also worth knowing: the default Pony prefix appends source_anime, - opinionated, and easy to miss if your target is furry or western-cartoon Pony derivatives.

Installing is trivial since the pack ships no dependencies:

cd ComfyUI/custom_nodes
git clone https://github.com/dokukobura/comfyui-usk-utils.git

Restart ComfyUI and search "USK Smart Checkpoint Loader" on the canvas. ComfyUI Manager finds it as comfyui-usk-utils if you'd rather click. One caveat from the wider ecosystem: any custom node installs via git clone with no review process, so don't auto-trust packs you found randomly - this one is small, by a solo author (Doq), and its whole job is reading headers and loading models, but the habit is cheap.

The one genuinely fiddly thing is customization: the aspect ratios and prompt prefixes live in aspect_ratios.json and prompt_profiles.json inside the node folder, so changing the base tag list means editing JSON and restarting rather than tweaking in the UI. For most people the defaults are fine. If you live in the anime checkpoint shuffle and keep forgetting the score tags, this node quietly eliminates a whole class of dumb mistakes.