Load Checkpoint (Umbra Lab)

Five loader nodes that all look the same but each know one model format - that's the stock ComfyUI situation, and it's exactly the friction this node removes. Load Checkpoint (Umbra Lab) is a unified loader with a model_type dropdown that switches between checkpoint, diffusers, diffusion_model, unet, and gguf. Pick the type, pick the file, and it returns model, clip, vae, plus a model_name string for metadata. One node in your graph instead of "which of the five did I use here."

How it works

The five modes map onto the loaders you already know:

• checkpoint - the classic merged SD1.5/SDXL/etc. .safetensors via load_checkpoint_guess_config, returning full model + CLIP + VAE.
• diffusers - a diffusers directory (one with a model_index.json), discovered automatically from your models/diffusers folder. Same full model+CLIP+VAE payload.
• diffusion_model and unet - bare diffusion weights (the flux1-dev.safetensors style single file in models/diffusion_models, or a UNet in models/unet). These come back with CLIP and VAE as None - you have to provide those from elsewhere, which is the standard Flux-style split setup.
• gguf - quantized .gguf diffusion models. This is the Flux-on-12GB path from the KB's GGUF panel (Q8 is basically fp16 at half the size). Important caveat: it requires ComfyUI-GGUF installed, because the node delegates to GGUF's UnetLoaderGGUF. Without that pack, the dropdown may be empty and GGUF mode raises a clear error telling you so.

There's also a weight_dtype selector with fp8_e4m3fn, fp8_e4m3fn_fast, and fp8_e5m2 options for the diffusion-model/UNet paths - a quick way to get fp8 loading without a separate fp8 wrapper.

What you actually set

• model_type - the format you're loading. This is the one that matters; everything else is keyed off it.
• Then exactly one of the four name fields - checkpoint_name, diffusers_model, diffusion_model_name, unet_name, gguf_name - fills in from your installed models. The others sit ignored.
• weight_dtype - leave default until you know you want fp8.

Installing it

Ships in Umbra-Nodes, the ComfyUI companion pack for Umbra Studio. ComfyUI Manager → search "Umbra Nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nocturne-Ai-Labs/Umbra-Nodes

Restart ComfyUI. No pip deps of its own. GGUF mode needs the separate ComfyUI-GGUF pack; everything else works against stock ComfyUI and whatever model files you already have in the standard folders.

Gotchas

The None CLIP/VAE on diffusion_model and unet modes trips people constantly - the node returns them as None, not error, so a downstream CLIP Text Encode fails confusingly until you wire in a separate CLIPLoader/VAELoader. That's normal for Flux-style stacks, not a bug. Second: gguf_name's dropdown only shows files ComfyUI-GGUF can see, so if your .gguf isn't appearing, the pack isn't installed or the file isn't in a diffusion_models/unet folder. And one nice behavior worth knowing: leave a name field empty and it quietly picks the first file it finds in the folder, which is a reasonable "just load something" default for testing.