Smart Model Loader

"Smart" is doing heavy lifting in that name, but for once it's earned. This is the flagship node of the ComfyUI Smart Model Loader pack: one loader that grabs your checkpoint or diffusion model, your CLIP, your VAE, your latent, your sampler settings, your seed, and your LoRAs, and hands the whole lot downstream over a single PIPE wire. If you've ever built a workflow and watched it become forty crossing noodles between the loader, empty latent, and sampler, you know exactly why that exists.

The trick is the combo-chip bar at the top. features is a comma-separated list of toggles - clip, vae, audio_vae, latent, sampler, lora, model_sampling, block_swap, memory_cleanup, integrity, seed, templates - and each chip you click reveals (or hides) that whole section of the node. Disabled sections don't just look tidy, they're not even put into the pipe. That's the design philosophy: start minimal with clip, vae, memory_cleanup, and grow the loader as the workflow needs it. The selected chip state is saved with the workflow, so a shared graph loads looking exactly like its author left it.

Pick a model_type and the loader adapts. Standard Checkpoint reads a full .safetensors with baked CLIP and VAE. UNet Model is diffusion-only and expects external CLIP and VAE. Nunchaku and GGUF are the quantized paths - FP4 for NVIDIA via SVDQuant, or llama.cpp-style GGUF - and both need their own ComfyUI extensions installed before the loader will even list those options. Each type surfaces only the widgets that matter to it: weight_dtype for UNets, data_type/attention/i2f_mode for Nunchaku, gguf_dequant_dtype for GGUF. GGUF is worth it if VRAM is the wall you keep hitting; Q8 is basically fp16 at half the size, and if it fits, just use it.

Two features genuinely set this loader apart. Templates (the templates chip) save your whole configuration - model file, chips, sampler, LoRA slots - to ComfyUI/custom_nodes/ComfyUI_SmartModelLoader/templates/ and restore it with one load. If a template references a model you don't have, the loader keeps the filename selected with a (missing) marker, restores the saved CivitAI AIR identity, and reveals a Download from CivitAI button - it never downloads on its own. Integrity (integrity chip) is where you paste a CivitAI AIR URN or SHA-256 into air_or_hash; the loader hashes the file, warns on mismatch, and can fetch a missing model into the folder you pick with download_target_role.

Output is one pipe that carries model, clip, vae, latent, dimensions, sampler settings, seed, model_name, and more. Feed it to the pack's IO Checkpoint Loader to unpack individual sockets, to Eclipse KSampler (Pipe) to sample and decode, or to optional ComfyUI Eclipse nodes (Context Image, Generation Data, Concat Pipe Multi, Smart Sampler Settings) that consume the same pipe contract.

Installing

ComfyUI Manager - search ComfyUI Smart Model Loader - or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_SmartModelLoader.git
cd ComfyUI_SmartModelLoader
python -m pip install -r requirements.txt

Restart ComfyUI. The [Eclipse] suffix is a compatibility identifier - ComfyUI Eclipse itself is not required, and existing Eclipse workflows load without node replacement. Nunchaku and GGUF support need ComfyUI-Nunchaku and ComfyUI-GGUF cloned into custom_nodes; without them everything else still works.

Where people get burned

• Legacy formats are denied by default. .ckpt, .pt, .pth, .bin (pickle-capable) are rejected until you enable Smart Model Loader → General → Allow Legacy Model Formats. Safe, but surprising if you live on old checkpoints.
• CLIP errors usually mean you picked the wrong source. UNet and quantized models need clip_source = External; baked only works for standard checkpoints.
• OOM with a big model? Enable block_swap (start around 10 blocks) or model_sampling → Flux for correct scheduler shift - the default settings won't be right for every architecture.