Model Loader

Not everyone wants the all-in-one. If the Smart Model Loader feels like a Swiss Army knife you don't need to carry, Model Loader [Eclipse] is the plain screwdriver: the same model-loading engine from the same pack, but it hands you raw sockets instead of one bundled PIPE. You get model, clip, vae, audio_vae, and model_name straight out, ready to wire into a normal KSampler, CLIP Text Encode, and VAE Decode like you've always done.

It shines for UNet-only and quantized setups. Flux, SD3, and AuraFlow are usually distributed as standalone diffusion models with no baked CLIP or VAE - so you need a loader that handles that file type directly, then pair it with the pack's CLIP Loader and VAE Loader. For a stock checkpoint with everything baked in, either loader works; the standalone version just keeps the rest of your graph conventional.

The inputs that matter

There are a lot of them, but you'll touch a handful. model_type switches the whole personality of the node - Standard Checkpoint, UNet, Nunchaku (Flux/Qwen/ZImage), or GGUF - and the relevant name field lights up: ckpt_name, unet_name, nunchaku_name, qwen_name, zimage_name, or gguf_name. Then:

• LoRA slots - lora_count (1–3), each with lora_switch_N, lora_name_N, lora_weight_N. Model-only patching, no separate CLIP weight.
• sampling_method - pick the architecture and the loader applies the correct scheduler shift for you: SD3 shift 3.0, AuraFlow 1.73, Flux max_shift 1.15, Stable Cascade 2.0, LCM distilled, LTXV for video. Leave on None for a standard model and forget it.
• blocks_to_swap - GPU→CPU transformer offload. The tooltip even gives suggested counts per architecture (Flux ~10 of 57, SD3 ~8 of 24, Wan ~10 of 40). Set to 0 to disable. offload_embeddings grabs another ~100–300MB.
• ltx_text_encoder - the special one. For LTX2/LTXV, point it at a Gemma text encoder (GGUF or safetensors) and it combines it with the model file's baked text projection to build a correct LTXAV CLIP.

Outputs

• model → KSampler
• clip → CLIP Text Encode (None for UNet/Nunchaku/GGUF - add the CLIP Loader)
• vae → VAE Decode (None for those same types - add the VAE Loader)
• audio_vae → LTXV/LTX2 audio decode
• model_name → handy for filenames or metadata

Installing

Same story as the whole pack - Manager, search ComfyUI Smart Model Loader, or:

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. Nunchaku and GGUF types need ComfyUI-Nunchaku and ComfyUI-GGUF in custom_nodes; without them the quantized options are just disabled and standard checkpoints still load.

Common issues

• clip comes out None on a UNet - that's expected, not a bug. The model file genuinely has no text encoder. Pair with CLIP Loader and set clip_type to match (Flux wants 2 CLIP modules).
• VRAM OOM - go GGUF or Nunchaku, or raise blocks_to_swap before you buy a new card.
• Weird sampler behavior on Flux - set sampling_method to Flux so the shift is computed for your resolution instead of using a generic default.