sum_load_adv

Most model families in ComfyUI need a different combination of loader nodes - a checkpoint loader alone for SD/SDXL, a UNet loader plus one CLIP for Wan, a UNet plus two CLIPs in a specific order for Flux, three for SD3.5, four for Hi-Dream - and remembering which combination goes with which model, then wiring it correctly every time, is exactly the kind of repetitive setup Apt_Preset was built to collapse into one node. sum_load_adv is that node: it's the pack's "Universal Loader," named directly in the README, and it covers every one of those model families from a single set of inputs.

The model combinations, straight from the author's own documentation. XL and SD models: load a checkpoint, or a UNet alone. Wan 2.1/2.2 and QwenEdit: UNet plus clip1. Flux and Kontext: UNet plus clip1 and clip2 - and the README calls this out specifically because the order matters: it's clip1 then clip2, not clip1 and clip3. That's an easy mistake to make if you're used to seeing three CLIP slots on this node and just filling in the first two you notice; skip straight to clip3 and Flux/Kontext loading will be wrong. SD3.5 needs clip1 through clip3, and Hi-Dream needs all four. clip_type (26 choices, including stable_diffusion, sd3, wan, hidream, pixart, and more) tells the node which text-encoder family it's actually dealing with, and unet_Weight_Dtype lets you force fp8 variants (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2) if you're trying to save VRAM on the UNet itself.

GGUF support rides along automatically. The README lists this as one of the loader's headline features, and it works the way you'd expect from any GGUF-aware loader: install ComfyUI-GGUF first, and quantized .gguf checkpoints and UNets become selectable in ckpt_name/unet_name right alongside full-precision files, no separate loader node required. GGUF is the standard way to fit big diffusion transformers - the ones this node targets, Flux, Wan, SD3.5, Hi-Dream - onto consumer VRAM by dropping from 16-bit down to a 4–8 bit quantization ladder; Q8 is close enough to full precision that the difference is basically noise, while Q4 is the accepted floor before quality loss gets obvious. If you're loading a quantized file through this node and it's not showing up, that's ComfyUI-GGUF's install you need to check, not sum_load_adv itself.

Two more things worth knowing. over_model and over_clip are override inputs - feed either one, and the node stops using its own internal loading for that piece and just passes through whatever you connected instead. And the whole node doubles as a preset system: preset must have something selected before the node will run at all (leave it on None and you get an error), you can freely tweak any parameter after picking a preset, and if you save your tweaks as a new preset, it needs a ComfyUI restart before it appears as a selectable option - annoying the first time, expected every time after.

Beyond loading, the node also bundles sampling settings directly - width, height, steps, cfg, sampler, scheduler, and starting pos/neg prompts - plus lora/lora_strength/lora_stack for LoRA. Outputs: context (RUN_CONTEXT, feeding the rest of the pack's Controller/Sampler stack), model (a plain MODEL wire), and preset_save (PDATA, the save-slot output).

Installing it. Standard for the pack - search "ComfyUI-Apt_Preset" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset.git

Run install.bat on Windows for dependencies, restart. If you need GGUF loading, install ComfyUI-GGUF too - that's a separate pack the README explicitly asks for.

Troubleshooting. Wrong output for Flux/Kontext is almost always the clip1+clip3 mistake above - check you're using clip1 and clip2. A node that errors instantly usually means no preset selected. A preset you just saved that isn't showing up means you skipped the restart. And a broader IMPORT FAILED on the pack itself is a documented, if separate, issue - check your ComfyUI console log for the actual missing dependency and reinstall via Manager rather than assuming this specific node is at fault.