sum_load_simple

Stock ComfyUI wants a different loader for every architecture: CheckpointLoader for SD, UNETLoader + DualCLIPLoader for Flux, a whole extra loader when you start quantizing. ComfyUI-Apt_Preset's sum_load_simple is the one-node answer - a universal loader that knows the model-combination rules for XL/SD, Wan 2.1/2.2, QwenEdit, Flux/Kontext, SD3.5, and Hi-dream, and produces the pack's RUN_CONTEXT bundle plus a model and clip. It's the "Loader" in the README's Loader → Controller → Sampler design, and it's the node you'll start almost every workflow with.

What it actually does

It's a stripped-down subclass of the pack's sum_load_adv - same loading engine, minus the preset system, the extra clip slots, and the inline prompt fields. You point it at the right files for your architecture, it loads model + clips + VAE, encodes nothing yet, and packs everything into a context with your sampling defaults baked in.

The architecture logic lives in the parent loader, and it matters more than the file dropdowns:

• XL / SD mode - a single checkpoint.
• Wan 2.1 / 2.2 / QwenEdit - Unet + clip1.
• Flux / Kontext - Unet + clip1 + clip2 (the README is explicit: in that order, not clip1 + clip3).
• SD3.5 - Unet + clip1 + clip2 + clip3.
• Hi-dream - Unet + clip1 through clip4.

Get that file combo wrong and the loader silently half-loads your pipeline.

The inputs that matter

• ckpt_name / unet_name - checkpoint for SD-class, or Unet for the DiT-era models. The dropdowns pull from models/checkpoints and models/unet - and crucially, the Unet list includes unet_gguf files, so quantized models show up here automatically once ComfyUI-GGUF is installed.
• clip_type - 26 entries covering stable_diffusion, sd3, wan, hidream, ltxv, pixart, cosmos, and friends. This is what tells the loader which combination rule applies and what latent shape downstream expects.
• clip1 / clip2 - text encoders from models/text_encoders (plus clip_gguf when ComfyUI-GGUF is present).
• vae - from models/vae.
• unet_Weight_Dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2 for fp8 Unet loading.
• lora / lora_strength - a single inline LoRA if you don't want the dedicated sum_lora node.
• width / height / steps / cfg / sampler / scheduler - the defaults stored into context; downstream nodes inherit them.
• over_model / over_clip - override inputs. The README's rule: when over_model is connected, whatever the loader was going to output for model/Unet is discarded and over_model wins; same for over_clip and all four clip slots.
• lora_stack - for the pack's stacked LoRA tokens.

The outputs that matter

context (the bundle - the thing almost everything else consumes), model, and clip.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset.git
cd ComfyUI-Apt_Preset
pip install -r requirements.txt   # Windows: double-click install.bat

Then, from the README's note list: ComfyUI-GGUF is required to actually load GGUF files - the loader lists them, but ComfyUI-GGUF provides the loader implementation behind the scenes.

Common issues

The parent loader caches loaded models by an MD5 of your settings, which makes re-runs fast but means a changed file with the same name can serve you stale weights until you restart or clear the cache. And the None samples in every dropdown are a footgun: leave clip_type at None and the loader can't pick a combination rule, so nothing sensible comes out. If your workflow "loads" but every generation is garbage, check that you didn't leave a clip slot on None when your architecture needs two.