sum_create_chx

Every ComfyUI workflow is a bundle of wires, and the more wires, the more ways to get them wrong. ComfyUI-Apt_Preset's answer is a "context" - a single RUN_CONTEXT object that rides down the graph carrying your model, clip, VAE, conditioning, latent, and all your sampler settings in one bundle. sum_create_chx is the node that manufactures that context from scratch when you don't want to use the full loader (sum_load_adv / sum_load_simple).

"chx" is the pack author's shorthand for context, and the design intent is in the README: a Loader → Controller → Sampler framework so workflows stay readable instead of becoming the spaghetti that sent everyone to rgthree and friends. sum_create_chx is the "Loader" end for people who prefer to bring their own model and just want the context scaffolding built around it.

What it actually does

It assembles a context dict (model, positive/negative conditioning, latent, vae, clip, steps, cfg, sampler, scheduler, guidance) and - the part that surprises people - it creates the starting latent by encoding a black image at your requested size. No random noise canvas: a literal black RGB image gets VAE-encoded into the latent you'll sample from. That's why it needs a VAE even if you're not planning to decode anything.

The inputs that matter

• vae - required. Pick one from your models/vae folder, or override with over_vae if you're feeding one in from elsewhere.
• width / height / batch - canvas geometry and batch size for the black-image latent.
• steps / cfg / sampler / scheduler / guidance - the sampling defaults stored into the context. Every downstream node that doesn't override these inherits them, which is the whole point of the pack: set once here, forget.
• pos / neg - text prompts, encoded through the connected clip into positive/negative conditioning. Defaults are the classic a girl / worst quality, low quality.
• model / clip - optional. Wire them in if you're loading elsewhere; leave unplugged and the context just carries whatever comes next.
• over_positive / over_negative / over_latent / lora_stack / data - the overrides. data is a wildcard passthrough, handy for shuttling arbitrary payloads through the graph.

The outputs that matter

Eight of them: context (the bundle), plus model, positive, negative, latent, vae, clip, and a data passthrough. Wire context onward; the rest are there when you want to grab an individual value for a non-pack node.

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

Or search ComfyUI-Apt_Preset in ComfyUI Manager.

Common issues

The biggest trap: context is a snapshot, not a live link. If you change a value on this node (say, steps 20 → 30) after it's already executed, the nodes downstream that already pulled values from the old context don't auto-update until you re-run the chain. ComfyUI handles that fine on re-execute, but mid-workflow edits can leave you scratching your head if you're staring at stale outputs. Also, remember it VAE-encodes a black image - if your vae field is set to None (the dropdown sample starts there), the node can't build its latent and you'll get an error the moment something downstream asks for one.