AC_FUN_SUPER_LARGE

AC_FUN_SUPER_LARGE is the flagship of the ComfyUI_AC_FUNV8Beta1 pack, and it is exactly what its name promises: a super-sized node that swallows a whole text-to-image pipeline. Checkpoint loader, three LoRA slots, two text encoders, an empty latent, and a KSampler all happen inside this one box. If you've built the standard SDXL or SD 1.5 workflow even once, you know the noise - a dozen nodes and twenty wires between "load the model" and "run the sampler." The pack author, a ComfyUI teacher who goes by Cc啊程 on Bilibili, made these mega-nodes to compress exactly that.

What it actually does under the hood

Don't let the size scare you. Reading the source, this is the same chain you'd build by hand: load_checkpoint_guess_config to load the model, load_lora_for_models for each LoRA, clip.tokenize + encode_from_tokens for the prompt, a torch.zeros latent, and the standard KSampler loop. There's no magic and no API call - it just runs the stock ComfyUI pipeline in one go and hands you everything it produced.

The inputs that matter

The node has a wall of widgets, but only a handful make or break your output:

• boolean (Single / Double / Triple) picks how many LoRA slots apply in order. Pick Single if you only use one - choose Double or Triple and every slot in that count applies at its own strength, so a stray 1.0 in a slot you thought was idle stacks weights you didn't intend.
• select_model - Prompt_to_Image generates from a fresh latent; Image_to_Image VAE-encodes the chosen image and samples from that instead. The img2img path crops the source to a multiple of 8 internally, which is the standard VAE requirement.
• dispatch (Custom / All_ready) - Custom uses your width and height; All_ready takes a preset from the 37-entry resolution dropdown, like 1024*1024 or 768*1280.
• steps, cfg, sampler_name, scheduler default to 10 steps, cfg 1.4, lcm. That's a fast-LCM setup, not a general one - see the gotcha below.

The outputs

You get MODEL, LATENT, Positive, Negative, width, height, and VAE. The important ones: wire LATENT + VAE into a decoder to see the image, or into this pack's AC_Super_SaveImage. The Positive/Negative conditioning outputs are live, so you can feed them into a second-pass sampler or a ControlNet downstream without re-encoding.

Installing it

It ships in the ComfyUI_AC_FUNV8Beta1 pack, installable via ComfyUI Manager (search "AC_FUN") or from a terminal:

cd ComfyUI/custom_nodes
git clone https://github.com/A719689614/ComfyUI_AC_FUNV8Beta1

Restart ComfyUI afterward. There are no model downloads - the node reads whatever checkpoints and LoRAs you already have. One heads-up: the README is Chinese-only, and requirements.txt pins torch==2.11.0. On a current ComfyUI install that's already your torch, so nothing happens; just don't hand-pip the requirements into an older venv and let Manager sort it out.

Where people get burned

The defaults are tuned for LCM-style fast sampling. Point this at a normal SDXL checkpoint without an LCM LoRA and 10 steps at cfg 1.4 comes out washed out - bump steps to 25+ and cfg to 5–7. The bigger trap is philosophical: because sampling happens inside the node, you can't insert anything between encode and sample. Want a ControlNet between text encode and KSampler? You can't, not here - you'd use the pack's separate nodes instead. Think of this as the "iterate prompts fast" node, not the "fine-tune a pipeline" node. For that, the smaller AC_Super_* nodes in the same pack are your tools.