AC_Super_KSampler

AC_Super_KSampler is the compact way to sample: one node that takes a model, a clip, and two plain-text prompts, and does the text encoding, the latent creation, and the sampling itself. No Empty Latent node, no CLIP Text Encode nodes, no conditioning wires flying around - type your prompt, hit go.

How it works

Read the source and you'll see it's three stock operations fused: it builds an empty latent with torch.zeros (sized by your resolution choice), runs clip.tokenize + encode_from_tokens on both prompt strings using the clip you wired in, then hands everything to the standard KSampler loop. The mechanism is identical to a hand-built chain - it's just all inside one node, with the same comfy.sample.sample call under the hood.

The inputs that matter

• model and clip - from a checkpoint loader, like AC_Super_Checkpoint, AC_Super_CKPT&LCM, or the core Load Checkpoint. Note the clip is required: unlike a normal KSampler, this node encodes text itself, so it needs a CLIP, not conditioning.
• positive / negative - multiline prompt strings, not conditioning. This is the biggest difference from the core KSampler.
• dispatch (Custom / All_ready) - Custom uses width/height/batch_size; All_ready pulls size from the 37-entry resolution dropdown.
• seed, steps, cfg, sampler_name, scheduler, denoise - the standard knobs, defaulting to 10 steps, cfg 1.4, lcm.

The outputs

LATENT (the sampled result), plus Positive and Negative as CONDITIONING. Those two are the useful bonus: they're the encodings the node already computed, so you can feed them into a second-pass node like AC_Super_UpKSampler without re-encoding the prompts.

Installing it

Ships in ComfyUI_AC_FUNV8Beta1 by Cc啊程 (Bilibili ComfyUI teacher; README is Chinese). ComfyUI Manager (search "AC_FUN") or:

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

Restart ComfyUI. No downloads.

Where people get burned

The defaults are LCM-flavored: lcm sampler, cfg 1.4, 10 steps. Those numbers only look good with an LCM LoRA on the model. Point it at a normal SDXL or SD 1.5 checkpoint and you'll get a washed-out smudge - raise steps to 25+ and cfg to 5–7. Second, because encoding happens inside, you can't interpose a ControlNet between encode and sample; for that, use AC_Super_CLIPEN + a core KSampler (or AC_Super_Controlnet on the conditioning) instead. And keep in mind denoise below 1.0 on a freshly made empty latent does nothing useful - it's there for when you feed it pre-existing latents.