Icy KSampler

Let's get the elephant in the room out of the way: IcyKSampler is ComfyUI's core KSampler. The ComfyUI-IcyHider pack (a preview-hiding privacy extension) auto-wraps every core node into an Icy subclass at startup - same inputs, same math, renamed class so the extension can hide its widgets until you hover. If you don't use the hiding, the plain KSampler is exactly this node. Fine. Now let's talk about the actual workhorse.

The KSampler is the heart of every diffusion workflow. It takes a latent canvas (usually from EmptyLatentImage), fills it with noise, and runs the denoising process - the loop that turns noise into an image over steps iterations. Everything else in your graph is setup for this moment: the model, the prompt conditioning, the resolution. This node is where pixels are born.

The inputs that matter (for a beginner, anyway)

• model (MODEL), positive / negative (CONDITIONING), latent_image (LATENT) - your loaded checkpoint, your encoded prompts, your starting canvas. The plumbing.
• seed (INT) - the noise's starting point. Same seed + same settings = same image. Change it to explore.
• steps (INT, default 20) - how many denoising iterations. The KB's practical guidance: 20–30 for quality, 10–15 for quick tests, and always check your model's own recommendation - distilled models want as few as 4–8.
• cfg (FLOAT, default 8) - classifier-free guidance: how hard the model obeys the prompt. 5–9 was the SD 1.5/SDXL range, but the KB notes CFG 1 is the modern default for many 2026 models (distilled and flow-matching checkpoints often want it near 1). The old 8 may be wrong for your model - read the model card.
• sampler_name (enum, 44 choices) + scheduler (enum, 9 choices) - the algorithm and the noise schedule. This pair is where people lose the most time. The KB's safe defaults: DPM++ 2M + Karras for SD 1.5/SDXL, Euler + normal/beta for Flux, and never Karras on flow-matching models - it's a genuine trap. If you're on a modern base, its card tells you the combo.
• denoise (FLOAT, 0–1, default 1) - how much of the latent gets regenerated. 1.0 is full txt2img; lower values preserve the input image's structure (img2img, inpainting passes). 0.3–0.6 is the inpainting sweet spot, per the KB.

Output: a single LATENT - wire it into VAEDecode to see the image.

Where people get burned

• Sampler/scheduler confusion. They're different knobs - sampler is the algorithm, scheduler is the noise schedule - and they get blamed for each other constantly. DPM++ 2M with Karras is one thing; "my Flux output is broken" is often "I left Karras on."
• Default 8 CFG on a modern model. On Z-Image/Flux-class distilled models that's usually way too high. Trust the model card.
• The Icy wrapper is not a new sampler. It's the same 44 samplers and 9 schedulers as core KSampler. It doesn't add RES4LYF or anything else - that's a separate pack.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/icekiub-ai/ComfyUI-IcyHider

Restart, or use ComfyUI Manager (search "IcyHider"). No pip dependencies, no models.

Common issues

• Missing node in shared workflows. Install the pack or swap in core KSampler.
• "Different result than the workflow's screenshot." Check seed, then sampler/scheduler, then CFG - in that order. One of them differs.
• Works but slow. Not a bug - it's denoising. Lower steps or move to a distilled variant of your model.