🔶 KSampler txt2img

A normal txt2img chain in ComfyUI is three nodes minimum: Empty Latent Image, KSampler, VAE Decode. This node is all three welded into one box that takes a width and height directly and hands you back a finished image - no latent wire to manage, no separate decode step to forget. The README calls this out as one of the pack's whole selling points ("VAE inclusion in KSampler"), and it's the clearest example of it.

What it does

Set Image_width and Image_height (both default 512, step 8, up to 8192 - step 8 because the underlying VAE works in 8-pixel blocks, same constraint as stock ComfyUI), and this node builds the empty latent internally, runs it through sampling, and decodes straight to IMAGE. There's no separate LATENT output on this node - if you need the raw latent for something else downstream, you want KSampler3 instead, not this one.

Everything else is standard KSampler territory: model, positive/negative (CONDITIONING), vae, seed, steps (default 20), cfg (default 8), sampler_name (34 choices - euler, the dpm++ family, ancestral variants, and so on), and scheduler (9 choices: normal, karras, exponential, sgm_uniform, simple, ddim_uniform, beta, linear_quadratic, kl_optimal). This node has no optional inputs at all - what you see is what you set.

Worth knowing before you touch the defaults: cfg=8 and steps=20 are SD 1.5/SDXL-era numbers, and they're fine for that era - the community's usual sweet spot there is roughly 5–9 with DPM++ 2M Karras as a safe default sampler/scheduler pairing. If you point the model input at something newer - a distilled or guidance-baked checkpoint like a Turbo or Lightning variant - 8 will badly overcook the image; those want CFG close to 1 and far fewer steps. This node doesn't know or care what kind of checkpoint you fed it, so that judgment call is on you.

Why you'd reach for it

Fewer nodes, fewer wires, same result - genuinely just that. If you're building a workflow with several checkpoint/prompt-change points chained together (the animation-style use case this whole pack targets), collapsing three nodes into one at each of those points keeps the graph legible. For a single straightforward txt2img image where you're not fighting canvas clutter, the stock three-node chain works identically and gives you the intermediate latent for free if you ever want it.

Installing it

Via ComfyUI Manager: search "Chaosaiart-Nodes," or "Install via Git URL" with https://github.com/chaosaiart/Chaosaiart-Nodes, then restart. Manually: cd ComfyUI/custom_nodes && git clone https://github.com/chaosaiart/Chaosaiart-Nodes, restart ComfyUI. Linux needs pip install opencv-python and pip install tqdm (venv active first) for the pack to import cleanly - other nodes in it need those libraries even though this one doesn't. Windows has a bundled Install_windows script. No models ship with the pack; bring your own checkpoint.

Troubleshooting

If output looks burned, oversaturated, or has deformed faces, that's classic too-high-CFG-for-the-model - drop cfg before touching anything else, especially if you loaded a distilled checkpoint into this node. If images come out muddy or the prompt seems ignored, CFG is probably too low for a non-distilled model (below roughly 3). And since Image_width/Image_height step in increments of 8, don't fight the widget trying to enter an odd number - it'll round, and generating well outside your checkpoint's native resolution (512² for SD 1.5, 1024²-class for SDXL) is what produces the duplicated-limbs, tiled-pattern look people usually blame on the sampler when it's actually a resolution mismatch.