⚡ KSampler (Illustrious)

Swap this in for a vanilla KSampler and the first thing you'll notice is the defaults are already right for Illustrious: CFG 5.2 instead of 8, euler_ancestral and karras instead of euler/normal. That's not decoration - the community's settled Illustrious range is CFG 4–7 with a sweet spot around 5, and this node ships inside that band instead of the generic SDXL default that tends to burn Illustrious output.

What the extra knobs do

Auto CFG keeps your CFG inside the sweet spot for whichever Model Version it detects (or you set explicitly), and Resolution Adaptive nudges step count for larger or smaller images so you're not under-stepping a 1536px render or wasting time over-stepping a 768px one. Color Safe Mode is the opposite instinct - turn it on and the node backs off its own auto-tweaks and uses your values more literally, useful once you've dialed in settings you like and don't want the node second-guessing you anymore.

Edge Feather (px), Edge Tone Match, and Tone Strength only matter when you're feeding this sampler a masked latent - an inpaint or regional workflow. They feather the mask edge in latent space before reblending the result, and optionally match tone near the seam before that reblend. This is the sampler-side half of the fix for the classic inpainting failure mode: a visible ring or color mismatch where the generated patch meets the original. If your IllustriousVAEEncode mask blur alone isn't cleaning up a seam, this is the next place to look.

The sigmas input lets a separate scheduler node hand this sampler a raw noise schedule directly, bypassing sampler_name/scheduler entirely - for when you want full manual control over the denoising trajectory.

The inputs and outputs that matter

• model, seed, positive, negative, latent_image - the standard KSampler set, all required.
• cfg defaults to 5.2, steps to 26, scheduler to karras - all already inside Illustrious's comfort zone.
• Model Version - leave on Auto Detect unless you know you're on a specific EPS/VPred build and want to force it.

Outputs go beyond just LATENT: it also passes through model, positive, and negative unchanged, so a multi-sampler graph (a second pass, a detailer) can pull the same wiring from this node's output rather than routing back to the original source nodes.

Installing it

ComfyUI Manager: search "EasyIllustrious," install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/regiellis/ComfyUI-EasyIllustrious
pip install -r ComfyUI-EasyIllustrious/requirements.txt

No extra downloads - it's a drop-in sampler for whatever model/VAE/CLIP you've already got loaded.

Common issues

Output still looks fried even with CFG 5.2. Check Auto CFG and Model Version - if you're actually on a v-prediction checkpoint (NoobAI vpred, an Illustrious vpred build), the fix isn't a lower eps-style CFG, it's a different regime entirely: CFG in the 1.5–2.0 range paired with a CFG++ sampler, or a rescale approach. Auto CFG here is tuned for standard epsilon models - set Model Version explicitly to a VPred option if that's what you're running, and don't expect this dial to compensate on its own.

Seams still visible after enabling Edge Tone Match. Raise Tone Strength gradually rather than jumping straight to 1.0 - too strong a match can flatten legitimate contrast right at the boundary, which reads as its own kind of artifact.