Skip Layers

Skip-layer guidance is one of those tricks that sounds like a bug and works anyway: you run a second, deliberately crippled pass through the model with a few of its middle blocks switched off, then steer away from the result. The crippled pass is worse in a specific way - usually anatomy and structure - and subtracting it firms up exactly the thing it broke.

ComfyUI has a stock node for this. It just doesn't work on most of the models you'd want it for, because it asks the model to skip a block through patches_replace, and only Flux 1, SD3 and Wan honour that request. Flux 3 and Krea 2 don't. Skip Layers goes around the problem entirely: it wraps the model's apply_model call and swaps the chosen blocks' forward out for one that hands its input straight back. No cooperation needed from the model at all - anything with a list of residual blocks gets the same treatment.

The fields

model in, model out. Everything else is a setting.

blocks is a string, 7, 8, 9 or 7-9, counting from 0. On Flux 3 those numbers are the 28 single blocks; the author's warning is blunt and worth repeating - avoid the first two and the last three, they tend to break the image rather than guide it. Middle blocks are where structure gets decided, which is why the default is 7–9.

mode picks between the two things this could mean. guidance is the real feature: one extra model call per step, with the blocks skipped, and the result steered away from it by scale (default 3), following the same rule as core's SLG - result + (cond - cond_skipped) * scale. remove skips the blocks in every call between start_percent and end_percent. That's not a production setting; it's for seeing what the blocks contribute, and the images look like it.

start_percent and end_percent (0.01 and 0.15 by default) fence the whole thing to the early steps. Skip-layer guidance helps most while the layout is forming; running it to the end of sampling usually just adds crunch. Because the window is short, the extra pass is cheap - you're paying roughly 10–15% more sampling time, not double.

rescale is the safety valve. High scale can burn contrast, so this pulls the result's standard deviation back toward the unguided prediction. Start at 0; reach for it if your image is coming out hot.

Where it earns its keep

The interesting case in 2026 is guidance-distilled models. On a Turbo or distilled checkpoint you're sampling at CFG 1, which means there is no unconditional pass - your negative prompt is being thrown away before it reaches the model. If hands are falling apart there, the usual lever ("raise CFG, add bad anatomy to the negative") is dead. Skip-layer guidance is a second lever that doesn't need CFG at all, because it computes its own degraded prediction. It's the same reason NAG got popular: it puts a corrective signal back into a model that has no CFG slot for one.

Install

Search Nynxz in ComfyUI Manager, or:

comfy node install comfyui-nynxz-nodes

From source:

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-NynxzNodes
cd ComfyUI-NynxzNodes
pnpm install && pnpm build

Then restart. No Python dependencies - this node is pure ComfyUI plumbing over core's patcher. ZenKit is optional.

The traps

Block numbers that are quietly ignored. The node looks for a block stack named single_blocks, blocks, transformer_blocks or layers, and indices outside that list are simply skipped - no error, no effect. Feed a 28-block Flux 3 model blocks: 40 and you'll conclude the node doesn't work, when actually it did nothing. If a run looks identical to no node at all, that's the first thing to check.

Wrong mode for what you wanted. remove at a wide window destroys images; people try it once, see garbage, and write the pack off. guidance is the one you leave in the workflow.

And don't stack it blindly with every other guidance trick. Skip-layer guidance, NAG and a high CFG on a distilled model all push in the same direction and the result oversaturates. One at a time, fixed seed, compare.