Sampler Custom Advanced Preview 🍌

ComfyUI's SamplerCustomAdvanced is the fully manual sampler-assembly node - you hand it a NOISE object, a GUIDER, a SAMPLER, a SIGMAS schedule, and a starting latent, and it runs the pass. It's what lets you swap in custom pieces like the ones this pack ships (TCDSampler, SamplerEulerAncestralFixedNoise, TextScheduler, the various custom_noise generators) instead of being stuck with ComfyUI's built-in options. This node is that same building block with one addition: a previews image output alongside the usual two latent outputs, so you get a look at the result without wiring a separate VAEDecode into the graph just to sanity-check a run.

The inputs, all required

• noise - a NOISE object, from ComfyUI's own RandomNoise or any of this pack's custom noise nodes (VariationNoise, SameColorNoise, and the rest of the custom_noise folder).
• guider - a GUIDER object, typically from ComfyUI's CFGGuider.
• sampler - a SAMPLER object.
• sigmas - a SIGMAS schedule.
• latent_image - your starting latent.

These four non-latent inputs are exactly the pieces ComfyUI split the old monolithic KSampler into specifically so third-party packs could swap in custom versions of each - which is the whole reason a pack like this one exists in that ecosystem at all.

The outputs

• output - the final sampled latent, same as SamplerCustomAdvanced's usual first output.
• denoised_output - the fully denoised prediction, useful for certain post-processing or CFG-adjacent tricks that want the model's clean prediction rather than the noisy sampling trajectory's endpoint.
• previews - decoded image(s), the feature this node adds over the stock version. Presumably an automatic decode step baked in, so you don't need your own VAEDecode wired in parallel just to see what you got.

Using it

If you already build sampling graphs manually with the native noise/guider/sampler/sigmas building blocks, this is a drop-in swap for SamplerCustomAdvanced wherever you want a quick look at the result inline. It isn't a replacement for a classic all-in-one KSampler node - it needs the assembled objects, not plain widgets, so you'll need the supporting pieces (native CFGGuider, KSamplerSelect, BasicScheduler, or this pack's own sampler/scheduler/noise nodes) wired in ahead of it regardless.

Installing it

Via ComfyUI Manager: search cgem156-ComfyUI, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/laksjdjf/cgem156-ComfyUI

Restart afterward. No model downloads, no extra dependencies - you still need a VAE loaded in your graph the way any decode step does.

Common issues & troubleshooting

Type errors on the four main inputs. Each of noise, guider, sampler, sigmas is its own object type - none of them accept plain widgets. If you're coming from a classic KSampler, you're missing the upstream nodes that produce these objects, not doing something wrong with this node specifically.

previews comes out empty, wrong-colored, or errors. Check for a VAE mismatch - if the graph's VAE doesn't match what the model expects, decode output (here or anywhere else in the graph) will look wrong. This isn't unique to this node; it's the same failure mode any decode step has.

You just want the final image, not a preview. output still gives you the standard sampled latent to decode normally downstream - the previews output is a convenience add-on, not a replacement for your normal decode/save chain if you want full control over that step.

Unsure which sampler/guider/sigmas nodes to actually use. This node is deliberately unopinionated about that - pair it with whichever of ComfyUI's native building blocks or this pack's own (TCDSampler, SamplerEulerAncestralFixedNoise, TextScheduler) fit what you're trying to do; it doesn't push you toward one combination over another.