Nodes/Nougan Nodes/Nougan Ultimate KSampler
ComfyUI Node

Nougan Ultimate KSampler

Nougan Ultimate KSampler

By Winnougan·Created 4 months ago·Updated 3 days ago· 31
Nougan Ultimate KSampler
  • model
  • positive
  • latent_image
  • negative
  • vae
  • audio_vae
  • sigmas
  • options
  • latent
  • denoised
  • image
  • audio
◄sampler_nameeuler►
◄schedulerbeta►
◄steps30►
◄steps_to_run-1►
◄cfg5.50►
◄cfg_enabledtrue►
◄denoise1.00►
◄seed0►
◄eta0.50►
◄sampler_modestandard►
◄bongmathtrue►
◄decodetrue►
◄tiled_decodefalse►
◄preview_fps12►
◄previewtrue►
◄preview_max_res1024►
◄preview_frames1►
◄preview_jpeg_quality80►
◄preview_tiny_vaenone►

The 🦈 node is the pack's attempt to collapse the boring half of every video workflow into one box: sampler, VAE encode, VAE decode, audio-VAE handling and a live preview, in one node with one click fewer to misconfigure.

Whether you want it depends on one question - do you actually use RES4LYF? Because that's the feature that isn't easy to get anywhere else. Everything else is convenience.

What it's really doing

The sampler dropdown is the union of ComfyUI's built-in samplers and RES4LYF's (ClownsharK) list, and the prefix is the switch. Anything listed as clown: <name> runs through ClownsharKSampler_Beta - at which point eta, bongmath, sampler_mode and the optional options chain actually do something. Anything without the prefix runs through the stock ComfyUI sampling path, and those same four fields do nothing. That's the single most important thing in this article.

The scheduler dropdown mixes the nine stock schedules with RES4LYF's own. If you pick an RES4LYF-only scheduler like beta57 while running a built-in sampler, it falls back to beta, which is documented behaviour and also why "why doesn't beta57 feel different" is a question people ask.

The I/O half: connect image (+ vae) and it encodes your frames into the starting latent, no separate VAE Encode node. Connect audio (+ audio_vae) and it encodes that too - and if you give it both, they're concatenated into one joint audio+video latent, the shape MiniMax H3 and LTX-2 workflows want. Then decode turns the result back into image (or a frame batch) and, if an audio VAE is connected, audio. Turn decode off and you get pure latents.

The preview is KJNodes' Model Preview Override wrapper attached to your model for the run, feeding the node's own DOM panel. Without KJNodes it falls back to ComfyUI's default preview and logs a warning - no crash.

The fields worth setting

steps, cfg, denoise, seed are what you expect. cfg accepts negative values (−100 to 100), which is for exotic distilled/CFG-zero setups rather than a general tool.

sampler_name defaults to euler, scheduler to beta. For clown samplers, eta (default 0.5) is noise added then removed each step, sampler_mode is standard / unsample / resample, and bongmath (on by default) is RES4LYF's own toggle.

tiled_decode is your VRAM escape hatch for big video latents. preview_fps, preview_max_res, preview_frames and preview_jpeg_quality control how pretty the panel is - preview_frames above 1 gives an animated frame-strip view, which is genuinely useful for spotting a clip that's falling apart at step 15.

Outputs: latent, denoised, image, audio. denoised is the pre-decode latent - feed it to a tiled or upscale decode path if the built-in one isn't good enough.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/Nougan_Nodes
# restart ComfyUI

Or ComfyUI Manager, search Nougan. Then, for the parts that make it worth using:

  • RES4LYF (ComfyUI_Res4lyf) - the clown samplers. It's discovered at runtime by scanning loaded modules, not imported at boot, so the node works fine without it; you just get a shorter sampler list.
  • ComfyUI-KJNodes - the live preview panel.

One thing to flag that the node can't: RES4LYF is AGPL-3.0 with a rider forbidding commercial image-generation services from using it without a separate licence - a harder line than any comparable pack takes (res4lyf.json). If you're building something you sell, that matters more than any sampler setting here.

Where it bites

You pick eta/bongmath and nothing changes. You're on a non-clown sampler. The tooltips say "Clown samplers only" and they mean it.

Audio connected, no audio_vae. It raises a clear error rather than silently dropping your track, which is the correct behaviour and still catches people.

decode off, and you wonder why the image output is empty. Obvious once you know, invisible when you don't.

The preview panel is dead. KJNodes missing, or the browser cache hasn't picked up the node's JS. Ctrl+F5 first, install KJNodes second.

Huge latent, instant OOM. Video latents at full resolution plus a preview at preview_max_res 1024 is more memory than you think. Drop preview_max_res to 512 and turn on tiled_decode.

My take: if you've ever tried to adopt RES4LYF, you know the friction is discoverability, not capability - most people meet it as one sampler substituted inside someone else's workflow and never learn what the rest of it does (clownsharkbatwing.json). A single node that puts 44 sampler entries and 9 schedules behind two dropdowns, with the VAE plumbing already attached, is a genuinely good on-ramp. Just go in knowing the shark-themed buttons only bite in clown mode.

CategoryNougan/sampling

Inputs (27)

NameTypeDefaultDescription
modelMODEL—
positiveCONDITIONING—
latent_imageLATENT—
sampler_nameCOMBOeulerBuilt-in ComfyUI samplers plus RES4LYF samplers (prefixed 'clown: ').
schedulerCOMBObeta'none' = the schedule comes from the wired sigmas input. RES4LYF-only schedulers (e.g. beta57) fall back to 'beta' on a built-in sampler.
stepsINT301–10000—
steps_to_runINT-1-1–10000Chainsampling: run only the FIRST N steps of the schedule, then stop partially denoised (a second sampler continues from there). -1 = all steps.
cfgFLOAT5.50-100–100Classifier-free guidance scale. Ignored when cfg_enabled is OFF.
cfg_enabledBOOLEANtrueOFF for CFG-free models: samples with positive conditioning only (BasicGuider) at half the per-step cost. negative and cfg are ignored while OFF. (MiniMax H3 has its own node: Nougan MiniMax H3 KSampler.)
denoiseFLOAT1.000–1—
seedINT00–18446744073709550000—
etaFLOAT0.50-100–100Clown samplers only: noise added then removed after each step.
sampler_modeCOMBOstandardClown samplers only.
bongmathBOOLEANtrueClown samplers only: RES4LYF's bongmath on/off.
decodeBOOLEANtrueDecode the result with the connected VAE(s). Off = latent only.
tiled_decodeBOOLEANfalse—
preview_fpsINT121–60Playback FPS of the live preview (ignored for clips with audio).
negativeoptCONDITIONINGOptional: leave unplugged for positive-only (CFG-free) sampling — cfg is ignored then.
vaeoptVAEVideo/image VAE: decodes the result into the IMAGE output and can drive the preview.
audio_vaeoptVAEAudio VAE: decodes the audio half of the result into the AUDIO output and drives the spectrogram preview.
sigmasoptSIGMASAdvanced: custom sigma schedule. Replaces scheduler / steps / denoise when connected.
optionsoptOPTIONSRES4LYF options chain (clown samplers only).
previewoptBOOLEANtrue—
preview_max_resoptINT10240–8192—
preview_framesoptINT11–1024Frames per step for the animated video preview (1 = single frame).
preview_jpeg_qualityoptINT8030–100—
preview_tiny_vaeoptCOMBOnoneTiny VAE decoder (models/vae_approx) for true-RGB previews, like KJ's Model Preview Override. Needs the decoder code: vendored tiny_vae.py or a KJNodes install.

Outputs (4)

NameTypeDescription
latentLATENT—
denoisedLATENT—
imageIMAGE—
audioAUDIO—