Nodes/RES4LYF/Legacy_SharkSampler
ComfyUI Node Runs on cloud

Legacy_SharkSampler

The older RES4LYF sampler, kept for old workflows

By ClownsharkBatwing·Created 2 years ago·Updated 2 days ago· 1,350
Legacy_SharkSampler
  • model
  • positive
  • negative
  • sampler
  • latent_image
  • sigmas
  • latent_noise
  • latent_noise_match
  • output
  • denoised
  • output_fp64
  • denoised_fp64
◄add_noisetrue►
◄noise_normalizetrue►
◄noise_stdev1.00►
◄noise_mean0.00►
◄noise_is_latentfalse►
◄noise_typegaussian►
◄alpha1.0►
◄k1►
◄noise_seed0►
◄sampler_mode▾►
◄scheduler▾►
◄steps30►
◄denoise1.00►
◄cfg5.0►
◄truncate_conditioning▾►

If you loaded an older RES4LYF workflow and hit a node called Legacy SharkSampler, here's the deal: it's the previous-generation version of the pack's core sampler. When ClownsharkBatwing shipped the newer sampler nodes, he didn't delete the old ones - he prepended "Legacy" to their names so existing workflows keep running. So this isn't a broken or deprecated-and-removed node; it's the old reliable, still there on purpose.

SharkSampler is one of RES4LYF's two foundational samplers (the other being the all-in-one ClownsharKSampler). What made it notable is the same thing that makes the whole pack notable: it's built for both rectified-flow and probability-flow models, it takes a custom SAMPLER and SIGMAS so you can drive it with any of the pack's 69 solvers and modified schedules, and it supports the noise-inversion img2img modes (unsample/resample) that a stock KSampler can't do. If you're starting fresh, use the current sampler nodes. If you're keeping an old graph alive, this is what belongs in it.

How it works

It's a full sampler: model in, latent out. You give it conditioning, a SAMPLER, and either a scheduler+steps or an external SIGMAS line, and it denoises. The unsample/resample modes disable ComfyUI's internal noise handling so RES4LYF's noise-inversion methods work correctly.

The inputs and outputs that matter

  • model, positive, negative, sampler (SAMPLER), latent_image - the standard sampler wiring.
  • output (LATENT) out - the sampled latent for your VAE decode. It also emits denoised, plus output_fp64 and denoised_fp64 high-precision copies for chaining into further high-precision passes.
  • sampler_mode (standard / unsample / resample) - standard for normal generation; unsample to invert an image back toward noise; resample for the pass after unsampling. Get this wrong and img2img inversion won't work.
  • scheduler (default includes beta57, bong_tangent, beta, …) and steps - the schedule, unless you override with the optional sigmas input.
  • cfg (default 5) - as in KSampler. On Flux with guidance, drop to 1.0.
  • truncate_conditioning - a compatibility setting; older workflows sometimes needed true here, and the current sampler exposes it via an extra-options box.

How to install it

ComfyUI Manager: search RES4LYF, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ClownsharkBatwing/RES4LYF
cd RES4LYF
pip install -r requirements.txt

then restart and hard-refresh (F5). Nothing to download.

Common issues & troubleshooting

Should I switch to the new sampler? For new work, yes - the current nodes are where development goes. But you'll have to recreate the sampler and re-enter settings by hand; the layout changed, so it's not a drag-and-drop swap. If the old workflow works, there's no urgency.

Unsampling produces garbage. Almost always sampler_mode. Unsampling needs unsample, and the resample pass needs resample; both disable ComfyUI's own noise addition, which is the whole point. Leaving it on standard breaks the inversion.

A custom scheduler ignores my shift. Same trap as the rest of the pack - bong_tangent applies its own sigma shift and ignores the workflow's. If shift seems dead, change the scheduler to confirm before assuming the node is broken.

CategoryRES4LYF/legacy/samplers

Inputs (23)

NameTypeDefaultDescription
modelMODEL—
add_noiseBOOLEANtrue—
noise_normalizeBOOLEANtrue—
noise_stdevFLOAT1.00-10000–10000—
noise_meanFLOAT0.00-10000–10000—
noise_is_latentBOOLEANfalse—
noise_typeCOMBOgaussian17 options: fractal, gaussian, gaussian_backwards, uniform, pyramid-cascade_B, pyramid-interpolated, +11
alphaFLOAT1.0-10000–10000—
kFLOAT1-10000–10000—
noise_seedINT0-1–18446744073709550000—
sampler_modeCOMBO3 options: standard, unsample, resample
schedulerCOMBO11 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +5
stepsINT301–10000—
denoiseFLOAT1.000–10000—
cfgFLOAT5.00–100—
truncate_conditioningCOMBO3 options: false, true, true_and_zero_neg
positiveCONDITIONING—
negativeCONDITIONING—
samplerSAMPLER—
latent_imageLATENT—
sigmasoptSIGMAS—
latent_noiseoptLATENT—
latent_noise_matchoptLATENT—

Outputs (4)

NameTypeDescription
outputLATENT—
denoisedLATENT—
output_fp64LATENT—
denoised_fp64LATENT—