Legacy2_UltraSharkSampler

The display name gives it away before you've even read the tooltip: Legacy2_UltraSharkSampler. RES4LYF unified most of its sampling into the newer ClownsharKSampler and prepended "Legacy" to the nodes it superseded, and this is one of them - kept around specifically because Stable Cascade and UltraCascade workflows still lean on it.

Stable Cascade never really got its moment. It had strong prompt adherence and good aesthetic scores at launch, but SD3's announcement landed while the ecosystem was still deciding whether to invest, and by the time SD3 disappointed, people had already started waiting for Flux instead. Most tooling moved on. RES4LYF's author didn't entirely - UltraCascade is a separate project built on top of Cascade, and this sampler is part of the Cascade-era plumbing the pack still carries.

How it works

It's a KSamplerAdvanced-style all-in-one node, purpose-built for Cascade/UltraCascade's own guide-based img2img system rather than the newer cross-architecture sampling framework the rest of the pack has moved to. You still pick your actual solver externally (a SAMPLER input, so res_2m and friends still apply) and your own sigma schedule - this node's job is running that combination against Cascade's guide mechanism.

The inputs and outputs that matter

• model, latent_image, positive, negative - the basics any sampler needs.
• sampler (SAMPLER) and sigmas (SIGMAS) - your chosen solver and schedule, fed in rather than picked from a dropdown.
• noise_type (17 choices - fractal, gaussian, several pyramid variants, and more) with alpha / k shaping it - this is the pack's noise-tuning system, part of the parameter-modulation-over-time approach RES4LYF is built around.
• cfg, add_noise / normalize_noise, noise_seed - standard sampling controls.
• guide_type (residual or weighted) with guide_weight, plus optional guide (LATENT) and guide_weights (SIGMAS) - this is the img2img-style latent guidance: feed in a reference latent and the sampler nudges toward it rather than starting from pure noise.
• Outputs: output (your sampled latent - send this to Stage B or your VAE decode), denoised_output (the sampling noise stripped out, worth grabbing separately if you want to preview the model's clean prediction), and latent_batch (bundles multiple sampled latents together, useful mainly when batching seeds or steps).

How to install it

Via ComfyUI Manager: search RES4LYF, install, restart.

Manually:

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

Portable installs: use the embedded Python's pip.exe. You'll need Stable Cascade or UltraCascade's own checkpoint stack - this node won't do anything useful with a regular SD or Flux model.

Common issues & troubleshooting

Wrong model family. This is architecture-specific - Cascade or UltraCascade checkpoints only. A standard SD/SDXL/Flux model plugged in here won't sample correctly.

Building a fresh Cascade workflow and unsure whether to use this or ClownsharKSampler. Check the README's current example workflows first - this node still works, but "Legacy" in its own name is the pack telling you where new development isn't happening. If you're starting from scratch rather than maintaining an existing Cascade workflow, it's worth checking whether the newer unified sampler now covers your case before building around this one.

Unsure which output to use. Default to output for a normal pipeline; reach for denoised_output only if you specifically want the clean prediction separate from the sampling noise, and latent_batch only if you're actually batching multiple latents together.