SharkSampler

SharkSampler is the "Shark" half of RES4LYF's flagship ClownsharKSampler - the part that actually does the work. Where ClownSampler just builds a sampler definition, SharkSampler takes a model, your conditioning, and a latent, and runs the denoise loop to produce an image latent. If you've split your sampling graph into modular pieces, this is the node everything else feeds into.

It's the natural anchor for a hand-built RES4LYF workflow. You wire a solver (from ClownSampler or ClownSamplerAdvanced) into its sampler input, optionally a schedule into sigmas, your prompts into positive/negative, and it runs the whole thing with the pack's high-order solvers and the beta57 schedule that made RES4LYF the default place sampler tuning happens for flow-matching models.

How it works

It's a KSampler in spirit - model in, latent out - but built around RES4LYF's solvers and with a few extra tricks. It'll generate null conditioning automatically if you leave positive or negative unplugged, which is handy for unsampling. And it has three sampler modes: standard for normal generation, unsample for running the process backward (noise inversion), and resample for the pass that follows an unsample. Those two extra modes disable ComfyUI's internal noise addition, which is exactly what unsampling-based img2img and editing need to work.

The inputs and outputs that matter

The ones a beginner sets:

• scheduler (default beta57) - the noise schedule. beta57 is RES4LYF's signature and the single most-borrowed thing in the pack; it's a fine default for flow-matching models. A sigmas input, if you connect one, overrides this.
• steps (INT, default 30) - step count. The pack's own advice is that ~20 is often enough with the res solvers.
• cfg (FLOAT, default 5.5) - same as any sampler. Set it to 1.0 on distilled/guidance-Flux setups; 3–5.5 suits dedistilled models with real CLIP conditioning.
• denoise (FLOAT, default 1) - full denoise for txt2img, lower for img2img.
• sampler_mode (standard / unsample / resample) - leave on standard unless you're doing noise-inversion editing.

Optional inputs include model, positive, negative, sampler (a SAMPLER from a Clown node), sigmas, and latent_image. Outputs are output (the result LATENT, decode this), denoised (the clean prediction LATENT), and options (an OPTIONS object for chaining).

Installing it

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

pip install in the venv or portable Python, restart, hard-refresh (F5). Install rgthree-comfy alongside it - the pack's nested sampler menus depend on it (check "Auto Nest Subdirectories" in rgthree settings).

Where people get tripped up

The recurring one: shift seems to do nothing. RES4LYF's bong_tangent scheduler applies its own sigma shift and ignores the workflow's, so if you've set a shift and see no effect, change the scheduler before you blame anything else - this exact trap gets flagged in every big sampler sweep.

Second, sampler_mode. If you're doing normal generation and left it on unsample or resample, you'll get garbage, because those modes deliberately turn off noise addition. Keep it on standard unless you're deep in an unsampling workflow. And the architecture rule stands - SharkSampler's solvers are a flow-matching advantage (Flux, Wan, Z-Image, Chroma); on SD 1.5 or an SDXL finetune, DPM++ 2M Karras is still the right call and this won't beat it.