ClownSampler

Here's the thing that confuses people about ClownSampler: it doesn't sample anything. Drop it in expecting a KSampler and you'll stare at a node with no model input, no steps, no CFG, and no image coming out the other end. That's not a bug. ClownSampler is a sampler definition - it packages up a solver and hands it to something else to actually run. Its only output is a SAMPLER object, and its whole job is to feed a node that has a SAMPLER input.

It's the "Clown" half of RES4LYF's flagship ClownsharKSampler. That all-in-one node fuses the sampler picker (Clown) with the thing that runs it (Shark). ClownSampler splits the picker back out for people who want to build their sampling graph by hand - wiring the solver into ComfyUI's own SamplerCustom / SamplerCustomAdvanced, into RES4LYF's SharkSampler, or into any custom setup that expects a SAMPLER. If the all-in-one feels like a black box and you want your scheduler, sigmas, and guides as separate, visible nodes, this is the modular route.

Why it exists

RES4LYF (by ClownsharkBatwing) is the pack that became where sampler tuning happens in 2026 for flow-matching models - Flux, Z-Image, Chroma, Wan, and friends. Those architectures denoise along a near-straight path, which is exactly what breaks the old Karras-based defaults and what the pack's high-order solvers are built for. Most people meet that capability through ClownsharKSampler. ClownSampler is for when you've outgrown the convenience node and want the solver as a first-class component you can reroute, reuse, and compose - the same reason ComfyUI splits out KSamplerSelect from the KSampler you plug it into.

The inputs and outputs that matter

Because it only builds a sampler, its widget list is short and specific:

• sampler_name - the solver, default res_2m. This is the one decision the node really makes. res_2m is the everyday pick: one model call per step (Euler speed) and it works on nearly everything. Solvers ending in s (res_2s, res_3s) take substeps for more accuracy at a real speed cost; the ones ending in m are the cheap multistep versions.
• bongmath - leave it on (the default). It aligns each substep's latents with the noise predictions as it goes, so denoising effectively runs forwards and backwards at once - a more accurate result for no extra VRAM and no extra time. It's a "set and forget" quality boost.
• eta (default 0.5) sets how much noise the SDE-mode solvers re-inject each step; seed seeds that noise.

Note what's not here: no scheduler, no steps, no CFG, no model. Those don't belong on a sampler definition - they live on the node you plug this into, which is the whole point of the split. There's one optional input worth knowing: guides, which accepts RES4LYF's composition/style guide images so the guiding travels with the sampler you're building. The single output is sampler (SAMPLER) - run a wire from it to the sampler input of your custom-sampler node.

Installing it

Two ways. In ComfyUI Manager, search RES4LYF, install, and restart. Or from a terminal: cd ComfyUI/custom_nodes && git clone https://github.com/ClownsharkBatwing/RES4LYF, then, if you use a venv, cd RES4LYF && pip install -r requirements.txt (portable ComfyUI points at its embedded pip). Restart and hard-refresh with F5.

Install rgthree-comfy alongside it. ClownSampler's sampler_name is the nested multistep/res_2m-style menu, and that browsable folder layout depends on rgthree's "Auto Nest Subdirectories" setting being on (gear icon, bottom-left, rgthree settings). Without it, picking a solver is a lot more painful.

Where people get burned

• Nothing renders and there's no error you understand. Remember this node produces a SAMPLER, not an image. It has to feed a SamplerCustom-style node that supplies the model, latent, sigmas, and steps. On its own it does nothing - that's by design.
• The shift setting seems ignored. RES4LYF's bong_tangent scheduler applies its own sigma shift and overrides the workflow's. If you've wired ClownSampler into a graph where shift does nothing, check the scheduler on the other node before assuming the sampler is at fault.
• Wrong architecture. The RES solvers are a flow-matching advantage. On SD 1.5 or an SDXL finetune (Pony, Illustrious, NoobAI), DPM++ 2M Karras is still correct and this won't help.
• Commercial use. The pack is AGPL-3.0 with a rider prohibiting paid image-generation services from using it without a separate license - a concern for building a product on it, not for generating your own images.