ClownsharKSampler

If you're running Flux, Z-Image, Chroma, Wan, or any of the newer flow-matching models and the stock KSampler dropdown feels like it stopped keeping up, this is the node people quietly swap in. ClownsharKSampler is the flagship of RES4LYF, a pack of high-order solvers and reshaped noise schedules built specifically for rectified-flow and probability-flow models - the architectures that made the old "DPM++ 2M Karras and forget it" advice actively wrong. It's genuinely mainstream now, not a power-user curiosity: it turns up credited in high-scoring Qwen-Image, Wan, and Anima workflow shares, and out-mentions the old default sampler in community discussion by roughly four to one.

The name is a portmanteau - it's the "Clown" sampler (the solver picker) and the "Shark" sampler (the thing that actually runs) fused into one node so you don't have to wire them together. That's the whole pitch: all the control, none of the plumbing.

Why you'd reach for it

Flow-matching models train on a near-straight path from noise to image, so there's much less for a clever schedule to fix and a lot more for it to break. Karras and exponential don't just underperform on these models, they fall apart. RES4LYF exists to fill that gap with solvers that suit straight trajectories, and ClownsharKSampler is where you use them without leaving one node. If you're on SD 1.5 or an SDXL finetune (Illustrious, Pony, NoobAI), skip it - those are DDPM-style, DPM++ 2M Karras is still correct, and this buys you nothing.

How it works, and the inputs that matter

Under the hood it's a Runge-Kutta solver zoo, but you only touch a few knobs:

• sampler_name - the solver. Default is res_2m, and honestly that's the one to leave it on: it's fast (one model call per step, Euler speed) and works on nearly everything. Samplers ending in s (res_2s, res_3s) take substeps, so they cost more model calls and run slower, but the accuracy jump can be dramatic. Names ending in m are the cheap multistep versions.
• scheduler - defaults to beta57, RES4LYF's own beta schedule (alpha 0.5, beta 0.7) and the single most-borrowed thing in the pack. It's a great starting point for flow-matching models.
• bongmath - leave this on (it's the default). It aligns each substep's latents with the noise predictions as it goes, so denoising runs forwards and backwards at once - a more accurate result for no extra VRAM and no extra time. There's no reason to turn it off.
• eta (default 0.5) controls how much noise gets re-injected each step in the SDE variants; cfg, steps, denoise, and seed behave like their KSampler equivalents.
• sampler_mode - keep it standard for normal generation. Switch to unsample / resample only if you're doing noise-inversion img2img.

The author's own recommended starting point, straight from the threads: 20-30 steps, res_2m or res_2s, eta 0.5, bongmath on, scheduler beta57.

Two optional inputs are worth knowing about. sigmas lets an external scheduler node override the built-in steps/scheduler. guides accepts RES4LYF's composition/style guide images - closer to img2img-with-high-denoise, great for steering layout and color. On the output side you get output (the finished LATENT, straight into your VAE Decode), denoised (the clean prediction, handy for previews), and options (an OPTIONS object you can chain into other RES4LYF nodes).

Installing it

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

One strongly-recommended companion: install rgthree-comfy too. The nested sampler menu that makes res_2m and friends browsable relies on rgthree's "Auto Nest Subdirectories" setting being on. Without it the solver list is a lot less usable.

Where people get burned

• The shift knob "does nothing." Some schedulers - bong_tangent especially - apply their own sigma shift and ignore the workflow's. If changing shift has no effect, change your scheduler before assuming anything's broken.
• You copied settings that name beta57 or res_2m and the workflow won't load. Those are RES4LYF names; the pack has to be installed or ComfyUI can't find the node.
• 69 solvers, no map. The pack gives you enormous choice and very little guidance on which suits what, and the honest community answer is still "it depends on the image." Don't sweep all of them - start at res_2m + beta57 and only reach for the slower s samplers when you're chasing the last few percent.
• Running a hosted service? The license is AGPL-3.0 with a rider that prohibits commercial image-generation services from using it without a separate license. That's aimed at services, not at you generating your own images - but worth knowing if you're building on top of it.