BongSampler

BongSampler is the RES4LYF sampler with the lid on. Its big sibling, ClownsharKSampler, throws 69 solvers, implicit-step controls, unsampling modes, and a wall of noise settings at you. BongSampler strips all of that away and gives you a node shaped almost exactly like ComfyUI's stock KSampler - same sockets, same handful of widgets - but with the RES family of samplers and the pack's schedulers baked in. If you looked at ClownsharKSampler and backed away slowly, this is the one to use.

It comes from RES4LYF (by ClownsharkBatwing), the pack that became where sampler tuning actually happens for flow-matching models like Flux, Z-Image, Chroma, and Wan. The whole reason the pack exists is that those architectures train on a near-straight noise-to-image path, which broke the old scheduler shortlist - Karras and exponential go from "safe default" to "actively wrong." BongSampler is the lowest-friction way to get the pack's flow-matching-friendly sampling without learning the rest of it.

How it works

The name is the giveaway. "Bong" is bongmath, the pack's signature trick, and here it's baked in rather than exposed as a switch. What bongmath does, in the author's own words: it aligns each substep's latents with the noise predictions as sampling proceeds, effectively running the denoising process forwards and backwards at once. The payoff is a more accurate result - better images, better video - for no extra VRAM and no added inference time. It's the kind of thing you'd leave on permanently, so BongSampler just does that for you.

Everything else is a Runge-Kutta solver picking its way from noise to your image, the same math the rest of the pack uses, wrapped in a KSampler-familiar face.

The inputs and outputs that matter

You wire it up exactly like a KSampler: feed in the model (required), your positive and negative conditioning, and a latent_image to sample into. The widgets you'll actually touch:

• sampler_name - the solver, and here it's a clean list of eight: res_2m, res_3m, res_2s, res_3s, and an _sde variant of each. The default is res_2s_sde. If you want the everyday workhorse instead, res_2m is fast (Euler speed, one model call per step) and works on almost everything; the s samplers take substeps, so they're slower but more accurate; the _sde versions re-inject noise each step for a bit more texture.
• scheduler - defaults to beta57, RES4LYF's own beta schedule (alpha 0.5, beta 0.7) and the single most-borrowed component in the pack. The list also includes the stock ComfyUI schedulers plus bong_tangent, so you're not locked in.
• steps, cfg, seed, denoise - behave like their KSampler equivalents. The author's suggested starting point for these RES samplers is around 20-30 steps.

The output is a single LATENT - send it straight to your VAE Decode. That's it. No options output to chain, no denoised preview tap; if you want those, that's what the bigger nodes are for.

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 run a venv - cd RES4LYF && pip install -r requirements.txt (portable ComfyUI users point at the embedded pip instead). Restart ComfyUI and hard-refresh with F5.

There are no model files to download; it's a sampler node, not a checkpoint. Because BongSampler's sampler list is a flat dropdown rather than the nested menu the flagship uses, you don't even strictly need rgthree-comfy here - though it's still worth having for everything else.

Where people get burned

• Wrong model family. BongSampler's advantage is specific to flow-matching models. On SD 1.5 or an SDXL finetune (Illustrious, Pony, NoobAI), DPM++ 2M Karras is still the right call and this won't help.
• The scheduler ignores your shift. bong_tangent applies its own sigma shift and disregards the one set elsewhere in your workflow. If a shift adjustment seems to do nothing, check whether you're on that scheduler before blaming anything.
• "Where did beta57 come from?" If you're copying someone's settings that name beta57 or res_2m, note those ship inside this pack - the workflow won't load them unless RES4LYF is installed.
• Commercial use. The pack is AGPL-3.0 with a rider forbidding paid image-generation services from using it without a separate license. Generating your own images is fine; building a product on it is the part to read carefully.