🐸 KSampler

Frog KSampler is the sampler that finishes the job. A stock KSampler hands you a latent and expects a separate VAE Decode node to turn it into a picture. This one does the sampling and the decode in a single node, so it spits out both the raw latent and a decoded Image ready for a preview or save. That's one node fewer in your graph, and on a workflow full of Frog nodes, it's the one that closes the loop.

It also carries the pack's signature trick: a beta57 scheduler. beta57 is a tuned variant of ComfyUI's beta schedule - this pack implements it with alpha=0.5 and beta=0.7 - and it's become the community's go-to for painterly Anima output. The pack injects it straight into ComfyUI's scheduler list at import time, so it shows up as beta57 in the dropdown alongside karras, simple, and the rest. If you're on Anima, that's the scheduler you want for the kind of soft, brushy result the model is known for; er_sde + beta57 is the pairing the KB's Anima guide recommends.

The inputs that matter

The familiar sampling controls are all here - seed, steps, cfg, sampler_name, scheduler, denoise - plus two extras worth knowing:

• Tiled - a boolean that turns on tiled VAE decoding. This is your high-resolution friend: at 2K+ outputs a single decode pass can blow past VRAM, and tiling decodes in chunks so the image still comes out whole. Default off.
• tile_size - controls the tile dimensions when tiled decode is active. Default 512; raise it if you have VRAM to spare and want fewer seams.

The model/conditioning inputs (model, positive, negative, vae) are all optional because there's a basic_pipe input that bundles them. The pipe is compatible with Impact Pack's ToBasicPipe and the pack's own 🐸 Pipe In, so you can carry the whole model stack down one wire and never think about it again. Wire basic_pipe and the individual inputs are ignored - which is usually what you want.

What comes out

Three outputs: latent (the raw sampled latent, if you want to do something else with it), Image (decoded, ready to preview or save), and debug - a string carrying seed, shape, and pixel value range. That debug output is genuinely handy for diagnosing washed-out or clipped results without leaving the canvas.

Installing it

Pack install, same as the rest. ComfyUI Manager → "Frog Node Pack", or:

cd ComfyUI/custom_nodes
git clone https://github.com/RabbitThatIsPink/FrogNodePack

Restart and hard-refresh. No extra dependencies. One gotcha specific to this node: because beta57 is injected when the pack loads, you won't see it in the scheduler dropdown until ComfyUI has been restarted after install. And if you're a RES4LYF user, you'll notice both packs offer a beta57 - that's a shared community convention, and the implementation here is the pack's own. If you see weird sampler behavior with both packs installed, that's the place to look.

Where people trip

The classic error is feeding this node a plain latent from EmptyLatentImage plus a denoise of 1.0 for img2img - that's a blank-canvas generation, not a re-imagining. Set denoise below 1.0 when you're feeding it an existing image's latent, and remember the decode respects your Tiled toggle: leave it on for big resolutions, off for normal ones, because tiling costs a little quality in exchange for fitting in memory.