☁️BizyAir SamplerCustomAdvanced

What it is

This is the cloud twin of stock ComfyUI's SamplerCustomAdvanced - the node you land on when a single KSampler dropdown isn't enough control and you want noise, guidance, sampler, and schedule as four separate, swappable pieces. If you're already running a custom sampling graph locally (Noise → Guider → Sampler → Scheduler, all feeding into this), this node is the same idea, just executing on BizyAir's hardware against a BIZYAIR_MODEL.

You'd reach for it instead of a plain KSampler-equivalent when you need to mix and match - a specific guider (like a CFG guider with unusual settings), a specific sampler algorithm, and a scheduler built independently, rather than the fixed combination a simpler node bakes in.

How it works

It's the terminal node in a custom sampling chain: it takes an already-constructed noise generator, guider, sampler, and sigma schedule, and runs the actual denoising loop, applying each sampling step according to those four pieces. Everything upstream of it (which sampler algorithm, which scheduler shape, how CFG is applied) is decided by separate nodes feeding into it - this one just executes the loop and returns the result.

Inputs and outputs that matter

All five inputs are required and all are typed sockets rather than dropdowns or numbers, which tells you this node exists purely to assemble other nodes' outputs: noise (a NOISE source, usually a random-noise generator node), guider (a GUIDER, which wraps your model plus conditioning and CFG behavior), sampler (a SAMPLER, the actual step-taking algorithm - Euler, DPM++, etc.), sigmas (a SIGMAS curve - this pack's BizyAir_BasicScheduler is the obvious source), and latent_image (the starting LATENT, either pure noise for txt2img or a partially-noised image for img2img-style work).

Two outputs: output, the final sampled latent, and denoised_output, the model's fully-denoised prediction at that point in the process - useful if you want to preview or inspect what the model thinks the finished image looks like independent of however many steps remain. Both are LATENT and both need decoding through a VAE before they're viewable images.

Installing it

Search "BizyAir" in ComfyUI Manager and install, or clone manually with git clone https://github.com/siliconflow/BizyAir.git into custom_nodes, then restart ComfyUI. This pack is dead without a BizyAir API key - set one up through the login prompt that appears the first time you run any BizyAir node.

Common issues

Because every input here is a typed socket with no defaults, this node is unforgiving about mismatches - a NOISE output from a stock ComfyUI node won't plug into this one, and a sigmas curve from a non-BizyAir scheduler won't either. If you're getting nothing but red wires, you're likely missing one of the four upstream BizyAir nodes this one depends on (a noise generator, a guider, a sampler, and a scheduler), rather than a problem with this node itself. Beyond that, sampler/scheduler mismatches are the usual source of bad output - a Karras-style schedule on a flow-matching model like Flux is a common way to get oversaturated, broken results, and that's a modeling choice made upstream, not something this node can fix.