🔧 Flux Sampler Parameters

This is the node you use to stop guessing your Flux settings and actually test them. Flux Sampler Parameters runs a whole grid of sampler/scheduler/steps/guidance combinations in one go and hands the results to a plot node so you can eyeball which settings win. If you've ever wondered whether guidance 3.5 is really right for your model or whether euler beats deis, this is how you find out instead of arguing about it.

Why you'd reach for it

Flux settings are contentious. The community default of guidance 3.5 is widely called too high - "the fact 3.5 is the default is why so many people struggle" is a real refrain - and the right sampler/scheduler pairing genuinely depends on the checkpoint and the LoRAs you've stacked. Testing that by hand means dozens of manual runs and squinting at outputs. Flux Sampler Parameters automates the sweep: give it lists of values, it generates every combination, and you compare them on a labeled grid. It's the empirical answer to "what settings should I use," and it's a signature node of cubiq's pack (cubiq = Matteo Spinelli, the IPAdapter author, whose Latent Vision channel is where a lot of this Flux-tuning knowledge got demonstrated).

How it works

The clever trick: almost every settings field is a string, not a number, so you can put a comma-separated list in it. Type euler, deis into sampler and 3.0, 3.5, 4.0 into guidance and the node runs all six combinations. Leave a field as a single value to pin it. seed accepts ? for random. It batches the whole matrix into one execution and outputs both the resulting latents and a params bundle describing each run - that params output is meant to feed the pack's companion Plot Parameters node, which lays the results out as an annotated XY grid so each cell is tagged with the settings that produced it.

Inputs and outputs that matter

You wire in the usual model, conditioning, and latent_image. The fields you sweep are sampler, scheduler, steps, guidance, and denoise, plus max_shift / base_shift (Flux's timestep-shift knobs - leave blank to use defaults). There's an optional loras input (LORA_PARAMS) so you can even sweep across LoRAs. Outputs are latent (the batch of results) and params (SAMPLER_PARAMS) for the plot node. In practice you connect params → Plot Parameters → Preview and read the grid.

Installing it

ComfyUI Essentials, via Manager (search ComfyUI Essentials, restart) or:

cd ComfyUI/custom_nodes
git clone https://github.com/cubiq/ComfyUI_essentials

then restart. It's 🔧 Flux Sampler Parameters under essentials/sampling, class FluxSamplerParams+. Maintenance-only pack since April 2025 - but this node was already the mature way to benchmark Flux.

Common issues

The big one is memory. Because it batches an entire matrix of generations, it's far heavier than a single KSampler - one Mac mini M4 user hit repeated "MPS backend out of memory" errors and had to restart between runs. Keep your grids small (sweep one axis at a time), and on tight VRAM add --disable-smart-memory or reduce how many combinations you request. Two more: this node is Flux-specific - don't point it at SDXL and expect sense - and if your plot comes out unlabeled, make sure the params output is actually wired into Plot Parameters, not just the latents into a decode.