FLUX.1-schnell Sigma Scheduler

FLUX.1-schnell is the distilled fast FLUX - guidance baked into the weights, CFG 1.0, and 1–4 steps instead of the dev model's 20–30. The catch is that when you run it through ComfyUI's normal scheduler, the "scheduler + steps" logic assumes a model that wants a long, reshaped denoising journey. Schnell does not. It wants a short, unshifted, exactly-pinned schedule. The FLUX.1-schnell Sigma Scheduler is a purpose-built node that emits precisely that.

This is one of the pack's flagship "explicit over inferred" moves. Instead of letting a generic scheduler derive sigmas and hoping, the node builds the pinned unshifted FLUX.1-schnell schedule with the profile's recipe, fingerprints it, and hands it to you as a SIGMAS tensor. The profile's guidance declaration says it plainly: host_cfg 1.0, model_guidance 0.0. Guidance is off because it's already baked in - that's the whole schnell story.

Inputs

• steps - default 4, and the official recipe lives in the 1–4 step range. If you're cranking it to 30 because that's what you're used to, you're fighting the model.
• strict_official - default true, and you should generally leave it there. It pins the schedule to the verified official recipe. Turning it off produces a modified-schedule variant flagged in schedule_info; useful for experiments, wrong for reproducibility.
• start_step / end_step - slice the schedule to a sub-range.

Outputs: sigmas (feed into a custom-sampling path that accepts external sigmas) and schedule_info (JSON with the recipe, fingerprints, and warnings).

The scope warning, read it

The pack is very specific: this node is for the schnell family only, and it does not imply anything about other FLUX variants. FLUX.1-dev, FLUX.2, and the rest have different guidance regimes and different schedule needs - schnell's unshifted 1–4 step schedule is not a universal FLUX cheat code. If you drop it on dev and wonder why everything's mushy, that's on you, not the node.

Installing

It ships with the whole Sigmax pack:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-Sigmax comfyui-sigmax

Restart ComfyUI and search Sigmax in the node menu; ComfyUI Manager also has "ComfyUI-Sigmax". Python 3.10+, ComfyUI 0.29.0+, no extra dependencies, no model downloads - the node builds numbers, not weights, so you still need your normal schnell checkpoint/VAE path.

Common gotchas

The pack-wide rule applies hard here: connect sigmas directly to the sampler and don't chain a BasicScheduler or another time shift after it, or you'll apply a schedule twice and get exactly the kind of garbage a 4-step model produces when its noise levels are wrong. And if your output looks off at 4 steps, check the sampler before the schedule - schnell is happiest with robust low-step samplers (Euler is the usual recommendation), not ancestral ones that re-noise every step.