Model-Aware Sigma Scheduler

The Model-Aware Sigma Scheduler is the version of Sigmax's Krea 2 scheduler that actually looks at the model. Where Krea2SigmaScheduler takes your word for which variant you're running, this node validates a Krea 2 MODEL directly, resolves the exact profile from what's really loaded, and emits capability-gated sigmas. Same family, same recipes - but the model is the source of truth instead of your memory.

This matters more than it sounds, because Krea 2's RAW and Turbo checkpoints are easily confused and the filename is not trustworthy. Loading the wrong variant's schedule onto a model is the fastest way to turn a genuinely good model into garbage, and this node is the pack's answer to that failure mode.

Inputs

• model - the MODEL object from your loader. This is the whole point of the node: it probes the loaded model's sampling class and capabilities to resolve which profile applies.
• variant - Auto, Turbo, or RAW. Auto tries to resolve from the model evidence. Here's the honest caveat the README gives: Auto may identify only the family, not the exact variant, so an ambiguous Auto result should be resolved manually. When in doubt, pick Turbo or RAW explicitly - the README's troubleshooting table calls this out as the fix for "Auto rejects Krea 2."
• steps - default 8 (the Turbo count). RAW wants 28 or 52.
• width / height - actual output dimensions; RAW's schedule is resolution-aware, so these matter.
• strict_official - default true, keep it for official recipes.
• start_step / end_step - slice.

Outputs: sigmas and schedule_info. Feed sigmas into your custom-sampling path; don't add a second scheduler or time shift.

When to reach for it vs the plain scheduler

If your workflow is locked to a known variant, the plain Krea2SigmaScheduler is simpler and there's nothing wrong with it. This node earns its keep when you're mixing models - swapping between RAW and Turbo checkpoints in one graph, or sharing a workflow where the model might be either. Then "let the model decide" beats "hope the person loading it picked the right dropdown." Just keep the Auto-is-advisory caveat in your head.

Installing

Same one-pack install:

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

Restart, search Sigmax. ComfyUI Manager has "ComfyUI-Sigmax". Python 3.10+, ComfyUI 0.29.0+, no extra dependencies, no model downloads.

Where people get burned

Two recurring ones. First, trusting Auto too hard: if the model evidence is ambiguous it resolves to the family, not the variant, and you can silently end up on the wrong recipe - read schedule_info before generating. Second, forgetting this node is schedule-only: it validates and builds sigmas, it doesn't load the checkpoint, choose a sampler, or fix a wrong CFG. If the output's still bad with the right variant resolved, the problem is downstream of the schedule.