Krea 2 Sigma Scheduler

Krea 2 is the 12B diffusion transformer that pulled an entire ecosystem off Ideogram 4 in a week, and it's also a poster child for why generic schedulers fail on flow-matching models. RAW is an undistilled base with no aesthetic finetune; Turbo is guidance- and step-distilled. They need completely different sigma schedules, and Turbo's resolution-dependence means even the "right" generic scheduler often builds the wrong curve. The Krea 2 Sigma Scheduler is Sigmax's flagship: explicit RAW/Turbo selection, resolution-aware sigmas, and pinned recipes - plus an experimental RAW-to-Turbo LoRA lane for the crowd that runs RAW with the Turbo LoRA at reduced weight (which, as the Krea ecosystem converged on, often looks better than Turbo alone).

The variants

• Turbo - 8 steps, CFG 1.0. Guidance is baked into the weights, so CFG 1 isn't a low setting, it's the setting. This is the fast lane.
• RAW - the undistilled base. Official 52-step recipe at CFG 4.5, or the framework-reference 28-step recipe at CFG 5.5. RAW is the one to train on and the one that wants your actual dimensions.
• LoRA Experimental (RAW mu) and LoRA Experimental (Turbo mu) - schedule modes for a RAW model with a compatible RAW-to-Turbo LoRA applied. Community starting point per the README: 12 steps, Euler, CFG 1.0. These force strict_official off and construct schedule mu values only - the node does not load or scale any LoRA. You load the LoRA yourself in the workflow.

Inputs that matter

• variant - the four above. The pack's rule is explicit selection; Auto lives on the sibling ModelAwareSigmaScheduler instead.
• steps - default 8, which matches Turbo. RAW wants 28 or 52.
• width / height - the actual output dimensions. These matter because RAW's schedule is resolution-aware; the node normalizes RAW to its supported grid, and the README's troubleshooting table explicitly says to check requested vs effective dimensions in schedule_info if your RAW output looks off.
• strict_official - default true. For the LoRA experimental lanes it gets forced off; keep it on for real RAW/Turbo.
• start_step / end_step - slice.

Outputs: sigmas (into your custom-sampling path) and schedule_info (recipe, effective dimensions, warnings).

Installing

Standard pack install:

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

Restart, search Sigmax. ComfyUI Manager also has "ComfyUI-Sigmax". Python 3.10+, ComfyUI 0.29.0+, no extra deps, no model downloads. You still need the Krea 2 checkpoint, its Qwen3-VL encoder, and the Qwen-Image VAE - the node only supplies the schedule.

Where people get burned

Three things, in order of frequency. One: forgetting the "no second scheduler" rule - feed sigmas straight to the sampler, or you'll shift twice. Two: running RAW at Turbo's 8 steps because the default fooled you; read the variant and step count together. Three: assuming the LoRA lanes load the LoRA - they don't, by design, so a "LoRA experimental" schedule with no LoRA loaded just gives you a weird RAW result. And if Auto on the model-aware sibling rejects your Krea 2, pick the variant explicitly; the filename alone isn't trustworthy.