A Sigma-locked two-stage sampler with separate inputs for models for each. The general thinking was to run several steps of a base/raw model for better variation between seeds, and then finish it off with an extracted turbo lora on the second stage for both speed and possibly higher quality. It also has support for running two resolutions, so you can run the first stage at a lower, faster resolution. The neat thing is that you can choose a percentage and any amount of steps for the first and second stage and it will change over with the original noise (if not upscaling) at just the right sigma values.

I've also included a dual resolution node—select the aspect ratio and the base and final megapixels. It includes random modes covering all ratios, vertical ratios, horizontal ratios, or a constrained set (1:1, 4:5, 5:4, 2:3, 3:2, 3:4, and 4:3). The included aspect ratios are specifically tailored for Krea 2.

The three-stage sampler adds a final pass that reuses all of the stage 1 settings (including its model, steps, CFG, sampler, and scheduler). handoff_percent controls the stage 1 to stage 2 transition, while stage3_handoff_percent controls the later stage 2 to stage 3 transition. This can be helpful if you want to apply negative conditioning to the beginning and end of a generation.

Both samplers accept stage2_positive_optional and stage2_negative_optional for separate stage 2 conditioning. Unconnected inputs reuse positive and negative; stages 1 and 3 use the original conditioning. At stage 2 CFG 1, the selected negative conditioning is zeroed.

The Krea 2 Model Sampling node provides raw_dynamic, turbo_fixed, and manual modes. raw_dynamic follows Krea 2's resolution-dependent Raw schedule (0.5 at 256x256 through 1.15 at 1280x1280). turbo_fixed pins the shift to 1.15, as expected by the distilled Turbo sampling regime. ComfyUI doesn't set the shift correctly for non-turbo Krea 2, so it might be worth a shot to use this when not using the turbo model or checkpoint - like you might for stage one.

The main knob you'll want to play with is handoff_percent, which sets the point in the denoising process where stage 1 hands off to stage 2. For example, at 25%, stage 1 handles the first 25% and stage 2 handles the remaining 75%. At 0%, stage 2 performs the full generation; at 100%, stage 1 performs the full generation. There's no single right answer for where it should be set. You can also set it to 0% or 100% to only to one stage or the other. That can be a little confusing as to which one is 0% and which is 100% so there's a tooltip to help you.

Installation: Put in the custom_nodes folder or grab from ComfyUI manager.

SesquiLSR latent upscaling

Install SesquiLSR alongside this node pack and restart ComfyUI. In either sampler, select sesquilsr_wan21 under the advanced upscale_method setting and set final_width / final_height as usual. It uses SesquiLSR's bundled models/upscaler_Wan21.safetensors for Krea 2's VAE, with exact target dimensions and no VAE round-trip. This option is for Wan 2.1 VAE latents, not Flux latents.

The resize happens after stage 1 produces a clean latent estimate, before stage 2 adds fresh noise at the handoff sigma. The three-stage sampler uses the same resize; stage 3 continues at the final resolution. Existing interpolation options and defaults are unchanged.

SesquiLSR is trained for 1–2x scaling per spatial dimension (1 MP to 2 MP is about 1.41x). Larger or smaller targets are passed through but are outside that training range. An extremely early handoff still starts from an immature image estimate, so compare several handoff settings with the same seed. Learned upscaling can preserve detail better but cannot guarantee detail that the first stage has not developed.

Here's an image with a sample workflow embedded:

Krea 2 raw-to-turbo LoRA workflow