Credit to alibaba-pai for more fantastic work!

New update 2/10/26 The new Fun-distill-2602 version by alibaba-pai is correctly converted for ComfyUI. Unfortunately there are new feed forward keys in this version. These keys affect inference, while they help to improve contrast and brightness, they also prevent any step inference from converging when sigmas are <0.500. This means that many schedulers that have steps below this, like exponential, bong_tangent, will not work.

I have vigorously tested this, and initially removed all feed forward keys but the result, although otherwise clean, would appear somewhat washed out. I have selectively removed the feed forward keys from the weights, keeping only main blocks 0, 20-25 intact. All refiner blocks have also been kept intact.

This allows for all schedulers and samplers to work properly, while also maintaining the benefits of the new version.

Refer to the "Tests" folder for comparisons.

This is the lora converted for ComfyUI compatibility. Meant to be used with the Z-Image base model and its derivatives.

Best run at 8 steps with CFG 1. Avoid sageattention. Samplers that add noise per step such as LCM seem to work well.

Download them in the Files & versions tab.