Here are the results from tests run on raw, unquantized weights.

No compression, no compromisesjust pure density and rendering depth.

These images are "straight off the assembly line," with no post-processing in Capture One. The entire pipeline and logic are baked directly into the preview workflow.

Generation parameters:

Resolution: 2048 × 2048 (4 true megapixels)

Grid step: 128

Steps: 65

CFG: 3.5

Sampler: UniPC

Scheduler: SGM Uniform

Training:

The model was trained on a custom dataset (768×768) using proprietary memory management and training optimization code, written from scratch.

Want to train at the same speed with zero OOM (Out of Memory) errors? Head over to my GitHubthe toolkit is already available in the repository.

In a few hours, I’ll be releasing an update that adds support for new training modes. Stay tuned for more drops, and let me know in the comments which styles or datasets you’d like to see me run next!

https://github.com/OrakulStudio/AI-Toolkit-Windows11