Sigma testing for Flux2Klein
New sigma schedules produce more stable shifts and fewer artifacts in the distilled Flux2Klein model.
A community researcher known as Capitan01R has published optimized sigma (noise) schedules specifically for the Flux2Klein image generation model. This work addresses a key technical challenge: Flux2Klein is a distilled version of the larger Flux2 Dev model, meaning it behaves differently and the default Flux2Scheduler is not ideal. The researcher's testing revealed that custom sigma values are required to achieve stable image edits without introducing artifacts or distortions. The findings include specific, step-by-step sigma schedules for workflows using 4, 6, 8, 10, 12, and 15 generation steps, with the 10-step configuration highlighted as most ideal for regular use.
These sigma schedules are designed to be used with the Klein edit scheduler (a bug-fixed version is provided) and the Euler sampling method. The primary benefit is more controlled and predictable image transformations, or "shifts," during the editing process. The researcher also provides tips for parametric mode, suggesting adjustments to sigma minimum, denoise, shift, and curve parameters depending on the step count. This optimization work is crucial for users looking to leverage the speed of the distilled Flux2Klein model while maintaining high-quality, artifact-free output, effectively bridging the gap between model efficiency and creative control.
- Custom sigma schedules are essential for the distilled Flux2Klein model, as it behaves differently than the base Flux2 Dev.
- The optimized 10-step sigma schedule (1.0000, 0.9997, 0.9994, 0.9900, 0.9818, 0.9200, 0.45, 0.44, 0.43, 0.0513, 0.0000) is recommended for regular use.
- The settings, used with the Klein edit scheduler and Euler sampler, reduce final-step movement and artifacts for more stable image edits.
Why It Matters
Enables stable, high-quality image editing with efficient distilled AI models, improving workflow for digital artists and content creators.