SD3_watermeloncarving

This is the test version of SD3,

training on diffusers allows it to train Lora with a minimum of 20GB of memory on a 1024+1024 resolution-based instantiation

The model is trained on a bs1 with 1024+1024 basis, with 20GB memory inside, using random sampling to increase generalization

version: v1

*Parameters do not represent the best, for personal use only

Sampling method(Sampler): euler normal

Magnification algorithm: 4x-UltraSharp | 4x_NMKD-Siax_200k

Sampling iteration steps(Steps): 25-30

Resampling magnitude(Denoising): 0.1

Width 768, Height 1024

Hint word relevance(CFG Scale): 4

*Make sure comfyUI is the latest version, otherwise SD3's Lora will not take effect