Moss Image FinalCut (12 - 25Step)
Use at 12-25 steps sgm_uniform/er_sde (Beyond 25 quality tends to taper off)
Up to 2048x2048
Onsite generations look drastically worse then local comfy output
Comfy Quants
NVFP4 has canvas issues and doesn't work with Lora's
INT8 has canvas issues both are inferior to universal formats
This is likely a quant issue as INT8 should outperform FP8
Trained with both 4096 caption length and short WD14 tags - English language reinforcement and Pony style training.