Sign In

Moss Image FinalCut (5GB - 6Step)

474

Updated: Oct 23, 2026

base modelmossmoss image

Download

2 variants available

Type
Checkpoint Trained
Stats

99

Reviews
Published

Oct 22, 2026

Base Model

MOSS Image

Training
Steps: 50,000
Epochs: 50
Hash
AutoV2
1630F46BED
Krea2_00001_.png

Moss Image FinalCut

I highly recommend Universal FP8 - Quality is 2x of INT8 and it works very well with loras. NVFP4 will not work with loras currently.

  • FinalCut was trained to increase performance at low steps with reduced weights.

  • 4 Step is possible however less anatomical errors appear at 6 steps

  • CFG 2-3 (If using comfy bypass) with Euler or Euler-A

  • If using onsite DO NOT SELECT TURBO LORA and USE CFG 1 (Onsite is ER_SDE)

  • Capable of native 1920x1280 or variant - 2048x2048 is possible but may repeat.

INT8

  • Reduce steps to 4 use SGM_Uniform/ER_SDE

  • INT8 suffers from canvas issue if paired with INT8 QWEN

  • INT8 works with lora's however universal works far better

  • Quality no longer scales at steps beyond 4 vs FP8 model


I did mess up the metadata and it throws a comfy error but this is not the cause of the canvas issue with convrot - it is not the turbo or the convrot size