Minimalist Flat Color Illustration Style版本No Caption Training (ID: 1079826)

Yet another minimalist flat color illustration style.

Weight of 1 seems to work fine, but you can decrease it if you see some deformation.

Several versions are provided. All seems to work well, but are somewhat different. You can run tests and draw your own conclusions about "No caption" vs "Florecence 2 Captioning", and "Network Rank 8" vs "Network Rank 4".

All the trained Epochs can be found here:

tensor. art/models/797945833707468705/Minimalist-NoCap-5Cos-D4A2-R20-2024-11-19-19:43:06-Ep-10

tensor. art/models/797930006752993394/Minimalist-V3-4Cos-D8A4-R22E10-2024-11-19-07:37:31-Ep-10

tensor. art/models/796965859609410879/Minimalist-V2-4Cos-D4A2-R22E10-2024-11-17-05:25:23-Ep-10

Training parameters:

FLUX.1 - dev-fp8

Trigger: i11ustmin Illustration

Repeat: 20 Epoch: 10, 16 synthetic 1024x1024 images

Unet LR: 0.0005 LR Scheduler: cosine Optimizer: AdamW

Network Dim: 4 Alpha 2 or Network Dim: 8 Alpha 4

描述:

Same dataset, but this time it was trained without any caption (I just want to see how well "no captioning" works.)

The trigger has been change from "i11ustmin" to "i11ustmin illustration".

U-Net LR was changed from 0.0004 to 0.005.

Steps per Epoch is 20

训练词语: i11ustmin illustration