SoteMix V3 is trained at 1024x1536 for high resolution image generations.

Also available at Hugging Face:

OpenVINO and ONNX are supported too.

https://huggingface.co/Disty0/SoteMixV3

This model is tested on SD.Next with Diffusers backend and HyperTile size set to 0 (Auto).

Positive Prompts:

masterpiece, best quality, highres, 1girl,

Negative Prompts:

(worst quality, low quality, lowres), zombie, interlocked fingers, water,

Do not use any negative embeddings.

Sampler: Euler a

Steps: 30-40

Clip Skip: 1 or 2

CFG: 4-7

Base Resolution: 512x / 768x / 1024x / 768x1280 / 960x1280 / 1024x1536 / 1920x1080

Model can still be chaotic at 1024x1536 and 1920x1080.

Second Pass / Hires:

Sampler: Euler / Euler a

Steps: 10 with Euler / 20 with Euler a

Upscaler: RealESRGAN 4x+ Anime6B / ESRGAN 4x-AnimeSharp with 0.2-0.3 denoise strength.

CFG: 6-11

Resolution: 2x of the base resolution.

Training:

My GPU couldn't handle full model training at these resolutions so i trained it as a 512 layer Lora with SoteMix V1 as the base.

Used highres as the trigger word. Also used raifu trigger word with my OC character.

Resolution: 1024x1536 with Bucketing

Batch Size: 1

GPU: Intel ARC A770 16GB

Bucket:

bucket 0: resolution (832, 1664), count: 49
bucket 1: resolution (896, 1280), count: 1
bucket 2: resolution (896, 1536), count: 2
bucket 3: resolution (960, 1408), count: 8
bucket 4: resolution (960, 1472), count: 57
bucket 5: resolution (960, 1536), count: 12
bucket 6: resolution (960, 1600), count: 537
bucket 7: resolution (1024, 1344), count: 266
bucket 8: resolution (1024, 1408), count: 349
bucket 9: resolution (1024, 1472), count: 1535
bucket 10: resolution (1024, 1536), count: 950
bucket 11: resolution (1088, 1280), count: 63
bucket 12: resolution (1152, 1216), count: 62
bucket 13: resolution (1152, 1280), count: 147
bucket 14: resolution (1152, 1344), count: 114
bucket 15: resolution (1216, 1152), count: 44
bucket 16: resolution (1216, 1216), count: 409
bucket 17: resolution (1216, 1280), count: 53
bucket 18: resolution (1280, 576), count: 20
bucket 19: resolution (1280, 640), count: 94
bucket 20: resolution (1280, 704), count: 217
bucket 21: resolution (1280, 768), count: 102
bucket 22: resolution (1280, 832), count: 118
bucket 23: resolution (1280, 896), count: 280
bucket 24: resolution (1280, 960), count: 137
bucket 25: resolution (1280, 1024), count: 32
bucket 26: resolution (1280, 1088), count: 27
bucket 27: resolution (1280, 1152), count: 61
bucket 28: resolution (1280, 1216), count: 24
bucket 29: resolution (1344, 1024), count: 17
bucket 30: resolution (1344, 1152), count: 38
bucket 31: resolution (1536, 896), count: 94
bucket 32: resolution (1536, 1024), count: 34
bucket 33: resolution (1600, 960), count: 196
bucket 34: resolution (1664, 832), count: 21
bucket 35: resolution (2048, 768), count: 3
bucket 36: resolution (2304, 576), count: 1
mean ar error (without repeats): 0.01257769833438255

Merge:

Merged SoteMix V1 with Lunar Radiance Light and then merged the Hires Lora i trained on top of it.

Merge ratio: (0.6 SoteMix V1 + 0.4 Lunar Radiance Light) + 0.7 Hires Lora

SoteMix V1:

A Stable Diffusion 1.5 anime model.

Also available at Hugging Face:

https://huggingface.co/Disty0/SoteMix

LCM version is available here:

https://huggingface.co/Disty0/LCM_SoteMix

Mege:

0.7(0.75(0.75(0.75(0.75(0.75(0.5(SomethingV2_2) + 0.5(meinapastel_v5Anime)) + 0.25(meinaalter_v2)) + 0.25(meinamix_meinaV9)) + 0.25(nyanMix_230303Intense)) + 0.25(mixProV4_v4)) + 0.25(breakdomain_M2150)) + 0.3(0.4(SomethingV2_2) + 0.6(meinapastel_v5Anime))

And baked in Blessed2 VAE