PixAI Membership

📢 Login Issue Resolved — Earlier today, some users may have unexpectedly been logged out of their PixAI accounts, and this was due to a known ser…

View details

SDXL

【SDXL / Artist】赤倉 - Akakura

【SDXL / Artist】赤倉 - Akakura - AI Model cover image
Uploaded at
Apr 11, 2024, 12:07 PM
Uses
1.1k
Reviews
No reviews
Sampling Steps

30

CFG Scale

7

Negative

(cropped:1.3), (worst quality:1.3), (low quality:1.3), poorly drawn, low resolution, portrait, gothic, dark, moody, monochromatic, ugly, gritty, western cartoon, normal quality, lowres, low details, oversaturated, undersaturated, overexposed, underexposed, grayscale, bw, bad photo, bad photography, bad art, watermark, signature, text font, username, error, logo, words, letters, digits, autograph, trademark, name, blur, blurry, grainy, asymmetrical, poorly lit, bad shadow, draft, out of frame, cut off, censored, jpeg artifacts, out of focus, glitch, duplicate,

Permissions
  • Allow image generation & sharing
  • Allow users to download this model
  • Allow commercial use of generated images

Description

本人使用 SDXL 架构训练的第一个二次元画风模型。以下记录训练参数：Anything XL → XL-1 : lr=1e-5, batch size = 1, repeat = 1, epoch = 4, optimizer : LionXL-1 → XL-2 : lr=1e-5, batch size = 2, repeat = 2, epoch = 2, optimizer : LionXL-2 → XL-3 : lr=1（即初始值1e-6，最终稳定在3.4e-6左右）, batch size = 1, repeat = 1, epoch = 15, optimizer : DAdaptLionXL-3 → XL-4 : lr=1（一直稳定在1e-6） , batch size = 1, repeat = 1, epoch = 8, optimizer : DAdaptLion这里选择第八个 epoch 作为成品，实际 XL-4 一共有 10 个 epoch，这个模型先试传。如果效果不好我会换成别的。分辨率建议 9681536左右，建议大于10241024，采样器选择euler最佳。提示词基本上兼容sd1.5架构。由于底模原因，您可能需要添加 "masterpiece" 等质量相关提示词。欢迎尝试。PS：Prodigy 太耗显存了，开启全精度训练，还将 batch size = 1，结果还是直接给我干爆到30多个 g，只能忍痛割爱用 Lion 了（本来第三步需要用prodigy寻找最优学习率）。另外，不清楚是否需要将学习率降低至e-7继续精调，有经验的伙伴可以交流一下。

Comments

Artwork made by 【SDXL / Artist】赤倉 - Akakura

1
HomeModelsMio.2Profile