NoobAI-SDXL-Vpred
Collection
9 items
•
Updated
•
3
该模型页面为 NoobAI XL 的 V 预测分支,无法在 AUTOMATIC1111 WebUI 中使用。 请通过 diffusers, ComfyUI[https://github.com/comfyanonymous/ComfyUI] 或 reForge 使用。
模型使用约 300k 图像,在 该检查点 训练而成。
训练的第一阶段,仅训练 UNet 的 OUT 层,以较低的学习率(3e-6)微调,直到 V 预测基本能够正常工作。 训练的第二阶段,恢复所有其它参数,训练一个完整的 epoch。
git checkout dev_upstream_experimental
;import torch
from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler
ckpt_path = "/path/to/model.safetensors"
pipe = StableDiffusionXLPipeline.from_single_file(
ckpt_path,
use_safetensors=True,
torch_dtype=torch.float16,
)
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
pipe.scheduler.register_to_config(
prediction_type="v_prediction",
rescale_betas_zero_snr=True,
)
pipe.enable_xformers_memory_efficient_attention()
pipe = pipe.to("cuda")
prompt = "best quality, 1boy, solo"
negative_prompt = "bad hands, worst quality, low quality, bad quality, multiple views, 4koma, comic, jpeg artifacts, monochrome, sepia, greyscale, flat color, pale color, muted color, low contrast, bad anatomy, picture frame, english text, signature, watermark, logo, patreon username, web address, artist name"
image = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
width=832,
height=1216,
num_inference_steps=28,
guidance_scale=7.0,
generator=torch.Generator().manual_seed(42),
).images[0]
image.save('image.png')