mirror of
https://github.com/modelscope/DiffSynth-Studio.git
synced 2026-04-14 13:48:20 +00:00
38 lines
1.3 KiB
Python
38 lines
1.3 KiB
Python
from diffsynth.pipelines.ernie_image import ErnieImagePipeline, ModelConfig
|
|
import torch
|
|
|
|
vram_config = {
|
|
"offload_dtype": torch.bfloat16,
|
|
"offload_device": "cpu",
|
|
"onload_dtype": torch.bfloat16,
|
|
"onload_device": "cpu",
|
|
"preparing_dtype": torch.bfloat16,
|
|
"preparing_device": "cuda",
|
|
"computation_dtype": torch.bfloat16,
|
|
"computation_device": "cuda",
|
|
}
|
|
|
|
pipe = ErnieImagePipeline.from_pretrained(
|
|
torch_dtype=torch.bfloat16,
|
|
device='cuda',
|
|
model_configs=[
|
|
ModelConfig(model_id="PaddlePaddle/ERNIE-Image-Turbo", origin_file_pattern="transformer/diffusion_pytorch_model*.safetensors", **vram_config),
|
|
ModelConfig(model_id="PaddlePaddle/ERNIE-Image", origin_file_pattern="text_encoder/model.safetensors", **vram_config),
|
|
ModelConfig(model_id="PaddlePaddle/ERNIE-Image", origin_file_pattern="vae/diffusion_pytorch_model.safetensors", **vram_config),
|
|
],
|
|
tokenizer_config=ModelConfig(model_id="PaddlePaddle/ERNIE-Image", origin_file_pattern="tokenizer/"),
|
|
vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024 ** 3) - 0.5,
|
|
)
|
|
|
|
image = pipe(
|
|
prompt="一只黑白相间的中华田园犬",
|
|
negative_prompt="",
|
|
height=1024,
|
|
width=1024,
|
|
seed=42,
|
|
num_inference_steps=8,
|
|
cfg_scale=1.0,
|
|
sigma_shift=4.0,
|
|
)
|
|
image.save("output_turbo.jpg")
|