NetaYume v35 Flashpack
Inference
from diffusers import AutoencoderKL, Lumina2Pipeline, Lumina2Transformer2DModel
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
from flashpack import FlashPackMixin
from flashpack.integrations.diffusers import FlashPackDiffusionPipeline
from flashpack.integrations.diffusers.model import FlashPackDiffusersModelMixin
from flashpack.integrations.transformers import FlashPackTransformersModelMixin
import torch
from transformers import Gemma2Model
class TransformerModel(Lumina2Transformer2DModel, FlashPackDiffusersModelMixin):
pass
class TextEncoder(Gemma2Model, FlashPackTransformersModelMixin):
pass
class Lumina2FlashpackPipeline(Lumina2Pipeline, FlashPackDiffusionPipeline):
def __init__(self, transformer: TransformerModel, scheduler: FlowMatchEulerDiscreteScheduler, vae: AutoencoderKL, text_encoder: TextEncoder, tokenizer):
super().__init__(transformer, scheduler, vae, text_encoder, tokenizer)
if __name__ == '__main__':
model_path = '/path/to/netayume-v35-flashpack'
text_encoder = TextEncoder.from_pretrained_flashpack(model_path, subfolder='text_encoder', torch_dtype=torch.bfloat16)
transformer = TransformerModel.from_pretrained_flashpack(model_path, subfolder='transformer', torch_dtype=torch.bfloat16)
pipeline = Lumina2FlashpackPipeline.from_pretrained_flashpack(
model_path,
text_encoder=text_encoder,
transformer=transformer,
torch_dtype=torch.bfloat16
)
pipeline.enable_model_cpu_offload()
image = pipeline(
'prompt',
system_prompt='You are an assistant designed to generate anime images based on textual prompts.',
num_inference_steps=40,
generator=torch.Generator().manual_seed(0)
cfg_trunc_ratio=6,
cfg_normalization=False
).images[0]
image.save('preview.png')
References
