digiplay

/

DucHaiten-Real3D-NSFW-V1

like 19

I'm new here and I can't run the model using the default guide for safetensors.

hi do you use stable-diffusion-webui(AUTOMATIC1111) on your PC?
or use Google colab + diffusers online? or something else?

I'm new here and I can't run the model using the default guide for safetensors.

This thread is old, I know. But I'm registering here the answer because I also started yesterday too. And I had a lot of troubles and I needed to read, to code and test a lot before do something.

import os
import torch
from diffusers import DiffusionPipeline
from datetime import datetime

# Set memory optimization
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

# Free up GPU memory
torch.cuda.empty_cache()

# Define maximum memory for devices (adjust based on your hardware)
max_memory = {0: "6GiB", "cpu": "40GiB"}  # Very low GPU memory to force CPU offloading

# Initialize the pipeline without moving to a device
model_name="digiplay/DucHaiten-Real3D-NSFW-V1"
pipeline = DiffusionPipeline.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    max_memory=max_memory
)

# Se ainda assim estiver sem memória, use pipeline.enable_sequential_cpu_offload() ao invés de pipeline.enable_vae_slicing
pipeline.enable_vae_slicing()
pipeline.enable_model_cpu_offload()
pipeline.enable_attention_slicing()

NOW = datetime.now().strftime("%Y-%m-%d_%H%M%S")

Prompt="Magazine ad, iconic, 1943, Cartoon, sharp focus, comic, 2k art on canvas by kyoani and ROSSDRAWS, princess"
Prompt2="cartoon, colored manga"

Negative_prompt="extra limbs, double torso, extra arms, extra hands, mangled fingers, missing lips, extra legs"

# Se alterar de 600x800 para 720x1280 dobra o tempo de computação, mas ainda cabe na GPU
height=688
width=896
num_images_per_prompt=5

output_image_file=model_name.replace('/', '_').replace('\\', '_')

last_image_file_saved=""
for index, image in enumerate(pipeline(
        Prompt,
        prompt2=Prompt2,
        negative_prompt=Negative_prompt,
        height=height,
        width=width,
        num_images_per_prompt=num_images_per_prompt,
        add_watermarker=False,
        num_inference_steps=35).images):
    last_image_file_saved=f"{output_image_file}_{NOW}_{index}.png"
    image.save(last_image_file_saved)

print("last_image_file_saved: ", last_image_file_saved)