Pedro21613/PS-IMAGE-1.3-LEVE

PS-IMAGE 1.3 LEVE — pessoas + automoveis + texturas (compacta e rapida)

LoRA r16 em fp16 (11,5 MB) derivado de Pedro21613/PS-IMAGE-1.3 (23 MB fp32). Pesos identicos, metade do tamanho.

Uso rapido (12 steps, ~5x mais rapido que o original de 30 steps)

import torch
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler, AutoencoderTiny
from peft import PeftModel

BASE = "stable-diffusion-v1-5/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(BASE, torch_dtype=torch.float16, safety_checker=None)
pipe.unet = PeftModel.from_pretrained(pipe.unet, "Pedro21613/PS-IMAGE-1.3-LEVE")
pipe.unet = pipe.unet.merge_and_unload()
pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taesd", torch_dtype=torch.float16)
pipe.scheduler = DPMSolverMultistepScheduler.from_pretrained(BASE, subfolder="scheduler", use_karras_sigmas=True)
pipe.to("cuda")

img = pipe("ultra realistic photo of a man's face, weathered skin, visible pores, "
           "detailed skin texture, natural light, portrait, high quality",
           num_inference_steps=12, guidance_scale=6.0, height=512, width=512).images[0]

Troque Pedro21613/PS-IMAGE-1.3-LEVE pelo id deste repo. Teste interativo: gradio_colab.py (Gradio, pronto p/ Colab). Script local: usar_leve.py.