Image-to-Image Translation with Diffusion

Module: Generative AI | Difficulty: Advanced

SDEdit (Img2Img)

1. Add noise to input image:
2. Denoise from to

Strength Parameter

• Low strength: preserve input structure
• High strength: creative variation

DDIM Inversion

Invert a real image to noise, then denoise with different conditioning.

def img2img(model, image, text, strength=0.75, steps=50):
    scheduler.set_timesteps(steps)
    init_timestep = int(steps * strength)
    t_start = steps - init_timestep
    noise = torch.randn_like(image)
    t = scheduler.timesteps[t_start]
    latent = scheduler.add_noise(vae.encode(image), noise, t)
    for t in scheduler.timesteps[t_start:]:
        latent = scheduler.step(model, latent, t, text)
    return vae.decode(latent)

Research Insight: SDEdit's quality depends heavily on the noise level. Too little noise preserves artifacts; too much noise loses the input structure. The optimal strength is typically 0.5-0.8 for style transfer.