Scribd
Upload
4 views2 pages

Diffusion Model Assignment: Q7 and Q8: Q7: Blade Runner 2049 - Reconstructing Memory Through Noise

The document discusses the diffusion model assignment focusing on reconstructing memory through noise in 'Blade Runner 2049' and conditional diffusion for Ghibli-style portrait stylization. It explains the forward diffusion process, the role of noise schedules, and the training of neural networks for denoising. Additionally, it highlights the advantages of diffusion models over GANs, including stable training and better fidelity and diversity in generated outputs.

Uploaded by

Monster
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
4 views2 pages

Diffusion Model Assignment: Q7 and Q8: Q7: Blade Runner 2049 - Reconstructing Memory Through Noise

The document discusses the diffusion model assignment focusing on reconstructing memory through noise in 'Blade Runner 2049' and conditional diffusion for Ghibli-style portrait stylization. It explains the forward diffusion process, the role of noise schedules, and the training of neural networks for denoising. Additionally, it highlights the advantages of diffusion models over GANs, including stable training and better fidelity and diversity in generated outputs.

Uploaded by

Monster
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Diffusion Model Assignment: Q7 and Q8

Q7: Blade Runner 2049 – Reconstructing Memory Through Noise


(a) Closed-form Forward Process q(xt | x0 )
We define the forward diffusion process as:
p p
xt = 1 − βt xt−1 + βt ϵt−1 , ϵt−1 ∼ N (0, I)

Let:
t
Y
αt = 1 − β t , ᾱt = αi
i=1

Then: √ √
xt = ᾱt x0 + 1 − ᾱt ϵ, ϵ ∼ N (0, I)
So: √
q(xt | x0 ) = N (xt ; ᾱt x0 , (1 − ᾱt )I)

(b) Role of βt and Noise Schedules


The choice of βt controls the noise schedule. Two common schedules:

• Linear Schedule: βt increases linearly with t. This results in fast early corruption and lower
sample quality.

• Cosine Schedule: A nonlinear schedule where ᾱt follows a cosine function. This maintains
more signal in early steps and leads to higher quality generations.

(c) Reverse Denoising and Learned Mean µθ


We train a neural network ϵθ (xt , t) to predict noise added to xt . The true posterior mean (if x0
were known) is: √ √
ᾱt−1 βt αt (1 − ᾱt−1 )
µ̃(xt , x0 ) = x0 + xt
1 − ᾱt 1 − ᾱt
But since x0 is unknown, we estimate it via:
√
xt − 1 − ᾱt ϵθ (xt , t)
x̂0 = √
ᾱt

Plugging this into the posterior, we approximate:


 
1 βt
µθ (xt , t) = √ xt − √ ϵθ (xt , t)
αt 1 − ᾱt

1
Q8: Conditional DDPM for Ghibli-style Portrait Stylization
(a) Conditional Diffusion Setup
We aim to learn a conditional model:
p(x | y)
where x is a Ghibli-style portrait and y is a real-world portrait.
We define a conditional denoising model:

ϵθ (xt , t | y)

The generation process samples xT ∼ N (0, I), then denoises via:

xt−1 ∼ N (µθ (xt , t | y), Σt )

(b) Denoising Objective


At training time, we minimize the denoising loss:

Lt = Ex0 ,y,ϵ ∥ϵ − ϵθ (xt , t | y)∥2


 

where √ √
xt = ᾱt x0 + 1 − ᾱt ϵ, ϵ ∼ N (0, I)
The full training objective:

L = Et,x0 ,y,ϵ ∥ϵ − ϵθ (xt , t | y)∥2


 

(c) Advantages of DDPMs over GANs


DDPMs outperform GANs in key areas:

• Stable Training: No adversarial game.

• No Mode Collapse: DDPMs model the full data distribution.

• High Fidelity & Diversity: Multi-step denoising enables fine detail.

• Easier Conditioning: Incorporating y is straightforward.

In contrast, GANs suffer from training instability, harder conditioning, and mode collapse.
DDPMs are thus preferred for tasks like Ghibli-style stylization.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy