//
Search
๐ŸŒ‰

[23CVPR] BBDM: Image-to-image Translation with Brownian Bridge Diffusion Models

Diffusion Series
Image-to-image Translation

BBDM: Image-to-image Translation with Brownian Bridge Diffusion Models

Bo Li, Kaitao Xue, Bin Liu (Nanchang Hangkong University, CN) [paper][code]

Intro & Overview

Prelinamaries: DDPMs

DDPM์— ๋Œ€ํ•œ ๊ฐ„๋žตํ•œ ์š”์•ฝ์„ ํ•˜๋ฉฐ ์‹œ์ž‘ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.
DDPM์˜ Forward Process์™€ Reverse process ๊ทธ๋ฆฌ๊ณ  Training Objective๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ •์˜๋ฉ๋‹ˆ๋‹ค.

DDPM: Forward Process

โ€ข
Forward Process๋Š” Data์— ๋…ธ์ด์ฆˆ๋ฅผ ์ ์ง„์ ์œผ๋กœ ๋”ํ•ด๊ฐ€๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค.
โ€ข
Data x0โˆผqdata(x0)\boldsymbol{x}_0 \sim q_{d a t a}\left(\boldsymbol{x}_0\right) ์—์„œ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค.
โ€ข
์ตœ์ข… Latent variable xT\boldsymbol x_T์ด๋ฉฐ platentย (xT)=N(0,I)p_{\text {latent }}\left(\boldsymbol{x}_T\right)=\mathcal{N}(\mathbf{0}, \boldsymbol{I})์ธ isotropic Gaussian distribution์ž…๋‹ˆ๋‹ค.
โ€ข
x0\boldsymbol x_0์—์„œ xT\boldsymbol x_T๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ๊ณผ์ •์€ Markov chain์œผ๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค.
q(x1,โ€ฆ,xTโˆฃx0)=โˆt=1Tq(xtโˆฃxtโˆ’1)\begin{equation} q\left(\boldsymbol{x}_1, \ldots, \boldsymbol{x}_T \mid \boldsymbol{x}_0\right)=\prod_{t=1}^T q\left(\boldsymbol{x}_t \mid \boldsymbol{x}_{t-1}\right) \end{equation}
โ€ข
ฮฒt\beta_t๋Š” ์•„์ฃผ ์ž‘์€ ์ƒ์ˆ˜.

DDPM: Reverse Process.

โ€ข
Reverse Process๋Š” Latent variable xT\boldsymbol x_T๋กœ๋ถ€ํ„ฐ ๋ฐ์ดํ„ฐ x0\boldsymbol x_0๋ฅผ ์ถ”๋ก ํ•˜๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค.
โ€ข
์ด๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ Markov chain์œผ๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค.
pฮธ(x0,โ€ฆ,xTโˆ’1โˆฃxT)=โˆt=1Tpฮธ(xtโˆ’1โˆฃxt)p_\theta\left(\boldsymbol{x}_0, \ldots, \boldsymbol{x}_{T-1} \mid \boldsymbol{x}_T\right)=\prod_{t=1}^T p_\theta\left(\boldsymbol{x}_{t-1} \mid \boldsymbol{x}_t\right)

DDPM: Training Objective

โ€ข
DDPM์€ Evidence Lower Bound (ELBO)๋ฅผ ์ด์šฉํ•˜์—ฌ ์ตœ์ ํ™”๋ฉ๋‹ˆ๋‹ค.
Ex0,ฯตโˆฅฯตโˆ’ฯตฮธ(xt,t)โˆฅ22\begin{equation}\mathbb{E}_{\boldsymbol{x}_0, \boldsymbol{\epsilon}}\left\|\boldsymbol{\epsilon}-\boldsymbol{\epsilon}_\theta\left(\boldsymbol{x}_t, t\right)\right\|_2^2\end{equation}
โ€ข
ฯต\boldsymbol{\epsilon}: xt\boldsymbol x_t ์˜ ๊ฐ€์šฐ์‹œ์•ˆ ๋…ธ์ด์ฆˆ. โˆ‡xtlnโกq(xtโˆฃx0)\nabla_{\boldsymbol{x}_t} \ln q\left(\boldsymbol{x}_t \mid \boldsymbol{x}_0\right).
โ€ข
ฯตฮธ\boldsymbol{\epsilon}_\theta: denoising model.

Conditional Diffusion.

โ€ข
๋Œ€๋ถ€๋ถ„์˜ conditional diffusion model๋“ค์€ condition์„ Objective์— ์ง์ ‘์ ์œผ๋กœ ์ฃผ์ž…ํ•ฉ๋‹ˆ๋‹ค.
Ex0,ฯตโˆฅฯตโˆ’ฯตฮธ(xt,y,t)โˆฅ22\begin{equation} \mathbb{E}_{\boldsymbol{x}_0, \boldsymbol{\epsilon}}\left\|\boldsymbol{\epsilon}-\boldsymbol{\epsilon}_\theta\left(\boldsymbol{x}_t, \boldsymbol{y}, t\right)\right\|_2^2 \end{equation}
โ€ข
ํ•˜์ง€๋งŒ, p(xtโˆฃy)p\left(\boldsymbol{x}_t \mid y\right)๊ฐ€ ์ด Objective์— ๋ช…ํ™•ํžˆ ์ •์˜๋˜์ง€ ์•Š์œผ๋ฏ€๋กœ ์ƒ์„ฑ ๊ฒฐ๊ณผ๊ฐ€ condition์— ๋ถ€ํ•ฉํ•˜๋Š”์ง€ ๋‹จ์–ธํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.

Prelinamaries: Brownian Bridge

โ€ข
Brownian bridge ๋ชจ๋ธ์€ continous-time stochastic ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.
โ€ข
์ด๋•Œ condtion์€ start point์™€ ending point์— ์˜ํ•ด ์ •ํ•ด์ง‘๋‹ˆ๋‹ค.
โ€ข
์ฆ‰, ๋…ธ์ด์ฆˆ์—์„œ ๋ฐ์ดํ„ฐ๋กœ ๋ณ€ํ–ˆ๋˜ ๊ธฐ์กด ๋ชจ๋ธ๊ณผ ๋‹ฌ๋ฆฌ ๋ฐ์ดํ„ฐ์—์„œ ๋ฐ์ดํ„ฐ๋กœ ๋ณ€ํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.
โ€ข
Brownian Bridge Process๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ •์˜๋ฉ๋‹ˆ๋‹ค.
p(xtโˆฃx0,xT)=N((1โˆ’tT)x0+tTxT,t(Tโˆ’t)TI)\begin{equation}p\left(\boldsymbol{x}_t \mid \boldsymbol{x}_0, \boldsymbol{x}_T\right)=\mathcal{N}\left(\left(1-\frac{t}{T}\right) \boldsymbol{x}_0+\frac{t}{T} \boldsymbol{x}_T, \frac{t(T-t)}{T} \boldsymbol{I}\right)\end{equation}
โ€ข
์‹์€ ๊ฐ„๋‹จํ•ฉ๋‹ˆ๋‹ค. T์— ๋Œ€ํ•œ t์˜ ๋น„์œจ๋กœ x0\boldsymbol x_0์™€ xT\boldsymbol x_T๋ฅผ Linearํ•˜๊ฒŒ ์„ž์–ด์ค๋‹ˆ๋‹ค.
โ€ข
๋…ธ์ด์ฆˆ๋Š” ์–‘ ๊ทน๋‹จ์—์„œ ์ตœ์†Œ๊ฐ€ ๋˜๊ณ , ์ค‘๊ฐ„์—์„œ ์ตœ๋Œ€๊ฐ€ ๋ฉ๋‹ˆ๋‹ค.

Methodology

Brownian Bridge Diffusin Model ์œ„ Brownian Bridge Process๋ฅผ ์ ์šฉํ•˜๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
qBB(xtโˆฃx0,y)=N(xt;(1โˆ’mt)x0+mty,ฮดtI)x0=x,mt=tT\begin{equation}\begin{gathered}q_{B B}\left(\boldsymbol{x}_t \mid \boldsymbol{x}_0, \boldsymbol{y}\right)=\mathcal{N}\left(\boldsymbol{x}_t ;\left(1-m_t\right) \boldsymbol{x}_0+m_t \boldsymbol{y}, \delta_t \boldsymbol{I}\right) \\\boldsymbol{x}_0=\boldsymbol{x}, \quad m_t=\frac{t}{T}\end{gathered}\end{equation}
โ€ข
์ด๋•Œ, ์‹ (4)์— ๋”ฐ๋ผ variance๋Š” ฮดt=t(Tโˆ’t)T\delta_t=\frac{t(T-t)}{T}์ธ๋ฐ t=12Tt=\frac{1}{2}T ์ผ๋•Œ ฮด12t=T4\delta_{\frac{1}{2}t}=\frac{T}{4} ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค.
โ€ข
์ด๋Š” T๊ฐ€ ๋งค์šฐ ์ปค์ง์— ๋”ฐ๋ผ ๋”์šฑ ํฐ ๊ฐ’์ด ๋ฉ๋‹ˆ๋‹ค.
โ€ข
๋”ฐ๋ผ์„œ variance๋ฅผ ๋ณด์กด (Preserving) ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋””์ž์ธํ•ฉ๋‹ˆ๋‹ค.
ฮดt=2s(mtโˆ’mt2)\begin{equation}\delta_t=2 s\left(m_t-m_t^2\right)\end{equation}
โ€ข
์ด๋•Œ s๋Š” scale factor ๋กœ์จ sampling diversity๋ฅผ ์ปจํŠธ๋กคํ•ฉ๋‹ˆ๋‹ค. (s=1s=1 by default)
Forward Process: qBB(xtโˆฃxtโˆ’1,y)q_{B B}\left(\boldsymbol{x}_t \mid \boldsymbol{x}_{t-1}, \boldsymbol{y}\right)
โ€ข
์œ„ ์‹ (5) ๋ฅผ ํ†ตํ•ด xt,xtโˆ’1\boldsymbol{x}_t,\boldsymbol{x}_{t-1}์„ ๊ตฌํ•˜์—ฌ ๋Œ€์ž…ํ•ด์ฃผ๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
qBB(xtโˆฃxtโˆ’1,y)=N(xt;1โˆ’mt1โˆ’mtโˆ’1xtโˆ’1+(mtโˆ’1โˆ’mt1โˆ’mtโˆ’1mtโˆ’1)y,ฮดtโˆฃtโˆ’1I)\begin{equation} \begin{aligned} q_{B B}\left(\boldsymbol{x}_t \mid \boldsymbol{x}_{t-1}, \boldsymbol{y}\right)=&\mathcal{N}\left(\boldsymbol{x}_t ; \frac{1-m_t}{1-m_{t-1}} \boldsymbol{x}_{t-1}\right. \\ & \left.\quad+\left(m_t-\frac{1-m_t}{1-m_{t-1}} m_{t-1}\right) \boldsymbol{y}, \delta_{t \mid t-1} \boldsymbol{I}\right) \end{aligned} \end{equation}
ฮดtโˆฃtโˆ’1=ฮดtโˆ’ฮดtโˆ’1(1โˆ’mt)2(1โˆ’mtโˆ’1)2\begin{equation} \delta_{t \mid t-1}=\delta_t-\delta_{t-1} \frac{\left(1-m_t\right)^2}{\left(1-m_{t-1}\right)^2} \end{equation}
Reverse Process xT=yx_T=y ์—์„œ ์ถœ๋ฐœํ•ด x0x_0์— ๋„๋‹ฌํ•˜๋Š” reverse process๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
pฮธ(xtโˆ’1โˆฃxt,y)=N(xtโˆ’1;ฮผฮธ(xt,t),ฮด~tI)\begin{equation}p_\theta\left(\boldsymbol{x}_{t-1} \mid \boldsymbol{x}_t, \boldsymbol{y}\right)=\mathcal{N}\left(\boldsymbol{x}_{t-1} ; \boldsymbol{\mu}_\theta\left(\boldsymbol{x}_t, t\right), \tilde{\delta}_t \boldsymbol{I}\right)\end{equation}
โ€ข
์ด๋•Œ, ฮผฮธ(xt,t)\boldsymbol{\mu}_\theta(\boldsymbol{x}_t,t): predicted mean value of noise ฮด~t\tilde{\delta}_t: variance of noise at each step.
Objective ELBO๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
ELBO=โˆ’Eq(DKL(qBB(xTโˆฃx0,y)โˆฅp(xTโˆฃy))+โˆ‘t=2TDKL(qBB(xtโˆ’1โˆฃxt,x0,y)โˆฅpฮธ(xtโˆ’1โˆฃxt,y))โˆ’logโกpฮธ(x0โˆฃx1,y))\begin{equation}\begin{aligned}E L B O & =-\mathbb{E}_q\left(D_{K L}\left(q_{B B}\left(\boldsymbol{x}_T \mid \boldsymbol{x}_0, \boldsymbol{y}\right) \| p\left(\boldsymbol{x}_T \mid \boldsymbol{y}\right)\right)\right. \\& +\sum_{t=2}^T D_{K L}\left(q_{B B}\left(\boldsymbol{x}_{t-1} \mid \boldsymbol{x}_t, \boldsymbol{x}_0, \boldsymbol{y}\right) \| p_\theta\left(\boldsymbol{x}_{t-1} \mid \boldsymbol{x}_t, \boldsymbol{y}\right)\right) \\& \left.-\log p_\theta\left(\boldsymbol{x}_0 \mid \boldsymbol{x}_1, \boldsymbol{y}\right)\right)\end{aligned}\end{equation}
โ€ข
์ฒซ ํ•ญ์€ constant๊ฐ€ ๋˜์–ด KL Divergence๋Š” 0์ด ๋ฉ๋‹ˆ๋‹ค.
โ€ข
๋‘๋ฒˆ์งธ ํ•ญ์€ ์‹ (5)์™€ (7) ๊ทธ๋ฆฌ๊ณ  Bayesโ€™ theorem, Markov chain property๋ฅผ ์ด์šฉํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
qBB(xtโˆ’1โˆฃxt,x0,y)=qBB(xtโˆฃxtโˆ’1,y)qBB(xtโˆ’1โˆฃx0,y)qBB(xtโˆฃx0,y)=N(xtโˆ’1;ฮผ~t(xt,x0,y),ฮด~tI)\begin{equation}\begin{aligned}q_{B B}\left(\boldsymbol{x}_{t-1} \mid \boldsymbol{x}_t, \boldsymbol{x}_0, \boldsymbol{y}\right) & =\frac{q_{B B}\left(\boldsymbol{x}_t \mid \boldsymbol{x}_{t-1}, \boldsymbol{y}\right) q_{B B}\left(\boldsymbol{x}_{t-1} \mid \boldsymbol{x}_0, \boldsymbol{y}\right)}{q_{B B}\left(\boldsymbol{x}_t \mid \boldsymbol{x}_0, \boldsymbol{y}\right)} \\& =\mathcal{N}\left(\boldsymbol{x}_{t-1} ; \tilde{\boldsymbol{\mu}}_t\left(\boldsymbol{x}_t, \boldsymbol{x}_0, \boldsymbol{y}\right), \tilde{\delta}_t \boldsymbol{I}\right)\end{aligned}\end{equation}
ฮผ~t(xt,x0,y)=ฮดtโˆ’1ฮดt1โˆ’mt1โˆ’mtโˆ’1xt+(1โˆ’mtโˆ’1)ฮดtโˆฃtโˆ’1ฮดtx0+(mtโˆ’1โˆ’mt1โˆ’mt1โˆ’mtโˆ’1ฮดtโˆ’1ฮดt)y\begin{equation}\begin{aligned}\tilde{\boldsymbol{\mu}}_t\left(\boldsymbol{x}_t, \boldsymbol{x}_0, \boldsymbol{y}\right) & =\frac{\delta_{t-1}}{\delta_t} \frac{1-m_t}{1-m_{t-1}} \boldsymbol{x}_t \\& +\left(1-m_{t-1}\right) \frac{\delta_{t \mid t-1}}{\delta_t} \boldsymbol{x}_0 \\& +\left(m_{t-1}-m_t \frac{1-m_t}{1-m_{t-1}} \frac{\delta_{t-1}}{\delta_t}\right) \boldsymbol{y}\end{aligned}\end{equation}
ฮด~t=ฮดtโˆฃtโˆ’1โ‹…ฮดtโˆ’1ฮดt\begin{equation} \tilde{\delta}_t=\frac{\delta_{t \mid t-1} \cdot \delta_{t-1}}{\delta_t} \end{equation}
โ€ข
์ด๋•Œ, x0\boldsymbol{x}_0๊ฐ€ unknown์ด๋ฏ€๋กœ reparametrization trick์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
ฮผ~t(xt,y)=cxtxt+cyty+cฯตt(mt(yโˆ’x0)+ฮดtฯต)\begin{equation}\tilde{\boldsymbol{\mu}}_t\left(\boldsymbol{x}_t, \boldsymbol{y}\right)=c_{x t} \boldsymbol{x}_t+c_{y t} \boldsymbol{y}+c_{\epsilon t}\left(m_t\left(\boldsymbol{y}-\boldsymbol{x}_0\right)+\sqrt{\delta_t} \boldsymbol{\epsilon}\right)\end{equation}
cxt=ฮดtโˆ’1ฮดt1โˆ’mt1โˆ’mtโˆ’1+ฮดtโˆฃtโˆ’1ฮดt(1โˆ’mtโˆ’1)cyt=mtโˆ’1โˆ’mt1โˆ’mt1โˆ’mtโˆ’1ฮดtโˆ’1ฮดtcฯตt=(1โˆ’mtโˆ’1)ฮดtโˆฃtโˆ’1ฮดt\begin{aligned}c_{x t} & =\frac{\delta_{t-1}}{\delta_t} \frac{1-m_t}{1-m_{t-1}}+\frac{\delta_{t \mid t-1}}{\delta_t}\left(1-m_{t-1}\right) \\c_{y t} & =m_{t-1}-m_t \frac{1-m_t}{1-m_{t-1}} \frac{\delta_{t-1}}{\delta_t} \\c_{\epsilon t} & =\left(1-m_{t-1}\right) \frac{\delta_{t \mid t-1}}{\delta_t}\end{aligned}
โ€ข
์—ฌ๊ธฐ์„œ ฮผ~ฮธ\tilde{\mu}_\theta๊ฐ€ ์•„๋‹Œ ๋…ธ์ด์ฆˆ ฯตฮธ\epsilon_\theta๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ํ•˜๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์€ Linear combindation์ด ๋ฉ๋‹ˆ๋‹ค.
ฮผฮธ(xt,y,t)=cxtxt+cyty+cฯตtฯตฮธ(xt,t)\begin{equation}\boldsymbol{\mu}_{\boldsymbol{\theta}}\left(\boldsymbol{x}_t, \boldsymbol{y}, t\right)=c_{x t} \boldsymbol{x}_t+c_{y t} \boldsymbol{y}+c_{\epsilon t} \boldsymbol{\epsilon}_\theta\left(\boldsymbol{x}_t, t\right) \end{equation}
โ€ข
๋”ฐ๋ผ์„œ, ELBO (์‹ 10)์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.
Ex0,y,ฯต[cฯตtโˆฅmt(yโˆ’x0)+ฮดtฯตโˆ’ฯตฮธ(xt,t)โˆฅ2]\begin{equation}\mathbb{E}_{\boldsymbol{x}_0, \boldsymbol{y}, \boldsymbol{\epsilon}}\left[c_{\epsilon t}\left\|m_t\left(\boldsymbol{y}-\boldsymbol{x}_0\right)+\sqrt{\delta_t} \boldsymbol{\epsilon}-\boldsymbol{\epsilon}_\theta\left(\boldsymbol{x}_t, t\right)\right\|^2\right]\end{equation}
์ˆ˜์‹์œผ๋กœ ๋ณด๋‚˜ ์ฝ”๋“œ๋กœ ๋ณด๋‚˜ ์‚ฌ์‹ค ๋ณ„๊ฑฐ ์—†์Šต๋‹ˆ๋‹ค.
mt(yโˆ’x0)m_t(y-x_0) ์ƒ˜ํ”Œ์— ๋…ธ์ด์ฆˆ๋ฅผ ๋”ํ•œ ํ›„ (target) L2 ๋กœ์Šค๋ฅผ ๊ณ„์‚ฐํ•˜๋ฉด ๋์ž…๋‹ˆ๋‹ค.
Sampling
qBB(xฯ„sโˆ’1โˆฃxฯ„s,x0,y)=N((1โˆ’mฯ„sโˆ’1)x0+mฯ„sโˆ’1y+ฮดฯ„sโˆ’1โˆ’ฯƒฯ„s21ฮดฯ„s(xฯ„sโˆ’(1โˆ’mฯ„s)x0โˆ’mฯ„sy),ฯƒฯ„s2I)\begin{equation}\begin{array}{r}q_{B B}\left(\boldsymbol{x}_{\tau_{s-1}} \mid \boldsymbol{x}_{\tau s}, \boldsymbol{x}_0, \boldsymbol{y}\right)=\mathcal{N}\left(\left(1-m_{\tau_{s-1}}\right) \boldsymbol{x}_0+m_{\tau_{s-1}} \boldsymbol{y}+\right. \\\left.\sqrt{\delta_{\tau_{s-1}}-\sigma_{\tau_s}^2} \frac{1}{\sqrt{\delta_{\tau_s}}}\left(\boldsymbol{x}_{\tau_s}-\left(1-m_{\tau_s}\right) \boldsymbol{x}_0-m_{\tau_s} \boldsymbol{y}\right), \sigma_{\tau_s}^2 \boldsymbol{I}\right)\end{array}\end{equation}
์‚ฌ์‹ค ํฐ ์˜๋ฏธ๋Š” ์—†์Šต๋‹ˆ๋‹ค. ์ƒ˜ํ”Œ๋Ÿฌ step = 200 ์— ๋Œ€ํ•ด DDPM ์ƒ˜ํ”Œ๋ง์„ ํ•˜๊ฒ ๋‹ค๋Š” ๋œป์ž…๋‹ˆ๋‹ค.

Experimental Results

Conclusion

โ€ข
๊ต‰์žฅํžˆ ๊ฐ„๋‹จํ•œ ๋…ผ๋ฌธ์ด์ง€๋งŒ, ๊ธฐ์กด์˜ Bridge Model๊ณผ ๋‹ค๋ฅด๊ฒŒ ๊ฐ„๋‹จํ•œ ์ˆ˜์‹๋งŒ์œผ๋กœ DDPM์˜ VP ํ”„๋ ˆ์ž„์›Œํฌ๋กœ ๋ช…๋ฃŒํ™”ํ–ˆ๋‹ค๋Š” ์ ์—์„œ ๋‹ค๋ฅธ Application์œผ๋กœ ๋ฐœํŒ์„ ๋งˆ๋ จํ–ˆ๋‹ค๊ณ  ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
โ€ข
์ตœ๊ทผ์˜ ๋งŽ์€ Architectural ํ•œ ๋…ผ๋ฌธ๋“ค์€ ๊ฑฐ์ง„ DDPM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๊ณ  ์žˆ๋Š” ๊ฒƒ ์ฒ˜๋Ÿผ์š”.
โ€ข
Bridge Diffusion Model์€ Condition์˜ ๊ฐœ์ž… ์—†์ด๋„ Image-to-image task๋ฅผ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์—์„œ ๋ชจ๋ธ ๋ณต์žก๋„, ํ›ˆ๋ จ ์‹œ๊ฐ„, ์ฐจ์› ๋ณต์žก๋„์—์„œ ํฐ ์ด๋“์ด ์žˆ์Šต๋‹ˆ๋‹ค. (๋…ผ๋ฌธ์—๋Š” ์–ธ๊ธ‰๋˜์ง€ ์•Š์•˜์ง€๋งŒ ์ด๋Ÿฌํ•œ ์ด์ ์ด Bridge Model์˜ ๊ฐ€์žฅ ํฐ ์žฅ์ ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.)
โ€ข
์ด๋ฅผ ๋ฒ ์ด์Šค๋กœ ํ•œ ์ €์˜ ๋…ผ๋ฌธ์ด ๊ณง publish ๋  ์˜ˆ์ •์ด๋‹ˆ ๊ธฐ๋Œ€ํ•ด์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌ๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค.