일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | ||||||
2 | 3 | 4 | 5 | 6 | 7 | 8 |
9 | 10 | 11 | 12 | 13 | 14 | 15 |
16 | 17 | 18 | 19 | 20 | 21 | 22 |
23 | 24 | 25 | 26 | 27 | 28 | 29 |
30 | 31 |
- 오블완
- NLP
- GAN
- cv
- 폰트생성
- animation retargeting
- RNN
- CNN
- Few-shot generation
- BERT
- WBP
- ddpm
- Generative Model
- 디퓨전모델
- 생성모델
- Font Generation
- deep learning
- UE5
- motion matching
- 언리얼엔진
- Stat110
- userwidget
- 모션매칭
- ue5.4
- Diffusion
- multimodal
- dl
- WinAPI
- 딥러닝
- Unreal Engine
- Today
- Total
목록Generative Model (7)
Deeper Learning

prerequisite: Diffusion Model에 대한 기초적인 이해 Diffusion Model $T$ = 전체 Timesteps 수 $x_T$ = forward process를 $T$ 번 적용한 마지막 Timestep $T$ 에서의 이미지 $x_0$ = 원본 이미지 Forward Process $$ q(x_t|x_{t-1}) = N(x_t, \sqrt{1-\beta_t}x_{t-1}, \beta_tI) $$ $\beta$ 는 noise(variance)의 강도를 조절하는 parameter로 DDPM 논문에서는 0.0001 ~ 0.02의 값을 사용 $\beta$ 가 선형적으로 timesteps에 따라(DDPM에서는 linear noise scheduler 사용) 0.0001에서 0.02까지 증가하..

Sebastian Nowozin, Botond Cseke, Ryota Tomioka. (2016) Abstract Generative neural sampler는 feedforward NN을 통해 샘플링하는 probabilistic model sampling과 도함수를 쉽게 계산할 수 있지만 likelihood, marginalization을 계산할 수 없음 generative-adversarial approach가 더 일반적인 variational divergence estimation approach의 special case임을 논문에서 제시 모든 f-divergence를 generative neural sampler를 학습시킬 때 사용할 수 있음 다양한 divergence function을 사용하는..

Mehdi Mirza, Simon Osindero.(2014.11) Abstract Generative Adversarial Nets(이하 GAN)은 최근(2014.06) 생성 모델을 학습하기 위한 새로운 방식으로 제시되었음. $y$ label을 generator와 discriminator에게 주어 GAN이 conditional 생성을 가능하게 함. MNIST를 class label에 따라 생성, multi-modal model로 training label이 없는 image에 대해 tagging 결과를 수록 1. Introduction GAN은 다루기 힘든 여러 확률적 계산을 근사하기 위한 생성모델 학습 프레임워크로 최근에 제시되었음. Markov chain을 사용하지 않고 역전파로만 gradient를 ..

Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, Timo Aila NVIDIA. (2019.12) Abstract StyleGAN은 unconditional data-driven 이미지 생성 SOTA 달성 모델 StyleGAN의 특징적인 아티팩트를 분석하고 이를 해결하기 위한 아키텍처와 학습방법을 제안 generator의 normalization을 새로 디자인, progressive growing을 재고(resolution을 증가시키며 학습하는 것), generator regularization을 통해 latent code에서 image로의 매핑을 개선 path length regularizer는 이미지 퀄리티를 개선할..

Chenhao Li, Yuta Taniguchi, Min Lu, Shin’ichi Konomi HDI Lab, Kyushu University. (2021) Abstract 적은 샘플로 언어 간 font style을 transfer 할 수 있는 FTransGAN을 제시 언어 간 font style transfer는 많이 연구되지 않았음 multi-level attention을 통해 local, global style을 capture English, Chinese를 모두 가진 847개의 font dataset을 사용 Introduction 폰트를 만드는 것은 매우 노동집약적, 대부분 하나의 언어에 대한 폰트만 만듦 폰트의 subset만 보고 나머지를 생성하는 것은 neural net으로 가능해졌으며 관련 ..

ABSTRACT few-shot generation(FFG)는 2가지 조건을 만족하여야 한다. target char의 global structure 보존 diverse local reference style의 표현 explicit 하게 component label을 주지 않고 MX-Font는 multiple style features를 추출한다. multiple expert들이 different local concepts을 표현한다. 각 expert들이 different local style에 특화될 수 있도록 weak supervision방식으로 component label을 이용한다. 각 expert들에게 component assign problem을 graph matching problem으로 정의..

Abstract 기존 Few-shot font generation 모델은 style과 content의 disentanglement을 하고자 하였지만 이와 같은 접근법은 다양한 "local styles"을 표현하는데 한계가 있다 이는 복잡한 letter system(ex. Chinese)에 적합하지 않다 localized styles = component-wise style representations을 학습하는 방식으로 font generation하는 모델을 제안 reference glyph의 수를 줄이기위해 component-wise style을 component factor와 style factor의 product으로 정의한다. strong locality supervision (component의 ..