fly51fly (@fly51fly)
ByteDance 연구팀이 'BitDance'라는 새로운 생성 모델 아키텍처를 공개했습니다. 이 모델은 이진 토큰(binary tokens)을 사용하는 자기회귀(autoregressive) 방식으로, 확장성과 효율성을 모두 향상시킨 이미지 및 텍스트 생성 성능을 보여줍니다. 대규모 모델의 경량화와 효율적 학습을 가능하게 하는 잠재적 돌파구로 평가됩니다.
fly51fly (@fly51fly)
ByteDance 연구팀이 'BitDance'라는 새로운 생성 모델 아키텍처를 공개했습니다. 이 모델은 이진 토큰(binary tokens)을 사용하는 자기회귀(autoregressive) 방식으로, 확장성과 효율성을 모두 향상시킨 이미지 및 텍스트 생성 성능을 보여줍니다. 대규모 모델의 경량화와 효율적 학습을 가능하게 하는 잠재적 돌파구로 평가됩니다.
Matt Shumer (@mattshumer_)
작성자가 Genie 3를 매우 강력하게 칭찬하며 '지금까지 써본 것 중 가장 미친 것'이라 표현한 반응성 트윗입니다. 강한 감탄과 함께 시연(또는 영상)을 보라는 권유가 포함되어 있어 새로운 모델 체험기·리뷰 성격입니다.
Matt Shumer (@mattshumer_)
Genie 3가 포트나이트(Fortnite) 게임플레이를 생성하는 모습을 보고, 그런 동작은 의도된 기능이 아닐 것 같다는 의문을 제기하는 내용입니다. 모델의 예상 밖 출력 사례로 해석될 수 있습니다.
I invented a new generative model and got accepted to ICLR
https://discrete-distribution-networks.github.io/
#HackerNews #generativemodel #ICLR #invention #machinelearning #hackernews
Spectral Labs releases SGS-1: the first generative model for structured CAD
https://www.spectrallabs.ai/research/SGS-1
#HackerNews #SpectralLabs #SGS1 #GenerativeModel #CAD #Technology #Innovation
Once again it’s time for a quick update of activity at the Open Journal of Astrophysics. Since the last update a week ago we have published four papers, which takes the count in Volume 7 (2024) up to 102 and the total published altogether by OJAp up to 217. This means not only that we have reached a century for the year but also that so far in 2024 we have published more than double the number of papers that we published in all of 2023. I blogged about the significance of the figure 217 here.
In chronological order, the four papers published this week, with their overlays, are as follows. You can click on the images of the overlays to make them larger should you wish to do so.
First one up is “A generative model for Gaia astrometric orbit catalogs: selection functions for binary stars, giant planets, and compact object companions” by Kareem El-Badry (Caltech, USA), Casey Lam (Carnegie Observatories), Berry Holl & Jean-Louis Halbwachs (U. Geneva), Hans-Walter Rix (MPA Heidelberg, Germany), Tsevi Mazeh (Tel Aviv, Israel) and Sahar Shahaf (Weizmann Institute of Science, Israel). This one is in the folder Solar and Stellar Astrophysics. The paper presents a forward method for estimating the selection function (i.e. the probability of a system with a given set of parameters being included in a catalog). It was published on November 4th 2024.
Here is a screen grab of the overlay, which includes the abstract:
You can find the officially accepted version of the paper on the arXiv here.
The second paper to announce, published on 5th November 2024. is “Primordial magnetogenesis in a bouncing model with dark energy” by Marcus V. Bomfim (Rio de Janeiro, Brazil), Emmanuel Frion (Western U. Canada), Nelson Pinto-Neto (Espírito Santo, Brazil), and Sandro D. P. Vitenti (Paraná, Brazil). This paper, in the section on Cosmology and NonGalactic Astrophysics, presents a discussion of the possible generation of magnetic fields on cosmological scales by in a model involving a scalar field coupled to electromagnetism
You can see the overlay here:
The accepted version of this paper can be found on the arXiv here.
The third paper, published on 6th November 2024 in the folder marked Astrophysics of Galaxies, is called “Evidence for large scale compressible turbulence in the ISM of CSWA13, a star-Forming Lensed Galaxy at z = 1.87 with outflowing wind” by Itzhak Goldman (Tel Aviv, Israel). It presents a statistical analysis of the spatial distribution and kinematics of nebular gas with discussion of the nature of the turbulence present.
Here is the overlay
The final version accepted on arXiv is here.
Last in this batch is “Star formation in the high-extinction Planck cold clump PGCC G120.69+2.66” by Anlaug Amanda Djupvik (Aarhus, Denmark), João L. Yun (Lisbon, Portugal), and Fernando Comerón (ESO, Garching, Germany). It was published on 7th November 2024 in the folder marked Astrophysics of Galaxies. The paper uses imaging and spectroscopy information to identify sites of star formation in a molecular cloud. This is the overlay:
You can find the official accepted version on the arXiv here.
That’s all for now. I will post another update in a week.
https://telescoper.blog/2024/11/09/four-new-publications-at-the-open-journal-of-astrophysics-9/
#240905329v3 #arXiv240515469v2 #arXiv241100088v1 #astrometry #AstrophysicsOfGalaxies #CosmologyAndNonGalacticAstrophysics #DiamondOpenAccess #GAIA #generativeModel #GravitationalLensing #Orbits #PlanckColdClump #primordialMagneticFields #SolarAndStellarAstrophysics #srtarFormation #starFormingGalaxy #TheOpenJournalOfAstrophysics
미니어처를 리디파인 했더니 오히려 질감이 떨어졌다. 도색가의 표현법이나 작업량을 생각하면 당연한듯도.
Topaz Gigapixel AI v8 - redefine (add realistic detils) 모델
AI의 모사력을 빌려 이미지 덧칠해준다.
옛날 사진이나 게임 스크린샷 크기 키워서 리터치 편하게 하려고 쓰던 앱인데 뭔가뭔가 늘어난다.
파이널 판타지 7 리메이크의 PS5 스크린샷을 넣어서 아무 입력 없이 creativity 값 1.
가죽 완갑과 급속 흉갑에 질감 넣어주고, 실 매듭을 강조하고, 얼굴에는 주름과 모공 질감 넣어주고, 머리카락은 거칠고 부산하게 만들었다.
작업 큐는 이전과 같이 로컬 GPU 쓰거나 클라우드 서비스 토큰을 쓸 수 있다. 매우 큰 해상도의 상업용에서나 쓰일?
I Will Fucking Piledrive You If You Mention AI Again
https://ludic.mataroa.blog/blog/i-will-fucking-piledrive-you-if-you-mention-ai-again/
#AI #LLM #ChatGPT #generativeModel #hype #overblown #overdeployed #overused