Dotphoton

Today at NeurIPS, we’re excited to share our third contribution, 'Is One GPU Enough? Pushing Image Generation at Higher-Resolutions with Foundation Models. Marco Aversa will present Pixelsmith, which is a breakthrough solution for generating gigapixel-scale images on nothing more than a single NVIDIA RTX 3090! What makes Pixelsmith stand out? 🔹 No Additional Training: Scale up to 1000× beyond the resolution your pre-trained diffusion models normally produce, all without retraining. 🔹 No Extra Hardware Costs: Forget large GPU clusters or hefty memory requirements. Pixelsmith does the heavy lifting on one GPU (<24 GB VRAM). What are the requirements?  🔸 A single GPU with under 24 GB VRAM  🔸 A pre-trained diffusion model like StableDiffusionXL  🔸 The Pixelsmith framework (just a handful of lines of code to start!) How does it work? 1. Start with a 1024×1024 base image created from your text prompt. 2. Upsample this base to the larger resolution you want. 3. Guide the final image generation with that upscaled version using patch-based denoising—no huge memory overhead. 4. Fine-tune details or encourage more creativity by adjusting a single parameter: the Slider. Ready to explore Pixelsmith further? The website, paper, and code 🔗 in comments 👇 This work was developed through the efforts of our colleague Marco Aversa and co-authors Athanasios T., Chaitanya Kaul, Roderick Murray-Smith, and Daniele Faccio! #NeurIPS #NeurIPS2024 #genAI #GPU

Today at NeurIPS in Vancouver our Head of AI/ML Luis Oala together with industry colleagues are showcasing Croissant 🥐 - metadata format designed to make working with datasets in Machine Learning simpler, faster, and more responsible. By making datasets more discoverable, portable, and interoperable across tools and frameworks, Croissant eases the friction points in ML data management. Backed by popular dataset repositories and compatible with major ML frameworks, Croissant is ready to unlock the full potential of your data. Learn more: NeurIPS 🔗: https://lnkd.in/d6CdqFGA   Github 🔗: https://lnkd.in/dK4zDZx2 Poster Session: Wednesday, December 12, 11 am. – 2 pm. PST The MLCommons Croissant working group, the team that 'baked' this new standard, brings together AI/ML experts from the following organisations: 🥐  Bayer 🥐  cTuning foundation (founding member of MLCommons) 🥐  DANS-Knaw 🥐  Dotphoton  🥐  Google 🥐  Harvard University 🥐 Hugging Face 🥐 Kaggle 🥐 King's College London 🥐 Luxembourg Institute of Science and Technology (LIST) 🥐 Meta 🥐 NASA - National Aeronautics and Space Administration 🥐  NASA IMPACT - UAH 🥐 Open Data Institute 🥐 Universitat Oberta de Catalunya 🥐 Eindhoven University of Technology Carole-Jean Wu, Costanza Conforti (Stanze), D. Sculley, Dan Brickley, Drew Duglan, Ph.D., Elena Simperl, Goeff Thomas, Grigori Fursin, Iksha Gurung, Joan Giner Miguelez, Joaquin Vanschoren, Jos van der Velde, Luis Oala, Manil Maskey, Meg Risdal, Michael Kuchnik, Mubashara Akhtar, Natasha Noy, Nitisha Jain, Omar Benjelloun, Peter Mattson, Philip Durbin, Pierre Marcenac, Pierre Ruyssen, Pieter Gijsbers, Quentin Lhoest, Rajat Shinde, Slava Tykhonov, Stefano Maria Iacus, Steffen Vogler, Sylvain Lesage, and Yuhan (Douglas) Rao #neurIPS #neurIPS2024 #croissant #metadataformat #ML

We just kicked off Dotphoton's streak of appearances at NeurIPS 2024, led by our AI/ML scientists Luis Oala and Marco Aversa. The first in line is the paper 'Generative Fractional Diffusion Models'. Think Brownian motion is the best noise driver for continuous-time diffusion models? Think again! Today (from left to right), our own Marco Aversa, Maximilian Springenberg, Gabriel Nobis, and Michael Detzel, presented a new approach that replaces traditional Brownian noise with a Markov approximate fractional Brownian motion (MA-fBM), unlocking exciting new properties for image generation. ⚡️Fractional Noise Control MA-fBM lets adjust the “mildness” or “wildness” of the randomness, bridging the gap between standard Brownian-driven SDEs, their deterministic ODE counterparts, and even rougher stochastic paths. ⚡️ Augmented Score Matching Learning a score model with data-matched input/output dimensions is enough to approximate the GFDM score function—no special architectural changes are needed. ⚡️ Better Results Experiments reveal that super-diffusive (smooth) MA-fBM regimes yield sharper images with fewer inference steps, richer pixel-level diversity, and improved distribution coverage compared to standard Brownian-based methods. Learn More: 🔗 Paper: https://lnkd.in/dqQavCMp 🔗 GitHub: https://lnkd.in/eTFGvt9K 🔗 NeurIPS: https://lnkd.in/ecAyNJ-r Gabriel Nobis, Maximilian Springenberg, Marco Aversa, Michael Detzel, Rembert Daems, Roderick Murray-Smith, Shinichi Nakajima, Sebastian Lapuschkin, Stefano Ermon, Tolga Birdal, Manfred Opper, Kristoffer Knochenhauer, Wojciech Samek #neurIPS #NeurIPS2024 #diffusionmodels #ML

This week at NeurIPS, the biggest conference for Machine Learning, we’re excited to share 3 papers that advance the fields of generative models, ML data management, and high-resolution image generation: 1. Generative Fractional Diffusion Models (GFDM) 🗓️ 11 December 📌 We introduce the first continuous-time score-based generative model using fractional Brownian motion to improve diversity and control in data generation. Co-authored with our colleagues working in AI Marco Aversa and Luis Oala, together with Gabriel Nobis, Maximilian Springenberg, Michael Detzel, Rembert Daems, Roderick Murray-Smith, Shinichi Nakajima, Sebastian Lapuschkin, Stefano Ermon, Tolga Birdal, Manfred Opper, Christoph Knochenhauer, and Wojciech Samek. Key Highlights: - Addresses mode-collapse and lack of diversity in diffusion models. - Offers flexibility through fractional Brownian motion. - Delivers superior image quality and diversity on real datasets. 2. Croissant: a Meta-Data Format for ML-Ready Datasets. 🗓️ 12 December 📌 Croissant was developed by scientists from various data-centric organisations like Dotphoton, Google, Harvard, NASA, Bayer, King's College London and others to help manage modern ML data workloads across multiple frameworks. streamlining the process of integrating, organising, and accessing ML data. 3. Pixelsmith: Achieving Gigapixel Image Generation on a Single GPU 🗓️ 13 December 📌 Our Senior Research Scientist in AI/ML Marco Aversa together with Athanasios T., Chaitanya Kaul, Roderick Murray-Smith, and Daniele Faccio developed Pixelsmith, a framework enabling gigapixel image generation from pre-trained diffusion models—on just one GPU. Key Highlights: - Scales diffusion models to gigapixel resolution by a factor of 1000. - Introduces the Slider mechanism for enhancing image details. - Reduces memory demands through patch-based denoising. Thank you to all the amazing collaborators for making this happen. If you’re at NeurIPS this week, let’s connect to discuss these exciting advancements! 💡 #NeurIPS #NeurIPS2024 #ML #Croissant #Pixelsmith #generativemodels

Missed our recent space news? 🚀 Thanks to the initial facilitation of the European Space Agency - ESA through the ESA Commercialisation Gateway, a groundbreaking partnership between SATLANTIS and Dotphoton delivered transformative results in Earth observation. Jetraw Core Compression Technology was successfully implemented on the GEISAT Precursor satellite earlier this year, achieving: 🛰️ 3x increased imaging capacity 🌍 Daily coverage expanded to 175,000 km² This innovation unlocks greater potential for environmental monitoring while improving data quality for AI-driven applications. Thank you to ESA's Spanish broker, Arribes Enlightenment, for documenting this exciting case study. Read the freshly published write-up below 👇 📷 Image acquired by GEISAT over the Tiber Valley (Italy) and compressed with Jetraw Core technology #SpaceInnovation #EarthObservation #ESA #Satellites #AI #Sustainability

Next week is shaping up to be an exciting one! Our work has been accepted as an Oral Presentation at the VH-RODA Workshop, an annual event by the European Space Agency - ESA. 🚀✨ Our Senior AI Scientist, Marco Aversa, will present “Design, Generate, Validate: Satellite Data-Free Machine Learning Model Evaluation from Earth to Space”! 🌍📡 If you’re attending, catch Marco’s presentation on Thursday, December 5th, at 15:30 (see the Agenda) as he unveils an innovative framework addressing the challenges of satellite imaging with limited labeled data. Key highlights include: 🔹 Leveraging drone-collected Earth-based aerial data for satellite model evaluation 🔹 Utilizing open-vocabulary techniques to generate labeled synthetic data 🔹 Enhancing satellite design and operations through downstream application insights This groundbreaking approach offers a cost-effective, scalable solution for assessing and benchmarking machine learning models in satellite imaging. Join us in exploring how this work shapes the future of Earth observation. Don’t miss it! 🛰️💡 #SatelliteImaging #EarthObservation #VHRODA #ESA #MachineLearning #SpaceTech

Are you in Bremen next week for the Space Tech Expo? Come meet Luke Poxon and Michael Desert to discuss our solutions for improving your data management—onboard, during downlink, and on the ground. If you missed the recent news, click the link below to learn how we helped SATLANTIS increase image acquisition capabilities and expand their daily surface coverage up to 4 times 🚀 📍Swiss Pavilion, Hall 5, booth # G47 Switzerland Global Enterprise Swiss Business Hub Germany Space Tech Expo Europe #spacetechexpo #spacetechexpoeurope #spaceevents

In just over 10 days, we will be at the Space Tech Expo Europe in Bremen. Join us at booth #G47 in the Swiss Space pavilion 🚀 Meet with Luke Poxon and Michael Desert to discover how Jetraw Core’s raw compression technology can boost your mission’s imagery and downlink efficiency by up to 6x. Plus, learn how Jetraw, our image data management platform, can cut your cloud costs and speed up data transfer on the ground. We're looking forward to seeing you there! #spacetech #spacetechexpo2025 #datamanagement #satelliteimagery

We're thrilled to have such an outstanding team led by Laura Batti and Stephane P. among our early adopters of the newly launched Jetraw. We look forward to gathering results and sharing exciting insights with you soon! Stay 𝖼̶𝖺̶𝗅̶𝗆̶ tuned, and 𝖼̶𝖺̶𝗋̶𝗋̶𝗒̶ ̶𝗈̶𝗇̶ try Jetraw today: https://lnkd.in/enJ-PwuR #lightsheet #microscopy #largeimagedata #datamanagement #imageanalysis

We are excited for our AI/ML team to present three (❗️) papers at NeurIPS this year! 🎉 One standout is a collaborative work titled 'Generative Fractional Diffusion Models (GFDM)' on improving generative image quality, pixel diversity, and distribution coverage, co-authored by our scientists Marco Aversa and Luis Oala with an exceptional global team, including Gabriel Nobis, Maximilian Springenberg, Michael Detzel, Rembert Daems, Roderick Murray-Smith, Shinichi Nakajima, Sebastian Lapuschkin, Stefano Ermon, Tolga Birdal, Manfred Opper, Christoph Knochenhauer, Luis Oala, Wojciech Samek This work brings together expertise from leading institutions Fraunhofer Heinrich Hertz Institute HHI, Dotphoton, Ghent University, University of Glasgow, Technische Universität Berlin, Stanford University, Imperial College London, Technical University of Munich. We're honoured and inspired to be part of this incredible community, advancing the future of imaging in AI. #NeurIPS2024 #DiffusionModels #AI #ML #GenAI #GenerativeAI