Show HN: Guiding LLM outputs using Zod https://ift.tt/FHS1KvI

Show HN: Guiding LLM outputs using Zod https://ift.tt/nV8qMsp July 22, 2023 at 02:32AM

Show HN: Datalake for Computer Vision Projects https://ift.tt/6uKTr1Y

Show HN: Datalake for Computer Vision Projects Buddhika, Kelum, and Chong Han here. We are building a self-hosted data infrastructure platform for computer vision. Our community page is https://ift.tt/swfbVh8 In the past, we worked on a couple of high-scale computer vision projects in retail, farming, and hospitals in various capacities. These projects involved 2D object sections, 3D object tracking, and more advanced 3D perception. Like other CV Engineers, we observed a common factor during these projects: one needs a large volume of high-quality data to build a production-deployable CV system. Our biggest challenge was not having a robust data infrastructure to handle large volumes of data. Our S3 buckets were like a data swamp; we had so much raw image and video in storage buckets without tracking. Instead of working on CV, we had to develop tools for data operations. We understand that many of us have our own custom scripts and stitch them together to make things happen in the CV pipeline. However, it is brittle and cumbersome to maintain. We wanted to build a system on top of the cloud buckets such as S3 that store all file indexes, labels, metadata attributes, inference outputs, model training outcomes, and literally anything related to machine learning/computer vision. This makes it possible for us to search for anything and consume efficiently. This behaves as a DataLake (by the way, "DataLake" is an overused term). All other downstream processes in the CV pipeline can access data more efficiently via SDK and can also return data back to the Lake (e.g., training/inference outcomes). The reason we made it self-hosted is to address data security and privacy concerns. Since data is fundamental to AI, we believe that companies and organizations should have complete control over it. Currently, we support AWS, GCP, and Azure cloud buckets; soon, we will support local storage. We ship this as a Docker container so you can just install it on any VM or local server. The installation script will do all the configuration automatically. The Python SDK and documentation are available but not perfect yet. We’ve launched this under MIT and Elastic licenses so any developer can use it. Our goal is not to charge individual developers. We make money by charging a license fee for things like multiple users, multiple buckets, scalability with K8, and providing support. Give it a try: https://ift.tt/swfbVh8 Let us know what you think. July 22, 2023 at 04:45AM

Show HN: TextToSample – Desktop App to generate audio with MusicGen locally https://ift.tt/kcN906P

Show HN: TextToSample – Desktop App to generate audio with MusicGen locally https://ift.tt/CxkM6rG July 21, 2023 at 09:23PM

Show HN: CopilotKit – a hackable OSS copilot for any react app https://ift.tt/CTMymx2

Show HN: CopilotKit – a hackable OSS copilot for any react app CopilotKit is a typescript library for adding a hackable copilot to any react app. You can let the copilot interact with your app via plain typescript closures , and give it (explicit) read access to app data. An example user interaction could look like: - "Which of these travel destinations has a rich architecture history?" - [Copilot answers] - "Great. Add these to my august trip folder, except the ones where it's typically rainy then" - [Added] Recursive agent integration (via Langchain) is in the works - if you have ideas I'd love to hear them here or on the discord. https://ift.tt/AnxHXCw July 22, 2023 at 12:29AM

Show HN: 3.0 Version of Invoke AI – open-source SD UI and Node-based Back end [video] https://ift.tt/qBrva7m

Show HN: 3.0 Version of Invoke AI – open-source SD UI and Node-based Back end [video] Hey all - Invoke started as one of the earliest Stable Diffusion UIs (you may remember it as “lstein”), and has evolved significantly into a full fledged react/typescript web app. We’ve been hard at work building a professional-grade backend to support our commercial move to serving businesses and enterprise with a hosted offering (invoke.ai), while keeping Invoke one of the best ways to self-host and create content as an open-core project. As of 3.0, all of the developments we’ve been working on and tweaking for our hosted environment are available to install and use locally, including an API and graph-based execution architecture - And, to demonstrate our commitment to free and open-source software, we’ve updated our license to the most explicitly permissive license available - Apache 2.0. — New SD Support in our 3.0 Version: - SDXL Support - We’ve implemented support the impending SDXL model architecture (And the current 0.9 model), and we’ll follow-up with streamlined SDXL support in the core UI interfaces once the 1.0 model is released. - ControlNet - Integrated support for the most popular ControlNet models, directly in the UI, with a simple processor preview and UI/UX. - Boards - Expanded gallery support to better organize and manage large scale images. Multi-select, drag & drop to anywhere in the UI, and backed by a local database to provide performant operations even when you’re thousands of images deep. - Expanded Schedulers - Rather than list all of them here… All the schedulers/samplers you know and love, with the ability to set favorites and disable those you’ll never use. - Model Flexibility - Swap your VAE on demand. Mix and match models as needed. Clip Skip. With the flexibility of the experimental Node Editor, you can even swap models mid-generation. - LoRA Enhancements - Full LoRA support (for all the Lo’s and Ly’s you can name), we’ve also added a mechanism which directly patches the model UNet on loading a LoRA. Test for yourself. - UI/UX Updates - Across the board, we’ve worked to clean up the UI, optimize the options panel for the most commonly used features, and left a number of the tiny little microinteractions across the app that make using Invoke easy for your workflow. It’s our best UI yet, hands down. And more is coming. - Node Editor (Experimental) - The main reason 3.0 took as long as it did (5 months!) is because we disassembled the entire backend of the application, and put it back together one “node” at a time. It’s clean, streamlined, and scalable. This sets us up to drive powerful advanced experiences for power users and gives an easy way for contributors to extend the capabilities of Invoke. The Node Editor that exposes all of the available functionality in the background is in an “experimental” status, for explorers and developers - Mainly because we have a lot of UI/UX polish we want to apply to it, and provide better ways to help less experienced users be successful with it! — *Up Next* More is coming. We have plans to extend on many of the core UI/UX experiences in the application, add more community resources for sharing new plug-ins/nodes, and will release the full version of the Node Editor in 3.1. Stay tuned! –- Whether you're a dev looking to build on or contribute to the project, a professional looking for pro-grade tools to incorporate into your workflow, or just looking for a great open-source SD experience, we're looking forward to you joining our community. You can get the latest version on GitHub https://ift.tt/xeAqhcj https://www.youtube.com/watch?v=A7uipq4lhrk July 21, 2023 at 10:33PM

Show HN: Open Video Game Data: A new approach to evaluating games https://ift.tt/JYMpHZj

Show HN: Open Video Game Data: A new approach to evaluating games > Introduction Our idea is to offer an alternative to well-known sites like Metacritic and OpenCritic, but with a different approach. Instead of being a score aggregator, we will be a list aggregator. Metacritic brings together reviews from multiple review sites in one place, providing a final score of 0-100 based on a weighted arithmetic average, where some critics carry more weight than others. An alternative to Metacritic is OpenCritic, where all critics are weighted equally in the final average. However, both still work with numeric scores. > Why relying on scores can be problematic? - Ratings only reflect the state of the game at launch Today, more than ever, games are constantly evolving. It is common to have "patch day one", that is, games released with bugs and incomplete content. However, with time and help from the community, these games can be improved, as was the case with No Man's Sky. When No Man's Sky was released in 2016, its average on Metacritic was just 61, due to the troubled release. However, over the years, the game has evolved significantly with updates, but its Metacritic score remains frozen at 61. Alternative: As the lists are constantly evolving and updating, they more accurately reflect the current quality of the games, tracking their improvements and changes over time. - The average score can be unfair as it is based on the amount of critics Sometimes, the amount of crits heavily influences a game's rating. An example of this is The Legend of Zelda: Ocarina of Time, with an average of 99 on Metacritic, based on 22 critics. While The Legend of Zelda: Breath of the Wild averages a 97, based on 109 critics. Getting a high average based on a large number of critics is extremely difficult, and this can influence the overall perception of a game. Alternative: When a final list is created, all games have an equal chance of appearing in different lists. For example, game A might be included in 3 out of 11 lists, while game B might be mentioned in 5 out of 11 lists. The total amount of lists will always be the same for all games. - Relying on an average can be inaccurate Metacritic converts the different rating scales of review sites into a single percentage-based quantitative scale. However, this conversion can be inaccurate and unfair, as each site uses different rating systems. This approach can result in important information being lost during conversion, affecting the accuracy of the final result. Alternative: With our ranked lists approach, we eliminate the need to convert rating systems, as all lists, regardless of site, follow the same common logic. In all lists, there will always be first place, second place, and so on. > A great alternative: *Open Video Game Data* Our site aims to be just another alternative to note-based sites. Our approach to aggregating lists allows users to have a more comprehensive and up-to-date view of games as these lists are constantly updated by the community. The calculation method is quite simple and transparent. All lists on the site have a maximum size of 15 games. When a game ranks first in a list, it is rewarded with 15 points, while if it ranks last, it only receives 1 point. > Conclusion Open Video Game Data seeks to provide gamers and game enthusiasts with a reliable tool to make informed decisions about which games to play, taking into account critics' opinions and the ongoing evolution of the gaming industry. With the active participation of the community, users can add critic lists and can also create personal lists that are also aggregated, we hope to build an inclusive and reference platform for the gaming community, promoting a more complete and updated analysis about the games that so much we love. Come be part of our community! Create an account and join us to explore the world of playlists. Welcome to Open Video Game Data! Visit us at: https://ift.tt/xoEZwUe July 21, 2023 at 05:56AM

Show HN: A fine-tuned Stable Diffusion model for generating Minecraft skins https://ift.tt/TER9mSU

Show HN: A fine-tuned Stable Diffusion model for generating Minecraft skins https://ift.tt/AhpyVxM July 21, 2023 at 04:10AM