Ardy Ala’s Post

To anyone who wants to use 3D as a guide for more control on AI outputs, Ardy Ala just dropped an amazing detailed tutorial on his youtube channel. Check it out!

A great presentation and Q&A of Cuebric by Pinar Demirdag. Cuebric and Unreal Engine allow me to quickly get an idea out of my head and into something visual. perfect for when you just need to get it out of your head. Lately the most fun I've had with Cuebric is letting my 4 year old tell me what he wants to see. The results are a lot of laughs and some pretty interesting results. The Cuebric team has made constant improvements and add great features. The latest is generating a translucent background for your image. Using that to create a simple 3D mesh. The beginnings of something pretty powerful I think. #cuebric #generativeai #AIInnovation

Well, Christmas sure came early this year. Last week I wrote about the LLama3.1 LLM model that you can freely download and use - and now we have another one. Another, really, really good one. We got the text-to-image and image-to-image generation model which is as good as Dall-E 3 or Midjourney and it's called Flux (https://lnkd.in/dsAHKcTG). And you can download it for free, run it locally, fine-tune it, or go nuts with it in some other way. It requires 24-44 GB of RAM (or VRAM) to work (30-40 GB is the norm but there are ways to reduce it to 24-ish by sacrificing performance and some quality), but that is within the realm of a decent gaming rig so definitely not out of reach. The model weights can be downloaded from Hugging Face and the project's GitHub page (with setup and inference instructions) can be found here: https://lnkd.in/dTWuphaH For reference, the picture below was generated using the *smallest* of the offered models (flux-schnell) and it took about 3-4 secs to generate. Yap, it's that good. #ai #imagegeneration #diffusion #flux

I was out of the loop for updates on Ready or not 1.0 quite some time. I was informed in the comment sections on one of my map mod's comments that it appeared briefly in the one of the developer 1.0 development videos. It was rather brief and not mentioned the map name, it was just talking about their AI still working on custom maps rather than a showcase of featured community maps, but still. Neat to see. (Timestamped) https://lnkd.in/e73J9U9W #gamedesign #unrealengine #gamedevelopment #gamedesigner #leveldesign #gamedev #unrealengine4 #gamemodding #modding

I had spent much of last year fine-tuning image models like VGG19 and recently started looking at LLMs. I was under the impression that fine-tuning LLM would be impossible since companies have not provided access to their models’ architecture. I was able to attend the LLM Fine-Tuning Hands-on Lab @ODTUG #Kscope24, and it was an eye-opener! Thanks to Adrian P. and Heli for their incredible insights. Not only were they great teachers, they were also kind enough to connect with me and talk about AI. It was great to explain how I want to use AI with people who are real practitioners and bounce ideas off of them! Their insights were important and got me thinking about ways to improve my app. #AI #LLMFineTuning  #OraclAPEX 

This weekend, I've dived in into ComfyUI. Another AI image generator but with mods. It kind-of works like mod system in Blender. I'm still really new to this, but I've managed to create a workflow that creates a character from a prompt. After that, the face swap steps in. In the picture, I've created a character from the Lora I've trained with Kohya ss for Stable Diffusion. Then I've used a custom node called Reactor for a face swap. You can see some results on the picture. You don't have to train a Lora for this, you can just use any face you want, but it may not look the way you want. #aiart #aigenerated #comfyui #stablediffusion #aiartcommunity

Viggle AI Experiment shows how it is a game-changer. We can basically replace anyone now. Things are about to become really weird.. Not a lot of work but some alpha blending compositing was involved plus #AnimateDiff and 'Differential diffusion'. Step-by-step on how this one was done: https://lnkd.in/gfGU_qmC The tool used (Discord Server): https://lnkd.in/gqp_7AV4

Era3D with Metaverse Applications: 🌟 Era3D: A new method to create high-resolution multiview images from a single image for 3D reconstruction. Addresses issues like camera prior mismatch, inefficacy, and low resolution. Uses a camera prediction module and an efficient row-wise attention layer. Generates detailed 3D meshes from diverse single-view input images. 🌌 Metaverse Applications: Virtual, immersive, and interactive 3D environments where users can engage in various activities, such as gaming, socializing, or e-commerce. Incorporates avatars, objects, and scenes created using 3D modeling techniques and tools. 3D content is crucial for creating realistic, engaging, and visually appealing experiences.

In the recent wave of open source breakthroughs huggingface dropped a revolutionary 𝗔𝗻𝗶𝗺𝗮𝘁𝗲𝗗𝗶𝗳𝗳 pipeline that allows a diffusion model to generate subsequent frames with help of pre-trained motion adapter weights. 🚀 To demonstrate the working of this modular approach towards AI, we have developed an open source huggingface space. The space is capable of generating 2 second long animations (16 frames). Following are customizations available: 𝗡𝗲𝗴𝗮𝘁𝗶𝘃𝗲 𝗣𝗿𝗼𝗺𝗽𝘁: Guides the diffusion process away from undesired features and styles. 𝗚𝘂𝗶𝗱𝗮𝗻𝗰𝗲 𝗦𝗰𝗮𝗹𝗲: Determines how strongly the model adheres to the provided guidance versus exploring randomness (7.5 is a good balance for the diffusion model selected in the space) 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗦𝘁𝗲𝗽𝘀: It is the number of iterations used during the diffusion process to gradually denoise a sample. It plays a crucial role in the quality of generated frames. (Our space provides a maximum of 24 steps) 𝗔𝗱𝗮𝗽𝘁𝗲𝗿 𝗖𝗵𝗼𝗶𝗰𝗲: These are LoRA weights that help define motion of perspective for the AnimateDiff pipeline Space link: https://lnkd.in/d83dkHVt #stablediffusion #huggingface #generativeai #huggingface

🚀 Era3D: A new AI model that creates high-res 🗿3D images from multiple viewpoints using just one input image!🌄 👇 - Generates high-quality images up to 512×512 pixels🎯 - Uses efficient row-wise attention to reduce computation⚡ - 12x more efficient than sota methods💪 Experience the Gradio demo on Spaces: https://lnkd.in/gT23u3sH Project Page: https://lnkd.in/gFcfnjUB 🚀Experience the magic of Era3D with the Gradio demo! Generate stunning high-res multiview images🖼️