newslounge.co · March 29, 2024

Good Morning. 

This is newslounge, We spread VFX-AI news like butter on toast – smooth, even, and covering every corner.

Before we dive into today's dose of madness, I wanted to remind you of something that people often forget:

*What is hard becomes easy - without upping the bar daily, you'll end up stagnating.
*Incremental improvements or changes can add up to significant progress over time.

I know many of our friends are struggling right now to find their path in life and navigate through all the changes coming their way. You should know that you're not supposed to have all the answers; sometimes, being open-minded about our situation is the best we can do. Change is necessary, and sometimes a bad decision is better than no decision.

On Today’s menu:

• Corvette Stingray - Houdini and AI - Case Study

• LoRA Training in Kohya_ss

• Large Language Models (LLMs) And ComfyUI

• Claude 3 or GPT-4?

• Exciting Updates in AI Video

-Ardy

was this email forwarded to you? you can sign up here.

🚨Just a quick note: You can share this newsletter with your friends using the unique referral link at the bottom of this email. By doing so, you'll earn points that can be exchanged for some cool prizes. Let's spread the word! 😎

CASE STUDY

1969 CORVETTE STINGRAY - HOUDINI & COMFYUI

💡I created a new project utilizing Houdini and ComfyUI. With each project, I aim to introduce at least one new element into my workflow, seeking new and more efficient methods. This time, my primary goal was to train an effective Lora model using my reference images. This was to ensure a consistent look throughout the project and to improve image quality, reducing my reliance on text prompts.

🙋‍♂️Here are some of the key points:

LoRA Training:

Training your own LoRA can significantly enhance your workflow, bringing you one step closer to having a reliable system for achieving desired results, yet still retaining the flexibility of AI.

I gathered a full gallery of the same car model, including the interior, and used Kohya_ss to train a LoRA. I used 55 images and trained my SD15 LoRA with 4 epochs, keeping my total steps around 2300. This process took about 25 minutes with my RTX 3090 Ti.

🚨Saving your LoRA at every epoch allows you to test each version and find the best one to use.

LoRAs trained from SD v1.x will only work with models trained from SD v1.x.

Generally, realistic LoRAs work better with realistic models, and anime LoRAs work better with anime models. However, this distinction isn't so clear-cut, as most popular models today are blends of realistic and anime/stylized models.

🚀If you're interested in training your own Lora and would like to go through the steps, I highly recommend watching these videos:

Stable Diffusion: How to train LoRA for Style | LoRA Part 1

Make a LoRA in Kohya like a PRODIGY

Checkpoint:

The only model that really delivered the results I was expecting was Juggernaut SD1.5. Epic Realism was too intense for my taste and tended to distort the car's interior.

Part of the animation uses the new Animatediff LCM setup for faster iteration, but Animatediff v3 with the v3 Adapter remains superior.

🚀 Youtube Link: here

🙋‍♂️Train LoRA in ComfyUI:

You can train a simple LoRA within ComfyUI using a specific set of tools. However, I strongly advise setting up a fresh installation of ComfyUI dedicated solely to this purpose. This is because you'll need to install a different version of PyTorch (PyTorch 2.2.1 with CUDA 12.1), which isn't compatible with some other ComfyUI tools.

💥One more thing: Please use the poll section below to share your feedback on this case study.

Large Language Models (LLMs) And ComfyUI

We are seeking:
full-stack software engineer
web developer

[email protected]

SNIPPETS

😲Autodesk: Interesting AI tools are coming to Autodesk Maya and they are looking awesome.

🤖Claude 3 surpasses GPT-4 on Chatbot Arena for the first time.

🌎The Sphere: The world’s biggest canvas, How Sphere’s pixel power brings art to the masses.

📈AI diffusion models for image generation are getting 30x faster.

🤿SAG-AFTRA ratified an AI protections contract for voice actors, requiring producers to negotiate with the actors’ union before using AI voices in TV animation.

🚀OpenAI recently showcased seven clips created by filmmakers and visual artists using Sora, its innovative text-to-video AI tool. Highlights include a short film about a man made of balloons navigating perilously through a store's cactus aisle, and captivating clips of a camera effortlessly soaring through space and time.

🚨Adobe claims that its Firefly model is commercially safe because it’s only trained on Adobe Stock images, openly licensed content, and public domain content, therefore presenting less risk of generating something that violates another company’s intellectual property. I imagine that sounds pretty appealing to organizations looking for viable ways to implement AI into their pipelines, but Adobe still needs to convince its investors.

😲With the new Adobe Structure Reference feature you can generate new images that match the structure of any reference image.

AROUND THE WORLD OF AI

EXCITING UPDATES IN AI VIDEO

What is LTX?
LTX Studio can generate complete video narratives based on simple ideas.
It simplifies the creative process from ideation to production, enabling users to effortlessly turn their concepts into compelling visual narratives.

Unique Approach:
In a groundbreaking collaboration, LTX Studio has partnered with the esteemed Ari Folman, an Oscar-nominated and two-time Golden Globe-winning director, for its inaugural project. This partnership epitomizes LTX Studio's mission to revolutionize storytelling by infusing it with the latest AI video generation technology.
The studio recently unveiled what many are hailing as the best commercial in the AI video generator category. Not only does the ad captivate with its emotional depth, but it also demonstrates the unparalleled potential of current tech to bring dreams vividly to life.
Aimed at a wide audience beyond professional creators, the commercial stands as a testament to the creative genius of the marketing and storytelling team at LTX Studio. Their efforts have resulted in a piece that doesn't just advertise but inspires.

SORA IS TRULY A GAME-CHANGER

OpenAI has recently teamed up with a select group of builders in the US to experiment with usage-based earnings for GPT.

-NL

Earn free gifts 🎁

5 referrals - “Water Bottle Stickers” 🔺15 referrals - “Mystery Box”
25 referrals - “Water Bottle” 🔺 40 referrals - “Nuphy Desk Mat”
60 referrals - “Logitech Mouse”