​

Good morning, this is Aituts. AI is like an onion - it many layers to unpeel and makes many people cry.

So let us do the peeling for you.

In today’s email:

• Nightshade: a potent or pointless poison?
• InstantID: the end of face LoRAs?
• Game Studio Survey: 50% of game studios are using generative AI

HEADLINE

Nightshade: a potent or pointless poison?

Researchers at the University of Chicago have unveiled Nightshade, a new tool that allows artists to "poison" their images so that AI models cannot be trained on them. [Research paper]

The announcement says that to human eyes, these poisoned images will not look any very different from the originals.

But models trained on these images will learn unpredictable behaviors; for example, a prompt that asks for "cow flying in space" might instead get an image of a handbag floating in space.

At the center of the issue is generative AI models use millions of images scraped from the web, without permission of the creators or copyright holders.

The Nightshade team hopes this project will force AI companies to license image data sets and respect opt-out requests.

Word on the street:

• Demand for Nightshade is very high. Searches for "nightshade" and "nightshade ai" have spiked in the past few days.
• Most technical commentators are not convinced of Nightshade's efficacy. They've called into question the sheer quantity of poisoned images that would be required to ruin a model, as well as whether poisoned images really do have a negative impact on the model
• Another commentator likens it to a pointless "arms race" of training and anti-training activity, similar to the music industry's fruitless crackdown on piracy in the 2000s.

The creator/publisher VS AI model trainer debate will continue to rage on against the backdrop of several lawsuits this year (OpenAI & Microsoft / Midjourney / Github / StabilityAI).

The practice of scraping images has not been definitively ruled on by US courts yet. (Unlike Japan, which has ruled that copyright does not apply to AI training.)

TRY IT

InstantID: the end of face LoRAs?

​InstantID is a new single-image fine tuning method for faces.

InstantID (not to be confused with IP-Adapter Face ID) can generate faces on par with those generated by LoRAs.

The team says that it integrates the face into the background better than existing face-swapping techniques (ROOP/Reactor).

You can try it by running the code, however an integration for Automatic1111 or ComfyUI has not been released yet.

LoRA will still be useful for styles and concepts that are not the face (which is most things).

But it seems like the trend is towards less training, which comes as a relief to anyone who has spent multiple hours on training a bad model.

HEADLINE

50% of professionals working at Game Studios are using Generative AI

The Game Developers Conference's 2024 State of the Industry Report was released last Thursday: it surveyed over 3,000 industry professionals since October.

Notable results:

• ~50% of professionals are using AI or have colleagues that use AI
• That figure may be misleading at a glance: different departments have different willingness to embrace AI; 44% of employees in business and finance use AI, compared to 16% in visual arts and 13% in writing
• 4 in 5 developers have ethical concerns about Generative AI
• 56% of developers are worried about future layoffs
• 12% of respondents said their company prohibited them from using AI. 21% of respondents working at AAA developers said the same.

SHOWCASE

Stable Diffusion Photobooth

​AI-Photobooth, created with Stable Diffusion img2img, running on Arduino and Raspberry Pi.

RESOURCE

Free Book: SDXL Magic

SDXL Magic is a quick and easy handbook on how to get started with SDXL. Featuring:

• My favorite, easy-to-use SDXL ComfyUI workflow
• Recommendations for SDXL checkpoint models, LoRAs, upscalers
• Example prompts for realistic and stylized generations

ROUNDUP

1/ Midjourney Web updates: filter & sort options for images, folder groups.

2/ TikTok releases Depth Anything, a very powerful model for generating depth-maps of images. You can use it as a ControlNet model: [download]

3/ Lykon releases AAM XL, an anime model for SDXL. Lykon is the creator of the massively popular Dreamshaper model and now works at StabilityAI. In the past, many people have commented on lack of good anime models for SDXL in the past, preferring SD v1.5 anime models.

4/ IP-Adapter Face ID Portrait: newest IP-Adapter model, FaceID Portrait, takes 5 face images as input for better face consistency.

5/ The Manga Whisperer, Magi, is a model that automatically generates transcriptions for Manga and comics. It detects panels, text blocks and characters.

6/ Editing AI images is a real job/business now.

That's a wrap.

This was a more news-y one that normal.

What did you think? Would you like more practical tuts than news? Maybe a fine balance?

Let us know!

Till next time,

Yubin & Crew

1223 Cleveland Ave #200, San Diego, CA 92103
​Unsubscribe · Preferences​