Well, with another 4 months comes another monumental leap in the quality and capabilities of AI image generation models. Recently, a new model called Flux was released, which is a huge step up over prior models.

While still it's still relatively new, work to enable training it on specific topics has already begun and I have been able to do one of the earliest full fine tunings of the model to train it on a new topic.

As with my prior image gen posts, the theme is maple syrup, though I imagine I'll train others as well in the future. Unique to this model are a few key elements:

1. Accurate hands (finally!)
2. Text.
3. Substantially higher fidelity
4. High degrees of prompt adherence, allowing for detailed prompts.

Let me know if there's anything specific you'd like to see!

these are amazing!

Not everyone's cup of tea, but this is also possible.

I still think AI art is an abomination, but I've always trusted you as someone who consistently brings to the community youtube and producer content that's among my personal favorite. So if anyone can turn this abomination into something admirable, you're definitely the man for the job!

Very Realistic - It's insane to think how quickly this continues to progress!

Not sure I'd call myself a fan of AI Wam, but much like the advent of 3D-Printing has enabled rapid prototyping and development of various products on a massive scale at a much lower cost...

AI generated images may one day soon allow a similar ability in the creation of conveying ideas for Custom Scenes by allowing an idea to be Story-boarded fully, thereby allowing the concept to be fully realized beyond just a detailed description.

Just finesse the variables and voila!

Gamma,

Really impressed with your work.

Couple of questions

1. Do you ever run into "terms of service" issues with creating messy content? Meaning do the programs you use refuse the outputs?

2. Have you tried running these images through any type of video processor like Runway or something similar?

3. Can we get a model in tights or pantyhose ina future update?

4. Love the anime girls. I'm curious if you've had any luck getting consistant images for anime characters. I've played around but not able to get the same girls in multiple images.

1. This is a model I run on my own computer that is custom-trained for this style specifically. Refusals and TOS aren't something that comes into play since I am running on my own hardware and have complete control over what I generate.

2. I have not (yet)

3. Brilliant!!! See attached. Can't believe I didn't think of this.

4. Haven't really messed around with consistency yet. A project for a future date.

I also realized that by slightly altering my definition of syrup to something more along the lines of "opaque chunky oatmeal like syrup" and adding in a variety of colors, I am able to achieve a very different effect.

Gamma

Thanks for you hard work! I can't believe these images are looking better than some of my old vintage wam videos at this point.

Do you mind sharing a bit about your hardware set up for all this? And a brief breakdown of what one would need to purchase in order to get started?

Also, do you have a specific version of flux on your hardrive? I wanna make sure I'm buying to proper stuff

Looking forward to seeing more. Thanks for all the pantyhose images, they really made my night!

Once you nail consistancy down you could have your own "cast" and then the world really opens up.

Thanks for sharing. This stuff is super impressive even outside of the wan space.

Haha I've noticed a lot of video generators are "playing" with sliming objects as a way to show capabilities lately. Runway is a good example.

Sure, when talking about any sort of machine learning modelling right now, the king is VRAM. I run an RTX 4090 with 24GB of VRAM, which allows me the ability to train models on specific concepts. Using the custom-trained model, I can create images like the ones above.

Starting off, you definitely don't need to buy your own hardware, as you can use cloud hardware, which will be cheaper and faster to get started with. Search for "Run Flux on Cloud" and "Train LoRA for Flux on Cloud" on YouTube.

For consistency, there are specific tools needed which don't exist yet today. But they're coming, I'm sure.

And yeah, it kind of blows my mind as well. I've been continually disappointed in the quality of existing AI generations, which look fake and uncanny. I think we are past that point now and truly moving into the realm of "indistinguishable from reality." If not perfect, they can be quite beautiful.

These are amazing! Great stuff!

Its absolutely great, thank you for thease pics.
wish more

Reminds me of that scene in The Testament of Sister New Devil, where she gets covered in maple syrup...