@Mdesign Brand designer No the other two aren't finished yet unfortunately. I wanted to release the SFW course first as it's by far the longest and most dense of the three, and I didn't want to delay the release of that for the other two courses :( I'm working on NSFW and LoRa course right now though! They will be shorter and roll into the SFW course, so you'll need that anyways in order to best understand the other two!

@Mdesign Brand designer That's a really interesting idea! So, to make sure I'm understanding correctly, you're essentially saying blending some more stylized character images within your dataset for LoRa training so that the model knows what your character looks like "IRL" and in a stylized format? Or are you just talking about stylized LoRa training in general? Would love to see an example if you have one or if you have seen the strategy you mentioned somewhere else before!

The TLDR since you're going to get a long response here, technically yes, with one very important caveat - our brains perceive real people differently than fully synthetic characters. Creating a LoRa dataset using synthetic data for a real person is a quick way to end up in the uncanny valley :) Believe it or not, in my opinion, synthetic characters are way easier to create for consistency than a real person is, especially yourself. The reason being that our "real" faces have nuance that anyone who looks at you will know. If I were to make a LoRa on myself, I wouldn't generate images of me. I would take the time to set up lighting and take the images myself somewhere for real, because a generated image of myself using even the highest quality reference will simply never 100% match what I actually look like in real life. In fact, I've noticed that for some reason most models have a hard time with my face, I'm not sure why. I like to think of it the same way that people hated to hear there voice in videos in the earlier 2000's, and even today. It's the same idea, but for visuals, not a voice. For real people, I definitely would advise just creating a dataset of real images of that person, that will get you 100% the best accuracy possible from your training data. Now, that's not to say that you can't do it with these workflows in the course, you absolutely could, but just know that given the way our brains are wired, the consistency is going to feel significantly more off than a fully synthetic character (but that's why we have the Face Analysis workflow). Take Taylor Swift again as an example from the course, you could do it with one single high quality image reference of her, but it's never going to be 100% perfect. Interestingly, I actually tried this with Sydney Sweeny just out of curiosity for a couple of images (attached), maybe you'll kind of see what I mean with the actual outputs. Sometimes you get bangers that really do resemble the person (like the attached images), but even then it's just slightly off enough where I would feel more comfortable with images of the person based in reality over synthetic.

@One Two You've got it! I wouldn't necessarily say though that you have to "wait until AI models can do this", because as you've mentioned yourself, they somewhat can depending on what you're doing. Nano Banana like you mentioned has a pretty good understanding of public figures, and can do generally pretty well with good reference images of anyone, but there will ALWAYS be that drift, even in LoRa training. I've yet to see a process with 100% character accuracy, but the method we show in the course is the best accuracy we've seen (mathematically determined with the Face Analysis workflow) largely because we spread across various models like Nano, GPT, Seedream, etc. To your point with Nano Banana though, there's the trade off of censorship and cost on top of the face drift, that's why I say that LoRa's would be the best option here. Yes, it would take time to make the dataset, but once you have that you can use it to train on any model going forward, and it's going to be the most accurate fidelity you could possibly get with very little restrictions since a model has to be open source to be trained on :)

Glad to have you here! Us too :)

Awesome!! Glad it's working! The workflow lock is specifically made to prevent re-distribution of the workflows. That in mind, it should already be in your ComfyUI install in Runpod. I'm guessing you're trying to run it locally? Let me know and if that's the case we can provide the install link for the nodes!