Patrefans
Xylix from patreon
Xylix patreon

Forum Handles and Miscellaneous Stuff

🕑 Added 2025-09-23 16:00:15 +0000 UTC

Comments

Steve Barbour

I mainly operate through an AI model aggregator poe. Technically, I could be calling the model through an API using the poe keys, which would give me far better control and setup. "For context: Automatic1111 itself is fairly lightweight, but when usage ready its at least 10GB, each model checkpoint is ~4–7 GB. " Sadly, my crappy old laptop only has 8 GB total. One of these days I'd like to get a real setup to really play around with AI models. "A LoRA is a small, supplementary model that can be loaded alongside a base model. It's great for teaching the generator to reproduce specific things, for example, you could train one for Sylvia, so you don't need long descriptions every time. That way, you can focus prompts on other aspects of the image while keeping the character consistent." That's rather cool. In my mind, the correct way to do this would be to build a library of character images. Extend those images into poses and postures. Add in clothing and examples of the character with that clothing. Then, finally, have a selection mechanism that programmatically pulls from the library based on labels, then feeds them through a chain of AI model activities that produces the final image based on the prompt. Finally, there would be an easy way to flow all this back into the library, breaking down the resulting image into labels to provide a flow of consistency from image to image. And perhaps a way to custom train the models into variants off that library. Many of the pieces to achieve this already appear to be there. But not all the parts have congealed into one easy to use IDE to make it happen yet. I'd really like to play around with this some day. But, the older I get the busier I get and the faster time seems to fly by. Still, you've given me lots of exciting points to branch off of if I ever do take that dive.

Daniverzum

About AI image generation, I'm not sure how familiar you are with it (maybe you simplified things for the post, or just didn't dive too deep). I'm not an expert either, but I thought I'd share a few tips. Flux isn't really an "AI" itself, it's a base model for image generation. You might be using it through a website, which is totally fine if you want to avoid the technical setup. Personally, I use Automatic1111 (an open source local program), and many people also recommend ComfyUI for more advanced, streamlined and customizable workflows. Some of the most popular base models are Flux, SD 1.5, SDXL (sometimes called Illustrious), and Pony. SD 1.5 is a bit older now, SDXL is very solid (my preferred choice), and Flux is newer but more resource-intensive. My favorite checkpoint is Nova Anime XL (which is SDXL based), but that's just a personal pick. Flux may have some anime-capable styles, but it doesn’t yet seem to have as many anime-focused derivatives as the SD 1.5 or SDXL families. For context: Automatic1111 itself is fairly lightweight, but when usage ready its at least 10GB, each model checkpoint is ~4–7 GB. On my RTX 2060, I can generate images decently, around a minute for a 1080p result. It's much faster to generate at lower resolution first, then upscale (or regenerate at higher resolution) the best outputs. If you have at least an RTX 3060, you can also train LoRAs fairly comfortably (though lower cards can still do it, just slower). A LoRA is a small, supplementary model that can be loaded alongside a base model. It's great for teaching the generator to reproduce specific things, for example, you could train one for Sylvia, so you don't need long descriptions every time. That way, you can focus prompts on other aspects of the image while keeping the character consistent. Speaking of prompts: it usually works better to think in keywords rather than full English sentences. You can group and emphasize words with parentheses, or adjust importance with weights (like keyword:1.4). Tools like inpainting are also handy, instead of regenerating an entire picture, you can select just the area you want to fix and re-prompt for that part. Finally, if you can't or don't want to bother with installing and setting up anything, or you want to explore different models, CivitAI is great. There is limited free generation and I'm pretty sure you can also train LoRas there. You can download checkpoints, LoRAs, and even browse other people's results. If you click an image there, you can see the exact settings and prompts they used, which is really helpful for learning.

Steve Barbour

That'd be a fun scene. I'm not sure if I'll ever have a good time in story to do it. The plot in the next several arcs is pretty tight. I might manage it in Arc X: Mortals.

Kurosov

On the topic of the forum comments i'd love to see a scene where the main posters all have their first post life meet up. Their reaction to loli sword god's inaccuracy being the lack of a sword and them complaining about finding out some of Sylvia's worst trolls in person such as those spiders not being edible.


More Creators