Work in progress, Ashara Zavros from Star Wars: The Old Republic. Future versions coming soon.
Description
Comments (7)
Doesn't seem to work for me. Did the trigger word actually get applied during training? I notice it's not in the sample picture prompts (and those innate traits+default costume really should be listed in description when they exist).
Also, don't know if you know this already but since I'm writing this and it's such a common issue I see, I'll just go ahead and tell you: When using 3D (polygonal) data in training, the image needs to get the "3D" tag. A LoRA made with 3D data works fine with that simple thing (at worst a LoRA made with only 3D data will need some kind of style specified for generations), but without tagging "3D" the LoRA will get a rather warped.
This isn't a model for creation to be specific, this is an initialization model solely for the purpose of generating additional references to create a more functional model. Its a process I've had to do in the past when there's a very limited image set to produce the model and have to do subsequent iterations refining from the very basic initial LoRA
To be specific I had under 20 images to work from and using the in-game model for additional images is somewhat pointless due to the low polygon count and resolution even when supersampled and upscaled to the maximum degree.
I'm currently working on version 2 and it should be far more functional.
@Arghal You do not need photorealism to use models as training data. I've done images based on sixth gen era (PS2/GCN) models as the sole training data (e.g., 1, 2, 3) just to see if I could and they absolutely work. I've even done some with sprites to see how they work out, and they work surprisingly well (only issue is it needs a face fix when genning full body since the face for full body stuff was 14x14).
If you're looking to make photorealistic generations of a character that's not really in my area of expertise, but the few times I have tried it was with a pair of models based on crude black and white illustrations, and they worked well enough given my total lack of experience in that section of generation (I've also seen a few stray photorealistic gens made with my stuff posted). My impression is that even if you want photorealistic generations out of a LoRA, the only requirements would be that there's an "actual" face geometry that can be interpreted as a real enough face (as opposed to e.g., a flat anime face with huge eyes). I consider 20 the practical minimum for characters, but 3D models offer such a variety of pictures I'd go to 30 (more if shes got alternate outfits). Just vary the angle (just combining the tags straight-on, from side, profile, from behind, high angle, low angle can get you loads of images), and preferably the pose+BG
@NanashiAnon Its less about photorealism and more about correct detailing. I ran around 1000 test images through this version until I was able to get accurate details like the montrals, lekku, abd markings so I could use them as an image set for the next version. This includes things like numerous emotion subsets and different angles/viewpoints
@Arghal You're not going to get correct detailing from throwing loads of synthetic data at training. Just take more screenshots of the model.