@Stable_Yogi I love your new Illustrious models! I've been thinking about starting the Lora training again because the Pony Lora's are sadly not usable, but these models are fantastic! My favorite is the Babes 1.0 DMD version, perfectly lifelike!

@Stable_Yogi I'm still waiting for the person Lora training. My first test seemed good at first, but after trying it on several models and variations, it hasn't been very stable. I really like your basic models (especially the DMD versions), but they don't seem stable enough for the Lora yet, so I'm waiting.

@Stable_Yogi I trained the Lora on FP16 model and tested it on DMD versions. I used the same parameters I use on Pony, and the faces are much more accurate on Pony. So sometimes it's perfect, sometimes it's weird. I don't know, I'm still training some Loras.
Even when I used the beta scheduler, the face was the closest to the original. The euler_ancestral_cfg_pp also works wonders with ddim_uniform in ComfyUI.
Otherwise, I'm not rushing you or the time, this model line has yet to boil itself, because it would be a fantastic change from Pony.

@Stable_Yogi I've made Prodigy and Adafactor-based Lora, and the models seem fine, it seems the first Lora was the one that could have been weaker.

So I love your DMD models and look forward to seeing more versions!

@Stable_Yogi  Im just trying to give feedback. I've tested both the old FP16 version and it does a lot better job at giving characters fiver finger hands. DMD2 version very rarely has correct hands and mouths but has much nice faces and detail over all. The hand and mouth issue though is big enough that I've stopped using it. If these issues could be fixed it be amazing.

@Stable_Yogi Hi Yogi, I'm running A1111 locally, further testing shows that the problem usually goes away after 5-10 image generations if I don't change the prompt. Which sounds ridiculous but I've notiiced on other models that if I change for example from a blue to white background it sometimes takes a few images for it to adjust, which shouldn't happen? I don't know how SD works under the hood.

@Wanquer Do you have the embeddings installed? The first keyword in positive and negative propmts are embeddings (not in generall but in the examples above). There may be more embeddings, you gotta take a look at the prompt.

Also there is some odd behavhiour of some of the webui if you use loras, they kinda may "taint" the loaded model (i.e. if you load a model generate an image, then use an lora, generate another pic, then use a different lora, generate an image and go back to prompt w/o lora it doesnt always looks same as the first generation. restarting may help in that case to ensure the model is unloaded and loadad newly. Quite annoying.

But I know the issue, some of the prompts wont reproduce no matter what, which is kind of frustrating as I use the sample prompts to ensure the model is configured correctly

@Stable_Yogi @Stable_Yogi Whenever I try a new checkpoint/model, I start with a fresh launch, then I set it up exactly as the creator recommends on the civit.ai page I downloaded it from. That includes any recommended embeddings and pos/neg prompt examples, but I no LoRA or anyhting else. Just the barebones recommended minimum. Then I use the exact same, simple prompt to give me a baseline to compare to other models: "waist up portrait photo of an attractive woman posing with her arms folded, in front of a plain black background." I then save the resulting image and use them as a baseline comparisoj to other models. This one (Babes_Illustrious...) was the first one that came up with tight close ups from the start, but as I said, it fixed itself later. No checkpoint swaps or anything.,