RDBT [Anima]
This page contains the "original LoRA files" for advanced users.
See the main page for more info. (ckpt version, usage, faq, int8 ckpt...)
See this page for update log and version info.
Sharing merges using this model is not allowed. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
Description
FAQ
Comments (9)
Why are you a god? This version just improves everything.
Firstly, thank you very much for the model. I was currently testing characters on anima_preview_rdbt_finetuned_cfg_distilled_v0.20b and the quality that the LoRA adds to the images is simply amazing. The only thing I can complain about it is that it has a severe overtraining on exaggerated proportions female bodies (I think "thicc" is the term). Any possibility of dialing this down on future releases? Also, please release versions without furry. I'm certain a lot of people do not care for it.
Testing on v0.19fd b and v0.12fd based on preview v1,with cfg1.0;3.0;4.0:
1. Using either tags (around70~120token)or natural language (around200~300,Describe lighting without depicting emotions, metaphors or atmosphere; in other words, it is a personal habit to deconstruct the latter into concrete composition and lighting.).Under this setting it seems to reduce the dynamic poses of characters to a certain extent, making them appear stiff
2. It does enhance relatively unrestrained and rough styles
3. It also appears to enhance specific body parts, which is particularly noticeable in the LoRA I trained myself; my LoRA performs extremely poorly on that body part
4.The textual expression is rather strange, or has deteriorated.
I think this LoRA constraint would be a more appropriate usage method for me in the last 20% of the denoising stage.
Gonna merge this one with P2? I'd prefer to not have to download the base model and use the lora on top 😅
Best one for Anima so far I've seen :)
About the trigger words you added, I think it's a perfect addition. Sometimes you just want anime stuff instead of realistic, but "anime coloring" tends to take out a lot of detail to make it look like the literal media, instead of the general style. It's good to have that pointer.
Although, from my initial testings, "digital art" is not going 2d at all, it's going full photorealistic. The anime one seems to be working fine.
Do I have to use the default Anima preview 2 checkpoint or can I use any preview 2 checkpoint with p2 v0.23fd?
Lora is only guaranteed to work on the base model it was trained on.
Lora is a small patch of the weights. Base model weights (that trained the lora) + lora patch = valid weights state that was optimized
Any change to the base model weights will cause deviation in the final weights, aka stability issues
It might work on other ckpt, but has cost, depends on how "far away" this ckpt is from the ckpt that trained the LoRA. "Trained/FFT" ckpts should be ok, their weights are also valid and optimized, usually very close to the original ckpt. But "merged" ckpts usually not, their weights are already f***ed up.
@reakaakasky oh okay thank you for your clarification as I have used it with other p2 Checkpoints like AnimaYume And it works fine. However it isn't as consistent which is probably why as I'm not using the base P2.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.