Note: I did not train these LoRA models myself. Original source: https://huggingface.co/stma/lora-dump
Recommended settings
Quick start
Model: AbyssOrangeMix2
LoRA weight: 0.6 (UNet: 0.6, TEnc: 0.25)
[realistic], [3d], (3dcg), ((octane render)), [fisheye]in positive prompt for realismsimple background, grey backgroundin negative promptCLIP skip: 2
Sampler: DPM++ SDE Karras (Euler a recommended for other models)
EasyNegative in negative prompt
Usable
LoRA weight range: 0.2 - 1.15 (UNet: 0.35 - 1.15 TEnc: 0.2 - 1.0)
UNet is required for the style. UNet over 0.65 will look like an anime figure product. Too high UNet weight will melt details.
Lower TEnc weight will improve generalization. High TEnc will affect composition to look more like an anime figure product photo. Too high will fry the details.
Start with lower weight and increase as necessary. Requires more weight when mixed with TI embeddings and other LoRAs.
Figma
Trained on 5,746 images scraped from GoodSmile's international catalogue, including the discontinued products. All images were tagged with figma first and the product name with figma stripped out second, so an image from a product named "figma Hatsune Miku" would be tagged figma, hatsune miku. All other tags came from the WD1.4 tagger. 3d and realistic tags seem to work well with it.
The alpha16 version is better at generalizing, while the 7e-5 version is better at capturing the figure-with-transparent-stand look.
Example images
Model
Counterfeit-V2.5 (TODO: update example images to AbyssOrangeMix2)
Textual inversion
Roxy Migurdia (character)
roxy-migurdiaMegumin (character)
megumin_1000Char Turner (composition)
chartunerEasyNegative
EasyNegativeBad Prompt
bad_prompt
Additional LoRA
Yelan (charater)
yelan \(genshin impact\)Torino Aqua (style)
torino aqua
Description
The alpha16 version has a stronger anime figure look and is better at generalizing and has better details.
FAQ
Comments (12)
大佬 可以去除那些可动关节的模型吗 求求了
那你去下AOAOKO [PVC Style Model]就行了，手办模型
这种关节不就是Figma的特色吗
hi, so im pretty new to this whole things and have some learning to do, but can someone explain to me how people are getting these super realistic gens with this? all my gens look really bad even when using the same prompts and settings, for example the faces look molten, its poor resolution, the figures don't have any joints, the limbs look weird, all that typical stuff basically, is it just based on luck if you get such a crazy gen or is it something else? it it inpainting? idk, anyway, any help would be appreciated!
Hello!
I'm not a professional or less on this, but maybe I can help
First: The Model, Which model/checkpoint are you using? Not all LoRas works well with any model/checkpoint. The model used on example images are Counrterfeit-V2.5, and also recomend using AbyssOrangeMix2 (Both can be find here on CivitAI)
Next: The LoRas/textualinversion, are you only using this LoRa? If that the case, you might need some more,
1.- This actual LoRa (To make them look like the figure)
2.- The LoRa/textualinversion for the character you want to get (To get a closer aparience on the character)
3.- LoRa/textualinversions for bad prompts (This page recommends EasyNegative but you can use anything that is good enough)
4.- Any other you can think. You get the point...
Oh! And don't forget to use the "trigger words" from each LoRa/textinersion
Now: The LoRas Weight, This might sound more difficult that it is, you have to "Balance" the weight of each LoRa to get a proper output, try to play with them. You may find the recommended weight on each LoRa page, if not, you will have to test them.
I just tried the LoRa and this conf worked to me:
Checkpoint: Counterfeit v2.5
LoRa (Android1818DragonBall_v1):
Weight: 0.65
LoRa (Figma Anime Figures)
UNet: 0.9 ~ 1.50
TEcn: 0.9~1.30
Steps: 47
CFG scale: 12
Clip skip: 2
Sampling Metod: Euler a
Promp: best quality, masterpiece, upper body, figma, android_18, and18, 1girl, blonde,white background, anime
Negative promp: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, badhandv4, verybadimagenegative_v1.3, easynegative, bad_prompt, (The last ones are other LoRa/TextualInversion)
I noticed every checkpoint has their own apropiate weight. If you try another you will have to play with the weights.
Hope I could help, if you have any question, I’ll try to answer ^^
Also I posted the images I made, take a look to the results if you want to ^^ :
https://civitai.com/posts/125637
@Tujfain04 sorry, can you explain what is mean UNET and TECN ???
"LoRA weight range: 0.2 - 1.15 (UNet: 0.35 - 1.15 TEnc: 0.2 - 1.0)"
What does UNet and TEnc mean here? LORA Weight I understand, that's just :0.2 to :1.15 after the LORA in the prompt, but I have no idea where to specify UNet and TEnc? Is it through an extension or are you saying that the weight range 0.2 - 1.15 equals around 0.35 - 1.15 on the UNet encoder and 0.2 to 1.0 on the Text Encoder?
did you find out what it means by any chance?
@vwzrd No idea 😔
It is in the Kohya Lora training 参数 I guess
i ask the same question...
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.