Type | |
Stats | 475 146 |
Reviews | (62) |
Published | Jul 2, 2024 |
Base Model | |
Training | Steps: 4,500 Epochs: 6 |
Usage Tips | Clip Skip: 2 |
Hash | AutoV2 F7A6C27EE9 |
These lora were trained on about 500 images from the game Rance 03 and some of 02. The idea was to capture the style that Onono Imoko used in the game. They can reproduce the GC style from the game's events as well as the full body character illustrations. (it was all in the data set used)
The V2 was trained with different settings and I used AutismMix as the base. This V2 is probably the most competent lora I've ever created.(so far) It listens to the prompt very well. It creates accurate hands and feet very often, and most of the images it creates are very aesthetic without taking any extra steps.
V2 Hadrian DeliceXL 2.0 was trained using 50 selected images from the V2.0. I haven't tested the lora on any other pony based model, but I speculate it won't work very well on anything but Hadrian. This version sits nicely in between the AustimMix and Ponymagine lora. It looks much closer to the source material like Ponymagine but has the prompt freedom of the AustimMix. I highly recommend you use this version.
The V2 of Ponymagine was trained using 50 selected images from V2.0. I also tested it on Animagine XL V3.1 and it seems to work (barely) but I highly recommend you just use it on Ponymagine since it's the entire reason I made this version. I do not guarantee support for anything else. Please read this if you're wondering why I made this.
The base for V1 uses Pony Diffusion V6 XL and it works on AutismMix as well. It also produces great images. Choose this one if you work exclusively in base Pony.
These are my settings that will give you the results you see in the V2 images above.
Prompt:score_9,score_8_up,score_7_up,score_6_up,source_anime,<lora:Cool_Rance_XL_V2:1>, [your prompt]
Negatives: mosaic censoring,bar censor,censored,source_cartoon,source_pony,source_furry,simple background,
Sample:DPM++ 2S a
Steps: 26
CFG Scale: 7
Extension: [sd-webui-color-enhance] set to 0.5 (this increases the saturation a bit)
(I have abandon Highres Fix for MultiDiffusion, please follow these steps to reproduce the results from above)
Note: The webUI I use is Forge (fork of Auto1111)
1.Gen your image in txt2img
2.Send image to Extra tab and upscale it by 2x using your choice of upscaler (I use 4x_NMKD-Siax_200k)
3. Go back to the txt2img tab while its upscaling and send the original image to img2img (we want the prompt and settings info)
4.Go back to Extra tab and send the upscaled image to img2img now, replacing the old image
5.Enable MultiDiffusion, Set it to Mixture of Diffusers
6.Set the Tile Width/Height to the original image dimensions (ex. if you gen at 1024x1024, that is what you put here, do not use the upscaled dimension from Extra tab)
7.Scroll to Denoising Strength and set it to 0.33 (anywhere between .25 to .45 is good, higher numbers results in more drastic changes)
8. Click generate until you get an improved image you like
9. If you like a gen but only want to slightly change the results then save the seed and increase the Denosing Strength
Why have I switched to this slightly more complicated method?
Because it is completely better than Hires fix. It improves the image with high denosing without ruining it like hires does. It's faster than hires. Once you understand the steps it's very easy to reuse. The only draw back is you have to reset the dimensions in MultiDiffusion each time you change them on the original gen image. Also, you might occasionally get unwanted object added to the new gen. (for ex, if you gen breast you might get a second pair of nipples, this is a very easy fix, just remove breast from the prompt in your img2img tab) But it's worth it. As you can see above, using this method not only improves but fixes bad hands, feet, eyes, distant objects, details, etc.
Enjoy!