LoRA trained on RizDraws' artstyle for ponydiff: https://civarchive.com/models/257749
Artist: https://x.com/Rizdraws | https://www.rizdraws.com
Because of my lack of experience and time, the version number may be based on changes that have made certain aspects worse. As such a previous version may perform better than a later version.
I'm still learning the ins and outs of the model and LoRA training in general, please leave reproducible feedback or tips so that I may improve over time.
Check the "About this version" section for any quirks or pitfalls I'm aware of or have experienced.
Some Q's:
What's your workflow?
Sampler from images is done with Restarts with segments of [3,2,0.06,0.30], which are values from the paper on Restarts
Initial resolution is typically 768 x 1152 or vice-versa and upscaled by x(5/3), I try to keep the pixel count around 1024 x 1024 for initial gens and adjust the upscale ratio as needed
Upscaler is usually 4x_fatal_Anime_500000_G or 4x_NMKD-Siax_200k
Second sample pass is the same as the first (params are on civit's meta interface) but at ~.35 denoise
A facedetailer is used
I use ComfyUI, sorry
What strength should I be using?
1
Description
Initial Version
Backgrounds have middling detail
I might just suck at prompting on pony, but '3d' in negatives improves stability
FAQ
Comments (5)
What strength should I be using?
Hi, the strength I used for the sample images is 1 which is also listed on the main page on the right. I'll answer your other comment here as well:
In the sample images I forgot to put in the quality tags that pony uses in the prompt when importing them and will remedy this soon.
I'm not sure if this is just a problem with me but there are some parameters that I can't add manually in Civit, for my previews assume that:
1. The sampler, DDIM, has restarts with segments of [3,2,0.06,0.30]
2. The initial image is made at 768 x 1152 or vice versa and then upscaled to target
3. The upscaler I use flip-flops between '4x_fatal_Anime_500000_G' and '4x_NMKD-Siax_200k' depending on the texture I want
4. The upscale pass is the same as initial, but with a denoise of 0.25-0.35
5. CLIP skip is not used
I'll update my lora descriptions with this information later, hope this helps.
This is my favorite
can you make a IL version?
Not likely, since Riz is already built in to Illustrious models and I found no egregious issues with the tag to warrant a lora. Sorry.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.