If you like my content consider supporting me! 😊
1 Cookie = Early access to future releases! 🍪
Lora version available for supporters! ❤️
This model attempts to transfer the style of SD1.5 based Rouramashin SR model, into a SDXL Pony based model. Its style sits between a '2.5D' animated model and a '3D render' model. You can use quality tags to help push the style either way. When using 2D/anime character LoRAs you might want to use tags such as (realistic) to help maintain its style (hair strands instead of flat anime hair for example), in addition to properly adjusting the LoRAs weights.
About v1.0
Version 1.0 consists of a custom block merge recipe between multiple Pony v6 based models and trained style lycoris models, with additional training made from carefully selected images generated from Rouramashin SR v2 model.
Recommended parameters:
Positive prompt: score_9, score_8_up, score_7_up, rating_safe BREAK
Negative prompt: score_4, score_5, score_6, simple background, white background, source_furry, source_pony, source_cartoon
Clip skip: 2
Sampling method: Euler a, DPM++ 2M Karras, DPM++ SDE Karras (suggested, use whatever you prefer)
Sampling steps: 20 (minimum)
CFG Scale: 7 (default during tests)
Hires. fix:
Upscaler: 4x-AnimeSharp (recommended)
Hires steps: 10 (minimum)
Denoise: 0,34
Upscale by: 1.5
For this model hires. fix is not strictly necessary, but if characters are far away, hires. fix or ADetailer (facedetailer) might be necessary.
Disclamer
All the models that I have developed and published are for entertainment purposes only. They are derived from the base Stable Diffusion 1.5 model, which means it also includes its inherent issues, limitations and unfiltered capabilities. With that in mind, I am in no way responsible for any content that the user creates with malicious intent while using any of these models. The user assumes complete responsibility for the misuse of these tools and any unethical content created with them. This disclaimer is subject to the laws of the United States of America and the state of California. I reserve the right to remove or report any content that violates the terms of service, ethical standards, or applicable laws.
License conditions
The model published here can be used for personal, non-commercial purposes only. It includes several publicly available resources in its development, each with various types of licenses. It is discouraged to upload this model for use in online generation services: I am not responsible for any license infringements.
Description
Initial release.
FAQ
Comments (4)
Controlnet models working for other similar checkpoints are ineffective here. Not sure what went wrong.
This is a symptom of PonyV6 based models as far as I'm aware. The model was trained so heavily that lost compatibility with sdxl controlnet models, similar to how sdxl loras are not compatible. I've read about people having some success when using the full version of the controlnet models, but I can't guarantee since I've never tested it myself. Also, they are much larger and using them are more computational intensive. You can find them here:
https://huggingface.co/lllyasviel/sd_control_collection/tree/main
Thank you for your message.
I always use the full controlnet, which is why I was surprised. I was somewhat able to increase the weight a lot, and it started following the composition of the reference image.
For reference, most checkpoints, including some derived from PonyV6, will require a weight around 0.9 to get decent results. In comparison with this one I had to go up to 1.6 to obtain a similar level of control.
Hi, could you pls enable this for on-site generator? Everything is opening up! 👍 Thanks for making!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.