Disclaimer
Financial support is very much appreciated, but please do so by donating to my Ko-Fi instead of sending me Buzz, as I currently have no uses for CivitAI Buzz!
Description
An SDXL 1.0 LoRa made by AI_Characters for producing images with an artstyle similar to screencaps from from the anime movies Your Name, Weathering With You and Suzume made by Makoto Shinkai.
Recommended generation parameters
Weight: 1.0 (adjust down as needed when used in conjunction with other LoRa's)
Model: SDXL 1.0 with 0.9 vae
Steps/Sampler/CFG/Resolution: 25 steps | DPM++ 2M Karras | CFG 7 | 1536x1536
'Kohya Deep Shrink' settings: 4 blocks | 2 downscale | 0 start | 0.35 end | skip = true
'CLIPTextEncodeSDXL' ComfyUI node resolution: 4096x4096
Positive prompt: high-definition 8k
Negative prompt: text, blur, low-resolution, low-quality, grainy, glossy, poorly drawn, amateur, poorly shot, compression artifacts
Trigger words (must be included in the prompt for the LoRa to take effect!)
'anime screencap in mnst artstyle'
Further usage notes
/
Usage permissions
You may not sell this model, merges containing it, or images generated with it.
Credit me when using this model in a merge or hosting it on a generation service.
Changelog
v1.4 'Skynet'
new training config leading to less overtraining, more flexibility and diversity
naming each version from now on after the training config used
for more info see my training guide linked at the top
v1.3
complete rework; higher epoch count, trained with EMA, reworked captions, reworked dataset preparation, tripled dataset count and added images from Suzume - overall much better likeness and quality at the cost of some slight bias and overtraining increase - final version for now
v1.2
retrained with new captions: slightly better style likeness and flexibility, less bias towards the subject matter of the training data; no longer as prone to generate young characters
v1.1
Initial version felt slightly overtrained, so I trained a new one and extracted epoch 90 instead of 100. Prompts should output slightly more diverse, coherent, and less biased content now compared to 1.0.
v1.0
Initial release