Model Info:
Anime screenshot composition style for Anima
Triggered by "@screencap_(anime), screenshot from anime TV series,". Strengh 0.85
Usage Recomendation:
Effect: Lora change image composition closer to anime screenshot. It is not consistent in style and probably eats out some details (because anime usually not so detailed as art) but bring a more lively look to you images
UPD: epoch 4 and epoch 9 on new dataset.
IMPORTANT NOTE ABOUT DATASET: Trained on 4,600 images, both modern anime in 16:9 and 21:9 and a small number of older anime in 4:3. Up to 1536px (so it should better handle Hires).
Dataset does not specify particular anime or studio styles. However, it is mostly consist of high-quality data from Chainsaw Man (MAPPA), Violet Evergarden Movie (KyoAni), and Monogatari Series (SHAFT). Therefore, in terms of lighting, detail, and composition, the results will generally resemble Chainsaw Man or Bakemonogatari, but with more detailed hair and clothing.
Because the dataset also includes older anime (20–30%), the level of detail will sometimes simpler, similar to that of K-on.
IMPORTANT NOTE ABOUT VERSIONS: Epoch 9 trained twice more than Epoch 4.
EP4: better preserves the the pose, anatomy, and backgrounds because it is less baked and has less influence on the model's prior knowledge. It also has slightly richer colors, but the composition is simpler, and the “screenshot-like” quality of real anime is conveyed to a lesser extent. However, this version was able to learn the cinematic black horizontal bars of the 21:9 frame and occasionally uses them.
EP9: much better captures the “screenshot” quality, much more closely resembling actual frames from the anime. This is good in terms of lighting and frame composition, but sometimes worse in terms of detail (you know the quality of non-important parts in anime, right?). So, simotimes the colors, backgrounds and even the character limbs are less detailed. The model DOES NOT DRAW (or draws very rarely) black bars, and overall, in terms of cinematic quality, this model is better.
In the Preview, for both checkpoints, I selected identical generations on the same seed, so they can be compared directly visually. In the additional illustrations in the comments, I will add less successful (and less SFW) examples for both checkpoints.
P.S. "Horny" stuff brings up simple old anime style more often because "horny" animation usually have lower quality and model remember it. I can't help with it.
Description
FAQ
Comments (6)
You might be able to boost the quality by tagging the screenshots by the decade, or giving the older screencaps an unique tag to negative them out
The same goes for tags for specific animation studios. I know.
The challenge is that LoRA is not a full finetuning, and mixing different key tags (different concepts) within a single LoRA is not a good idea.
At best, it will ignore such tags or cause the concepts to bleed into one another (i.e., the model will differ little from its current state); at worst, it will degrade the overall quality of the model by hindering generalization.
4,600 images is already too much for Lora with Rank 16 (and a higher rank will start to memorize not only the concept but also the content of the images, degrading the model's general knowledge).
Hi.
base-v1.0 ver please.
我使用这个lora一直出现赞助者水印，使用负面词也无法阻止这种他出现，我该怎么办
具体是什么水印？数据集的大部分内容取自无水印的BD（蓝光）资源，或者（在Patreon上的18+视频片段的情况下）经过了去水印和清理处理。创作者的水印痕迹最多存在于10%的图像中，且位于难以清理的位置，这些图像都已进行了相应的标注（captioned），按理说不应该产生稳定的水印。可能是你的提示词（prompt）中有某些内容触发了水印的出现？
我目前正在完善用于1.0版本微调的数据集，这可能会改善情况。
@desm0nt 我该怎么把图片发给你？他是文字水印，黑色的大字母
字样往往是：
PATRO-N'DOTREOSY
Malatreor PATIFO ONLY FONEOST
PATROCIY DOTROST
上面那些是我用QQ的提取文字将那些文字提取出来。