这个模型使用了280万的图片进行训练
目前它还并没有训练完成 预计还需要20天训练完成
目前发布的仅是它训练进度1epoch-step47%的检查点保存
不过即使它没训练完成，它的表现似乎也出乎了我的意料
大部分的我常用的画师风格它都能很简单的复现出来，
不过依然有欠拟合的问题（因为它连第一个epoch都没有训练完成）
它基于animagine-xl-v31经过12次全量微调后的基础模型
不过依然保留了animagine的大部分角色
不过理论上并不兼容animagine的lora模型-（当然，也不建议你这个时候进行训练就是了）
这个模型的主要方向是多个画师风格，还有nsfw图像，还有更多图像
它的质量词为Best-A,Medium-B,Low-C,Bad-D当然，你也可以使用animagine的质量词，
模型的词频统计在模型卡上，模型内置了约1700位强画风风格（单独进行加强训练的）
请等待20天后的第一个epoch版本。
The model has been trained using 2.8 million images.
Currently, the training is not yet complete, and it is expected to take another 20 days to finish.
The currently released version is only a checkpoint saved at 47% progress of the first epoch.
However, even though it hasn't completed training, its performance seems to have exceeded my expectations.
It can easily reproduce most of the art styles I commonly use.
However, there are still underfitting issues (because it hasn't even completed the first epoch of training).
It is based on the animagine-xl-v31 model after 12 rounds of full-scale fine-tuning.
However, it still retains most of the characters from animagine.
Theoretically, it is not compatible with animagine's lora models (of course, it is not recommended to train at this stage either).
The main focus of this model is on multiple artist styles, NSFW images, and more images.
Its quality tags are Best-A, Medium-B, Low-C,Bad-D Of course, you can also use animagine's quality tags.
The model's word frequency statistics are on the model card. The model has about 1,700 strong art styles built-in (with separate intensive training).
Please wait for the first epoch version after 20 days.
Description
FAQ
Comments (9)
I know it's still early, but any chance of an artist list for this? Model has a ton of potential so I've got my eye on it, would be nice to know which artists is in it so I don't end up re-training them as LoRAs with the full release if good enough.
Keep up the great work either way!
It's still in training, and most of the artists' styles don't fit very well at the moment. After the training is complete, I might train it for another epoch (depending on whether my financial resources allow it).
Regarding the list of artists, I will provide it later. However, don't have too much hope for this model.
You're pretty much the only one doing a proper fine tune on animagine as far as I can tell. The only way I will lose hope is if you tell me it's a LoRA merge, but I don't think it is from the training config you shared before on 220k. The biggest problem with animagine is it's undertrained but it's so full of potential so any amount of training to it is a positive thing in my book.
@Miyuutsu I don't know how you figured out that I was using full fine-tune or LoRA, but I need to tell you that both full fine-tune and LoRA are essentially doing the same thing: fine-tuning.Whether it's this model or the 220k version, the main training was done using the LoRA method, but with a very large DIM. Although the base model might have undergone full fine-tuning, the effects you are seeing now are all achieved through LoRA, which is fundamentally no different from full fine-tuning.For the next epoch of training, I might use full fine-tuning (since I now have a spare A100 80G GPU). However, personally, I still prefer LoRA because it's faster and works well.
@Miyuutsu Compared to the training algorithm, the dataset used for training might be more important. Theoretically, LoKR would be more suitable for this million-level training task. However, the purpose of my current version is to train a good base model so that my subsequent fine-tuning won't be as labor-intensive. Once the complete version of this model is trained, I will decide whether to continue training and may use LoKR or full fine-tuning. Of course, my training parameters are quite unusual and counterintuitive, but their performance is very good.
The 'config' file in the model card has been updated to include the 'artist list' and 'word frequency statistics'.
From testing, the current version of the model has a tendency to hallucinate sweat. It is recommended to refine the training data and specify this detail.
I did not encounter this issue during testing. Could you share the images you generated so that I can more accurately identify the problem?
The current model training is not yet fully completed; it will take another 400 hours (until July 25) to finish the first epoch. However, once the training is complete, I will use two A100 80G GPUs to fine-tune for another epoch using the SDXL-Finetune algorithm instead of the current version's LoRA algorithm.
The reason I didn't use the SDXL-Finetune algorithm from the beginning is that I didn't have two A100 80G GPUs at the time. Therefore, I trained on a 4090 GPU and used the LoRA algorithm to save memory and achieve higher performance.
The current training can be understood as laying a foundation for subsequent training, building a base model. Of course, this will be a matter for a month from now.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.