Warning: includes both safe content and unsafe content.
This model was trained by me with Dreambooth on 2859 pictures and then merged with 28 other photorealistic models.
Works best on realistic portraits, but also knows some other styles.
All examples are generated without face restoration and hires fix and they don't seem to be necessary, but you can use them when things get really bad. Negative embeddings work well too, like NG_DeepNegative_V1_75T with 0.8 strength.
The faces may be similar, but there is definitely no overtraining on a particular face here.
Contains a noise offset.
Contains a moderate number of booru tags.
This model involves dreamlike photoreal, so here is the license that you must abide by.
警告：包括安全内容和不安全内容。
这个模型是我用 Dreambooth 对 2859 张图片进行训练，然后与其他 28 个逼真的模型合并。
在写实的人像上效果最好，但也知道一些其他风格。
所有的例子都是在没有 face restoration 和 hires fix 的情况下生成的，它们似乎没有必要，但当事情变得非常糟糕时，你可以使用它们。Negative embeddings 也很好用，比如 NG_DeepNegative_V1_75T ，强度为 0.8。
这些脸可能是相似的，但这里绝对没有对某张脸进行过度训练。
包含一个 noise offset。
包含适量的 booru 标签。
这个模型涉及 dreamlike photoreal，所以这里是你必须遵守的 license。
Предупреждение: включает как безопасное, так и небезопасное содержимое.
Эта модель была обучена мной в Dreambooth на 2859 фотографиях и затем объединена с 28 другими фотореалистичными моделями.
Лучше всего работает на реалистичных портретах, но также знает некоторые другие стили.
Все примеры сгенерированы без face restoration и hires fix, и они не кажутся необходимыми, но вы можете использовать их, если дела совсем плохи. Негативные эмбеддинги тоже хорошо работают, например, NG_DeepNegative_V1_75T с силой 0.8.
Лица могут быть похожи, но здесь точно нет переобучения на конкретное лицо.
Содержит noise offset.
Содержит умеренное количество тегов booru.
Эта модель включает в себя dreamlike photoreal, поэтому вот лицензия, которую вы должны соблюдать.
Description
Inpainting weights
FAQ
Comments (4)
is this f333 in a noise offset high contrast direction ?
And most importantly you are back ?
The Asian Bias here looks high much higher then used from f222
No, f222 is not present here at all.
This is my first checkpoint, in which I made a lot of mistakes, so I don’t think there is any point in continuing to suffer with it.
yeah but by now it got pretty much merged into everything partly inter generational tests show this clearly, how its weights jumped around from one to the next merge
And i use it very often to create another Kind of reality more old analog Video focused even up to this day and compare it vs most current DSLR realities.
And the difference is often surprising
it still has it's very unique view of reality in some cases, whereby others look more and more the same, though some are on the way back it seems :)
i find f222 very natural very pure it highly focuses on keeping high frequencies alive even taking artifacts into account but when pushing it through further upscaling it doesn't matter so much, but it can retain unbelievable perceptual details for medium distances.
Where a lot of others Models just flatten out in Artificialness
it tries not to go so heavy on saturation as other models we can argue of course about it it wasn't neither contrast shock focused as when noise offset became popular of course perceptually contrast is important but it's so much overdone now.
but we can clearly see models who try to go away from artificial lens to EYE perception as well now again
it's just sad if you compare a photo of this century sometimes and how many details get lost vs some Sony BetaCam Recordings and it's VHS transfers.
And you sometimes wonder for what did you do all those resolution research and preservation of high frequencies when everyone fights it artificially ?
@openfire You can see that I have another photorealistic checkpoint on version 2.1 768, even though its dataset is more than three times smaller, but it behaves more predictably, and I use its much expanded version in my daily work, which most likely will never be posted on civitai, but maybe there is a chance that I will someday post at least a trimmed version of the expanded version