Trained on a small dataset, mainly used to learn the training environment of Flux.
Try online before download: https://tensor.art/models/xe-anime-flux
It seems that I still can't understand the concept of nsfw
Next step is to try training on a larger hentai anime dataset
Description
FAQ
Comments (10)
Well.. if you train for 251 steps on one epoch... you cannot expect much... how many pics was your dataset?
But it's fine if you are trying to pry into the waters of training... I would recommend more epochs though if you want to see changes... finetunes usually have datasets that consists of 1K - 10K pictures if not more
My local hardware cannot meet the training requirements, so I use a rented graphics card.
I need to estimate how much data is needed based on the rental time.
Next, I plan to use a spot instance for about 6 hours of training.
I tried a dataset of over 25000 danbooru sourced pictures with terrible results.
Tags heavily overlap and cause everything to bleed together.
Seriously good luck. You may have a better process than I used, but I don't plan to retain that size again until I'm certain it'll work.
@AbstractPhila I think tags with flux need to have a reference. You can tag an image with something, but (especially with nsfw stuff) flux has no idea what it is. Like if you tag 'pussy' it may not be as effective as if you say 'the pussy is located between her legs'. The T5 actually did that for me on a prompt, after I gave it a definition it was better able to understand a simple unattached tag from a lora.
I think auto booru taggers are going to not have great quality with flux.
Could you drop more sample generations?
You can experience it online: https://tensor.art/models/769916937011682047
There is a ~21 hours task on my device, and I can't provide more examples for the time being.
@XEZ it says the model is not public
May i ask if it's possible to upload the checkpoint as a single file, with the text encoders and VAE embedded in it?
It's inefficient but conventional. xD
(Thx for the upload btw!)
This will only be done in official versions after 1.0, because I don't label any anime character, so there's no point in training CLIP.
The main problem is that there is no tagger that can identify anime characters. I heard that the author of pony developed one using GPT-4o.
@XEZ Ok no worries! Keep it up m8. Looking really good, and relatively consistent compared to the competition too.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.