This LoRA was trained on 550k images of normal to hyper sized anime characters. It focus mainly on breasts/ass/belly/thighs, but now handles more general tag topics as well.**If you're wondering where the v8 LoRA is, read the <a target="_blank" rel="ugc" href="https://civarchive.com/articles/3320">Changelog Article</a>.**Also a back up <a target="_blank" rel="ugc" href="https://huggingface.co/thojm/HyperFusion">HuggingFace link</a> for these modelsUploaded 1.4 million custom tags used in hyperfusion <a target="_blank" rel="ugc" href="https://huggingface.co/datasets/thojm/hyperfusion_classified_tags_export">here </a>for integrating into your own datasets Recommendations: negative (depends on base model, use what works for you): <code>lowres, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, (worst quality, low quality), normal quality, jpeg artifacts, signature, watermark, username, blurry, monochrome, [3d], comic, (sketch), transparent background, artist name</code> cfg: 9-12 resolution: 768 to 1024, but prefers to be closer to 768 clip skip: 2sampler: With v7+ try to avoid Karras samplers. Training with --zero_terninal_snr makes karras samplers problematic, but they still work somewhat.<h2 id="changelog-article-link-yjoglzv7l"><a target="_blank" rel="ugc" href="https://civarchive.com/articles/3320">Changelog Article Link</a></h2><h3 id="tag-info-(you-definitely-want-to-read-the-tag-docs-see-:training-data)-teknn3are"> Tag Info (You definitely want to read the tag docs, see :Training Data)</h3> Because hyperfusion is a conglomeration of multiple tagging schemes, I've included a tag guide in the training data download section. It will describe the way the tags work (similar to Danbooru tags), which tags the model knows best, and all my custom labeled tags. For the most part you can use a majority of tags from Danbooru, Gelbooru, r-34, e621, related to breasts/ass/belly/thighs/nipples/body_shape.The best method I have found for tag exploration is going to one of the booru sites above and copying the tags from any image you like, and use them as a base. Because there are just too many tags trained into this model to test them all. <h3 id="tips-dhzs03vgw">Tips</h3><ul><li>If you are not getting the results you expect from a tag, find other similar tags and include those as well. I've found that this model tends to spread its knowledge of a tag around to other related tags. So including more will increase your chances of getting what you want.</li><li>Using the negative "3d" does a good job of making the image more anime like if it starts veering too much into a rendered model look.</li><li>Ass related tags have a strong preference for back shots, try a low strength ControlNet pose to correct this, or try one or more of these in the negatives "ass focus, from behind, looking back". The new "ass visible from front" tag can help too.</li><li>...more tips in tag docs</li></ul><h3 id="extra-32zf2s09s">Extra</h3> This model took me months of failures and plenty of lessons learned (hence v7)! I would eventually like to train a few more image classifiers to improve certain tags, but all future dreams for now.As usual, I have no intention of monetizing any of my models. Enjoy the thickness!<hr /><h3 id="training-hurdles-ir7pqkf6k">Training Hurdles</h3>-Tagging-The key to tagging a large dataset is to automate it all. I started with the wd-tagger (or similar danbooru tagger) to append some common tags on top of the original tags. Eventually I added an e621 tagger too, but I generally only tag with a limited set of tags and not the entire tag list (some tags are not accurate enough). Then I trained a handful of image classifiers like breast size, breasts shape, innie/outie navel, directionality, motion lines, and about 20 others..., and let those tag for me. They not only improve on existing tags, but add completely new concepts to the dataset. <s>Finally I converted similar tags into one single tag as described in the tag docs</s> (I stopped doing this now. With 3m images it really doesn't matter as much).Basically any time I find its hard to prompt for a specific thing, I throw together a new classifier, and so far the only ones that don't work well are ones that try to classify small details in the image, like signatures.Starting in v9 I will be including ~5% captions along side the tags. These captions are generated with CogVLM.I used this to train my image classifiers <a target="_blank" rel="ugc" href="https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification">https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification</a>Ideally, I should train a multi-class-per-image classifier like the Danbooru tagger, but for now these single class-per-image classifiers work well enough.-Poor Results-For a long time I was plagued with sub par results. I suspected maybe the data was just too low quality, but in the end it just ended up being poorly tagged images. Sites like r-34 tend to have too many tags describing an image like "big breasts, huge breasts, hyper breasts" all on the same image. This is not great for a model where you want specific sizes. Using the classifiers I mentioned above I limited each image to a single size tag for each body part, and the results were night and day.2023/08/13 Coming back to this after even more experience in labeling/training, I still agree with the above statement. As I tag more and more images, the model becomes more reliable with prompts. You can see this clearly with the new bottomheavy, topheavy, bellyheavy tags. It makes it easier to generate specific body types, and helps the model understand what you want from your prompt. I didn't have to add additional images to make these tags work. I just improved tagging.-Tag Bleeding-An example of tag bleeding is using the tag "gigantic breasts", but you end up with everything being gigantic, breasts, ass, thighs. It's been an annoying problem.2024/03/15 After training some larger models, I find bleeding is less of an issue. It seems the answer is more data.-Testing-In order to determine if a new model is better than the last, it's important to have some standard prompts that you can compare with. x/y plot is great for this. Just keep in mind that the seeds between models will be totally different, and you likely need to compare dozens of images at a time and not 1 to 1.It's also important to compare new models against the base model output to make sure what you are training is actually having an overall positive effect compared to the original model (obvious, but often overlooked).2023/08/13 The hardest part about testing is trying to determine when you have overcooked your text encoder. At some point the models ability to interpret the prompt starts to degrade when training the text encoder at a high enough LR. I've seen it happen with smaller and larger models alike. Unfortunately I still don't have a good way to test this other than comparing prompts to older models. I've tried training without the text encoder many times, but the results are always sub par or too slow to train. Concepts that are foreign to the base model, are so much better understood with TE enabled in training.-Software/Hardware-The training was all done on a 3090 on Ubuntu. The software used is Kohya's trainer, since it currently has the most options to choose from.

hyperfusion LoRA 550k images

Added ~60k more imagesLots more image classifiers trained, and existing classifiers improvedTraining Notes:<ul><li>~300k images from Danbooru, r-34, e621 etc...</li><li>LR 2e-4</li><li>TE_LR 5e-5</li><li>batch 4</li><li>GA 32</li><li>dim 64 (resized with kohya, with sv_ratio=10, and max dim 64)</li><li>conv_dim 32</li><li>alpha 32</li><li>conv_alpha 32 (I meant to set this to conv_dim/2, but ohh well)</li><li>scheduler: cosine_with_restarts, 33 restarts</li><li>base model NAI</li><li>No --vae</li><li>flip aug</li><li>clip skip 2</li><li>225 token length</li><li>bucketing at 1024 max 1024 (I should have stuck with max_bucket_size=768 for reduced training time)</li><li>tag drop chance 0.1</li><li>network_dropout 0.33</li><li>scale_weight_norms 1 (either this or dropout resulted in less blown out results, which allowed for longer training at higher LR)</li><li>tag shuffling</li><li>50,000 unique tags, but about 500 frequently occurring ones (see attached training data for tag list)</li><li>about 14 days training time (slower than last because of bucket size choice)</li></ul>

Files with SHA256: 24405c29500cf42e1b2a1939055e550401ae81b850d38aea8c6fb71030ff38d9