3 ms·
If you like anime artists, there is this: https://rentry.co/NAIDv3artisttagtest https://rentry.co/NAIDv3artisttagtest, a list of 470 artists using the last mode
by GaggiX 3y ago
If you like anime artists, there is this: https://rentry.co/NAIDv3artisttagtest https://rentry.co/NAIDv3artisttagtest, a list of 470 artists using the last model of NovelAI (this time they trained their model without removing artist names).
Fun fact, this model is based on SDXL, finetuned probably on a cluster of H100s (from coreweave).
- Der_Einzige 3y agoThe consensus seems to be that the lack of open source SDXL anime fine tunes is simply due to laziness and lack of resources from the interested parties in the community. Datasets like from the boorus are already floating around online to train on. Always funny to see that furries are ahead of weeaboos in the war for the best open source nsfw generator
- Aeolun 3y agoI think it’s more that SD 1.5 is already fine for most purposes (and much, much faster) Just don’t see any difference in results except them taking 4x longer.
- ttul 3y agoIt’s an interesting example of a new technology that is inferior to the old one because the old one is superior in important ways (speed, ease of manipulation, and established toolsets).
- GaggiX 3y agoWell to be honest I don't really know what's the best furry SD XL model out there but recently this model has been released: https://huggingface.co/cagliostrolab/animagine-xl-3.0 https://huggingface.co/cagliostrolab/animagine-xl-3.0, it's probably far from the NovelAI checkpoint but I would be curious to know how it compares to the furry model you're talking about.
- echelon 3y ago> Always funny to see that furries are ahead of weeaboos in the war for the best open source nsfw generator That's hilarious. It'd be funny if they had a rivalry going and it resulted in better open source tooling.
- numpad0 3y agoI think that's just reflecting different levels of ethics standards - a bit like different stances on street graffiti
- makomk 3y agoLast I heard, everyone was struggling to fine tune SDXL on furry art too. The only community fine tunes that seem to have succeeded are ones aimed at taking things it was already designed to generate and trained on and making it actually good at those, or improving the aesthetics of its output.
- ben_w 3y agoThat's mildly surprising, given how large and well tagged e621 in particular is. Though I assume most of the major furry art hosts are blocking scraping by this point, given how divisive GenAI is?
- GaggiX 3y agoThe dataset is not the problem, neither with anime or furry art, the problem is the computation required for a good finetuning.
- brucethemoose2 3y agoLots of people (including potential trainers) struggle with SDXL I think. It just doesn't give you good results if you shove it into SD 1.5 frameworks and prompt it like SD 1.5.
- GaggiX 3y agoIf you're talking about anime, every popular SD v1.5 model out there has a backbone based on the leaked NovelAI model, so what SD XL lacks the most is people with compute power finetune it, that being said, this model: https://huggingface.co/cagliostrolab/animagine-xl-3.0 https://huggingface.co/cagliostrolab/animagine-xl-3.0 was recently released and it was finetuned for +500 A100-hours.
- Jackson__ 3y agoThe devs behind NovelAI have been talking about their training on the discord a little bit. According to their messages, it took about the same amount of compute to finetune the model, as it took for StabilityAI to train SDXL itself. No wonder other people struggle to train it, if it takes >=128 A100s to get the model to perform better than SD1.5 on only slightly OOD tasks (anime pictures, in this case.)
- sorenjan 3y agoIs there a big difference in the result from fine tuning a model compared to using a LoRA? I thought the idea with LoRAs was that updating all model weights are unnecessary?
- GaggiX 3y agoWell this is a very extensive finetuning, LoRA are usually convenient if you just want to add a character, a style or a concept to the model.