E621 dataset?
While I agree that it makes sense to include something like Ye-Pop for general knowledge, I don't see any benefit in using the DeviantArt dataset. If you're going to add additional data that will distort the model's aesthetics (and the DeviantArt data certainly will), why not add something that might introduce new concepts into the model? (For example, the E621 dataset, which has been proven to improve the NSFW aspect of NoobXL models, compared to plain Illustrious models). Given it's properly tagged to not bleed into the model, of course.
I think 2B model is kinda limited in term of dataset people can cram in it. Once it have a decent base model, people can bake their own loras
Deviantart has no place in an anime model, photos could potentially help with stuff like backgrounds. e621 at least has something good going for it.
e621 is poison
I wouldn't say E621 is straight poison if tagged and filtered properly. There are some useful tags that only E621 offers and it does somewhat help with NSFW, particularly with genitals and camera angles. I feel like that viewpoint towards including E621 in models is a holdover from Pony when that awful fried thing was still being used. Bottom line though, Deviantart definitely has no place in the model, and I can't even think of a single positive thing it would bring overall. If anything else needs to be added besides Danbooru and ye-pop, it should just be E621, not Deviantart.
+1. Given that NoobXL was very successful I think it would be worth a try. There's a lot of potential.
+1. Given that NoobXL was very successful I think it would be worth a try. There's a lot of potential.
noobxl dropped E621 later and gained from it.
+1. Given that NoobXL was very successful I think it would be worth a try. There's a lot of potential.
noobxl dropped E621 later and gained from it.
Which Noob models dropped E621 later on? I'm not sure about the earliest EPS models (they aren't public I believe) but I do know that 0.5 onwards had E621 in the dataset, though it wasn't really prevalent until they started training the TE in subsequent versions. But honestly, it's hard to say if E621 was even the sole reason for Noob's issues. We don't have any fully trained Noob model that never had the TE trained or a fully trained Noob model without E621 entirely for reference, and I'm not sure if the E621 tags were even aliased to Danbooru tags when the TE training started.
Either way, I do believe that Anima should just be anime-only, but I would much rather take E621 over Deviantart if something really has to be added. The quality of data and tagging is undoubtedly far better than whatever Deviantart would have anyways.
+1. Given that NoobXL was very successful I think it would be worth a try. There's a lot of potential.
noobxl dropped E621 later and gained from it.
Which Noob models dropped E621 later on? I'm not sure about the earliest EPS models (they aren't public I believe) but I do know that 0.5 onwards had E621 in the dataset, though it wasn't really prevalent until they started training the TE in subsequent versions. But honestly, it's hard to say if E621 was even the sole reason for Noob's issues. We don't have any fully trained Noob model that never had the TE trained or a fully trained Noob model without E621 entirely for reference, and I'm not sure if the E621 tags were even aliased to Danbooru tags when the TE training started.
Either way, I do believe that Anima should just be anime-only, but I would much rather take E621 over Deviantart if something really has to be added. The quality of data and tagging is undoubtedly far better than whatever Deviantart would have anyways.
Noob basically abandoned E621 after around eps ~0.75 (internal builds). Of course, I think adding some general datasets was necessary. Anima is closer to reality in terms of photo backgrounds and image composition, and the quality is better than Noob.
+1. Given that NoobXL was very successful I think it would be worth a try. There's a lot of potential.
noobxl dropped E621 later and gained from it.
Which Noob models dropped E621 later on? I'm not sure about the earliest EPS models (they aren't public I believe) but I do know that 0.5 onwards had E621 in the dataset, though it wasn't really prevalent until they started training the TE in subsequent versions. But honestly, it's hard to say if E621 was even the sole reason for Noob's issues. We don't have any fully trained Noob model that never had the TE trained or a fully trained Noob model without E621 entirely for reference, and I'm not sure if the E621 tags were even aliased to Danbooru tags when the TE training started.
Either way, I do believe that Anima should just be anime-only, but I would much rather take E621 over Deviantart if something really has to be added. The quality of data and tagging is undoubtedly far better than whatever Deviantart would have anyways.
Noob basically abandoned E621 after around eps ~0.75 (internal builds). Of course, I think adding some general datasets was necessary. Anima is closer to reality in terms of photo backgrounds and image composition, and the quality is better than Noob.
I see, it's unfortunate though that it'll never truly be known if it was the TE training that resulted in issues since the dataset was dropped so soon (I think 0.75 eps was also when they froze the TE again? Though I could be confusing it with the vpred models). The quality of Anima so far is definitely better than Noob when it comes to colors and backgrounds, which I'm guessing part of that comes from ye-pop, as a big issue with (base) Noob was that its background capabilities were nonexistent lol. I do hope something bigger than a 0.6b LLM is used for the final model though, it probably won't happen but even something like 4B would just be a small penalty to prompt processing with offloading and would still fit in 8GB VRAM too.
+1. Given that NoobXL was very successful I think it would be worth a try. There's a lot of potential.
noobxl dropped E621 later and gained from it.
Which Noob models dropped E621 later on? I'm not sure about the earliest EPS models (they aren't public I believe) but I do know that 0.5 onwards had E621 in the dataset, though it wasn't really prevalent until they started training the TE in subsequent versions. But honestly, it's hard to say if E621 was even the sole reason for Noob's issues. We don't have any fully trained Noob model that never had the TE trained or a fully trained Noob model without E621 entirely for reference, and I'm not sure if the E621 tags were even aliased to Danbooru tags when the TE training started.
Either way, I do believe that Anima should just be anime-only, but I would much rather take E621 over Deviantart if something really has to be added. The quality of data and tagging is undoubtedly far better than whatever Deviantart would have anyways.
Noob basically abandoned E621 after around eps ~0.75 (internal builds). Of course, I think adding some general datasets was necessary. Anima is closer to reality in terms of photo backgrounds and image composition, and the quality is better than Noob.
I see, it's unfortunate though that it'll never truly be known if it was the TE training that resulted in issues since the dataset was dropped so soon (I think 0.75 eps was also when they froze the TE again? Though I could be confusing it with the vpred models). The quality of Anima so far is definitely better than Noob when it comes to colors and backgrounds, which I'm guessing part of that comes from ye-pop, as a big issue with (base) Noob was that its background capabilities were nonexistent lol. I do hope something bigger than a 0.6b LLM is used for the final model though, it probably won't happen but even something like 4B would just be a small penalty to prompt processing with offloading and would still fit in 8GB VRAM too.
But here's the worst part right now: in the model's current arch (Qwen3 0.6B + LLM Adapter (may be trained on T5-generated data) + MiniTrain DiT), the LLM Adapter not only seems to severely cripple Qwen's capabilities, but also underperforms the T5 TE. On top of that, this model is still using the T5 tokenizer — that's a huge problem.
@FHFanshu
Who told you, or where did you learn, that E621 data was abandoned in noob after EPS 0.75?
I'm also curious, I've seen some people say they trained on less images than they originally started with (obviously besides the fact that it was heavily curated in the first place), but this is also the first time I've heard of it being completely dropped after 0.75. Which would be strange, since 0.75 is when it began actually showing E621 concepts and tags, and that knowledge stayed pretty consistent even up to EPS 1.0 and in the Vpred models. Never heard of it just being straight up dropped entirely though
@FHFanshu
Who told you, or where did you learn, that E621 data was abandoned in noob after EPS 0.75?I'm also curious, I've seen some people say they trained on less images than they originally started with (obviously besides the fact that it was heavily curated in the first place), but this is also the first time I've heard of it being completely dropped after 0.75. Which would be strange, since 0.75 is when it began actually showing E621 concepts and tags, and that knowledge stayed pretty consistent even up to EPS 1.0 and in the Vpred models. Never heard of it just being straight up dropped entirely though
I must have remembered it wrong.However, the subsequent versions of NoobAI's V-PRED have encountered issues with instruction following and concept coupling for generation tasks within certain scopes, and I personally find them quite difficult to use.
This is the user documentation for Noobai. https://d0xb9r3fg5h.feishu.cn/docx/WWOHdr6RMoQZxQxCZRGc5KlEnUi
@FHFanshu
Who told you, or where did you learn, that E621 data was abandoned in noob after EPS 0.75?
和你这种蠢逼说话费劲，懒得翻译成英语了。你但凡多看看noobXL中文社区里那些训练人员自己说的话，都不会坚信e621的数据集从头到尾参与了训练
同时，但凡你有一个正常的测模型水平，也能感觉到noob后续版本里e6相关数据的劣化😅
但你屁都不知道，也不了解e6数据训练后有多么拉低美学质量。甚至你都不愿意看看使用手册里，e621的使用建议是作为负面prompt写进去提升出图质量，这么喜欢e621，你可以去玩ponyV7
@FHFanshu
Who told you, or where did you learn, that E621 data was abandoned in noob after EPS 0.75?和你这种蠢逼说话费劲，懒得翻译成英语了。你但凡多看看noobXL中文社区里那些训练人员自己说的话，都不会坚信e621的数据集从头到尾参与了训练
同时，但凡你有一个正常的测模型水平，也能感觉到noob后续版本里e6相关数据的劣化😅
但你屁都不知道，也不了解e6数据训练后有多么拉低美学质量。甚至你都不愿意看看使用手册里，e621的使用建议是作为负面prompt写进去提升出图质量，这么喜欢e621，你可以去玩ponyV7
Who was this supposed to be directed towards? I'm not sure what they said wrong, it was just a question asking if E621 was actually dropped past EPS 0.75. It is still true that the exact issue for Noob is hard to say; tag dropout likely wasn't even used, we don't know about tag aliasing which also leads into the lack of actually tagging the E621 dataset as a whole with "e621" (furry is also a tag on Danbooru after all), and there's the whole TE training issue. Can't really speak on EPS 1.1 or Chenkin though, never used them.
The fact is though that it's irrelevant how exactly Noob turned out or what caused the degradation; the main point is that Deviantart makes no sense to be in Anima. The data is highly questionable and who knows how it was even tagged, compared to something like E621 where at the very least it would have actual tagging and higher quality data. Yes, it would obviously be better to be Danbooru only, but there was clearly some motivation to include an extra dataset, and there's no reason to use such a poor one instead of one that's actually tagged and could improve generations in some cases. At the very least, Anima is tagging the extra datasets as a whole and tag dropout is being used.