SoulTide 灵魂潮汐 Silenus 西勒诺斯
时装和纯样式提示词请参考标准提示样例图片
For fashion and pure style prompts, please refer to the standard prompt sample pictures
Description
FAQ
Comments (5)
You seem to be releasing a lot of different loras, I've noticed, and I want you to continue to do so and be successful at doing so. So I would suggest a tip; You should try to train at conventional dim and conventional alpha at 8/4 respectfully so as you have a much smaller file size without losing any quality in your gens.
Smaller file size makes it more enticing for viewers to want to download more of your models. More people that download your models equals more popularity for you. Best of luck in your future ai endeavors!
Thank you for your suggestion!
I've actually been using 32/32 for quite some time now. I’m aware it results in a larger file size, but I haven't tested the quality of models trained with lower dim/alpha parameters yet. Before successfully releasing at least one stable version, I prefer to stick with this configuration as it’s something I know works well. For my own use, I find the current size (around 200MB) to be manageable.
If I decide to update the versions in the future, I might consider adjusting the parameters. However, for now, both ILV1.0 and PonyV3.0 will continue using the current configuration.
Looking ahead, do you think we might eventually transition away from LoRA to Dora/LyCORIS? Several studies suggest these newer methods outperform LoRA in many use cases, but I’m curious about your thoughts—have you explored them, or do you think they offer enough advantages to justify the switch?
I do, actually I believe many already use lycoris, it's just the civitai system still labels it as lora, unless you throw it in a metadata reader. There's a clear difference between the two in quality. Though for most characters it isn't needed too much, but lycoris will help with minute detail on clothing and such. Dora... I don't know TOO much on it, but it's more to learn and understand and most people don't like leaving their comfort zone of what works and what doesn't, so I see lycoris becoming more widely used, but not Dora.
I understand your reasoning for continuing to do it your way of 32/32, as back since 1.5 that's been the tried and true, but it's just only recently been discovered that it's unneeded. If you ever have the time you can always try and see for yourself with exact same parameters and only changing the dim and conv size. I've been a 'collector' of sorts of lora since the beginning of this bleeding edge technology and I too didn't care too much it was 217mb... but now as I keep collecting them, definitely starting to run out of space along with 7gb models and control-nets, video models, large language models and such. lol... it eats space very quickly. So it is definitely me trying to save space and get more models collected without having to sacrifice hard drive space, but also I know not everyone is aware of the fact it's even an option and I try to pick random people on here and explain such to them, then it's up to them.
@Light7799
I recently conducted a comparison under a 100-epoch training condition (with approximately 5000 global steps for 1 batch, selecting results with the lowest loss between steps 80-100). I tested both 16+16+4 (100MB) and 32+32+4 (200MB) Lycoris models.
Using the same seed, I generated images for an XYZ chart with 24+24 samples per group.
At a strength of 1.0, the 32+32+4 model showed significantly better adherence to the training set in terms of background details, though the difference in character clothing was minor.
At my usual strength of 0.7, the differences in clothing became more pronounced.
To rate them: the 16+16+4 set had 8 images where features were not well-represented, compared to just 1 such image in the 32+32+4 set.
I repeated the experiment with three different -1 seeds and reached the same conclusion.
It seems smaller-dimension models may not lose performance at lower steps, but larger models still exhibit greater potential for understanding the dataset at higher steps.
@zwa73 Interesting. Though I can see this being a case by case basis then if the larger mb lora gives better background, as some want just the character and I've even been told if the background starts to bleed into gens from the character dataset then it's slightly over-trained (I don't think this is particularly true, myself).
Though knowing this then I can see an argument for using a larger dim size when wanting to get the character pretty much spot on in the media that it exists in, say for like a genshin character, where you'd WANT the background to be cohesive with the character, but if you're just wanting the character per-say, and you don't need it in the exact media it's taken from, lower dim size works best... I see, I appreciate you informing me on your experiment and outcomes! I will keep this in mind, for sure.