santa hat
deerdeer nosedeer glow
Sign In

Why this lora can not bring good result? 这个lora为什么寄了?

Jun 28, 2023

Warning: I'm not native English speaker. There might be impolite, ironic, rude and offensive words in this article.

Quality of Lora

Making lora for a single character's single costume is not that hard as long as you can find enough suitable images as datasets. Well though most people can learn how to cook instant noodles easily, there are still kitchen-destroyers. I do not recommend to give a image too many repeats. Just image that you just one click to enable the lora and do not have to set it 0.6 or even 0.4 to avoid ruining the picture. A slogan of a chain-restaurant in the city I lived in is "If you dare not serve your dishes to your parents, do not serve it to your costumers." If you find your lora hard to use even for yourself, do not publish it.

If this lora often brings some unwanted and even uncontrollable elements in results, it is likely to be "polluted" or have learnt something wrongly. You should check the tags in datasets, or clean these patterns by photoshop.

As we know that a lora is trained on a certain based model. Most anime loras are trained on NAI, though some trainers might forget choosing that option "other" when uploading. Anyway NAI is from SD1.5. Basically, according to a experiment carried out by Qinglong(, the recommended based model for anime lora is NAI. And merged models are not recommended.

Use Models Wisely

Models are like cells and SD1.5 is like a stem cell (NAI is like an "anime stem cell"?). Merged models are like…… cells which have mutated and evovled for serveral generations. They are excellent in doing something but inevitably lose the ability to do something too, and some errors/pollutions may have remained in their DNAs. They have their own thoughts. Someone might even think "What an amazing model! I just input 1girl solo and the outputs are fantastic!"——after trying more tags and prompts, he or she might finally find that this model actually does not respond to certain tags.

Someone might have seen wonderful AI coser images. However result of using anime character loras on photorealistic models are not always satisfying. As mentioned before, cells have evovled into different forms, if you are expecting that a kind of "enzyme" can be used by all the cells, you are really lack of scientific spirit.

Do we really need so many tags for quality?

Several months ago we are using NAI. At that time, tags like "masterpiece" lower down the possibility of The Old Ones coming out. But now, after trainning and merging, tags for quality are not necessary and even bring negative effects. Firstly, you do not know what effects they may bring, thus making the the output hard to control. For example, "detailed eyes" is likely to make colors of your characters wrong. Also, putting too many tags in prompt is like active ingredients being diluted by too much "water".

Oh, remember to check your negative prompt too. Actully nowadays a too long negative prompt won't give you a perfect result but even change the model's style greatly. Some old so-called universal negative prompts even have "succubus" and "french braid", isn't that confusing. I guess "succubus" was added when someone was using a kedama_milk style dreambooth model or embedding in order to avoid the unwanted demon wings.

The Two Geniuses

The first one takes one model for treasure, he regards it as a perfect test paper for loras and regards himself as a lora scientist. He likes using anime loras on his dearest photorealistic model and said "Hey this is not a good lora, it do not work well on photorealistic models. Come and thank me for my hard work on lora testing."

The second one is a "good student" who has a big notebook (or grimoire) full of spells. While The "good student" do not care about what these tags, loras and words of spells mean and where do they come from. He just open the grimoire and read, and even read here and there, thinking he is a great wizard who can mix magic spells easily. I even used to see someone used an eye close-up lora together with a clothing lora, then there is only an eye in the picture. Well did he really know what he was doing? Anyway I do not know.