santa hat
deerdeer nosedeer glow
Sign In

Example images should NOT use other embeddings/LORAs/etc

First I know this isn't really a question but it's the only area for open discussion here outside of comments.

It's very frustrating looking through example images for resources (like checkpoints, textual embeddings) and seeing LORAs used in all example images for some model/TI authors.

Your example images should showcase what YOUR embedding/checkpoint does on it's own, without the help of other's creations. I'm not looking at your embedding because I also want to download 4 other models to accompany yours at the same time (with the exception of whatever checkpoint you use because those are required standard base files.)

But 10 example images using 4 LORAs in every prompt to showcase your textual embedding/model tells us nothing about what your file achieves. If you can't generate high enough quality content for basic examples with your textual embedding/model by itself to show what YOU have created, then why upload it?

12 Answers

Absolutely, just like how a chef offering you a free meal should not use seasonings. If their meal is worth eating it should not need any.

You are saying that Coke ads shouldn't show the Condensation on the packs. Pizza shouldn't have any sauce since it isn't mentioned. Mostly all of the neg embeddings are in Civit AI if you search it.

While I mostly agree with this...

To be fair where do you draw the line? Does that also include textual inversions? and negatives?

Therein lies the problem.

What I would suggest is an area in the image upload to include information and an easy & searchable way to link to the lora's and extensions used in the image.

TI embedding aren't that easy to detect in a prompt and don't forget a lot of them are used for negative prompts. I totally get the point, but I don't find it realistic.

By using other LORAs I show that my style LORA is very flexible and can work in conjunction with other LORAs, both styles and characters. Often a mix of the two styles will produce great results.

Don't like it either that's why my LoRA are always standalone ... but live and let live 99% of all Modele and Neg are here also civitai

A good model will help LORA shine in the prompt. If the model can handle the extra "spice", then it could be more versatile. I think its fine how it is now. Restricting the uploads will be the death of civitai.

Mygirl embedding will work exactly the same as embedding in a prompt. Not sure if it will change result of a seed though, but we can have a rule to have embeddings as in prompts regardless, then it's easy to highlight them here with CSS or javascript

Most of my LoRAs are designed to be used with other LoRAs and TIs, because that's how pros use them--mix them with other great models to make amazing art. So I'm going to keep using as many other LoRAs and TI as I need to make my art. If you want an AIO solution, you should check out Midjourney. I make LoRAs for other production artists like myself, who aren't afraid to download a few files to get the look they're looking for.

I partially agree. At least the cover image shouldn't include any other models. BUT some negative embeddings are used universally and are kind of in a grey zone where their impact is minimal. Why should a creator stop using their basic prompt setup when testing their model? Still, things like "easy negative" tend to be a crutch that fix unoptimized models.

Using nothing but the base model in all the images doesn't make sense. You can mix style models for cool effects but most importantly there should be examples of how the model interacts with character models and vice versa. I can't count how many character LoRA i downloaded that were overtrained to a point that they were basically style LoRAs as well and could barely be used in conjunction with other models.

I completely agree at least in regard to checkpoints. I also don't care how many Loras are in the gallery images. My only concern is that the sample images posted initially with the model should be unadulterated by Loras or Embedding. Those initial images are there to show off the model.
On my models, I don't use any Loras, TIs or even any negative embeddings in any of my images that I post with a new model. In fact, I think in all the pics I have posted even later after the model is released, I might have one image where I use the darkening LoRa. I don't even use Loras and TI or negative embeddings most of the time for any images. I would love to see at least the initial images posted with a model be free of any of those things so we can see what the model itself does. For just showing off other art, that can be posted later in the gallery.

I share the same viewpoint as the original poster. It indeed conveys a misleading impression of the potential achievable use in its default state. On several occasions, I've encountered models that already employ LoRa's within their initial example images. At present, I utilize a model from one author and push it to its maximum capacity without relying on LoRa's, embeddings, and the like.

When I share images generated solely by the model using a well-crafted prompt, there should be no necessity for LoRa's to enhance their appearance.

If an author builds a model, then he/she should be capable of generating a prompt that effectively demonstrates its creative potential.

From my perspective, I examine the utilization of LoRAs within images and analyze their specific purposes. Afterward, I attempt to replicate the same impact directly within the prompt without resorting to additional models, as these often increase the processing time for each run.

Your answer