- This topic has 23 replies, 4 voices, and was last updated 2 weeks, 4 days ago by
.
-
Posts
-
Some Loras, allegedly, don't need trigger words, so I decided to try a wee experiment.
I used this Lora https://civitai.com/models/824739/flux-3d-animation-style-lora which didn't provide any trigger words, but I noticed that nearly all of the sample images had prompts beginning with the words "3d animation style" which looked like trigger words to me, so it seemed like a reasonable guinea pig.
I went with a prompt befitting the magnitude of this experiment … "Wonder Woman in a fighting stance pose" and jumped in. The first image, without trigger words, was a random seed, but I kept that seed for subsequent images with the trigger words.
2nd image, trigger words at the beginning of the prompt.
3rd image, trigger words in the middle of the prompt.
4th image, trigger words at the end of the prompt.
I definitely prefer the image without the trigger words, but when you do use them, where you use them clearly affects how the AI interprets the prompt.
I think "3d animation style" is a prompt to style as well as a trigger ie, the Flux model knows what the phrase means (the concept) and it is also used heavily in the lora training captioning (below). I guess without the phrase it produces images that are less 3D model looking and can move away a bit from the the dataset images.
I sometimes think that certain concepts are so well known by Flux that you don't really need a lora for them.
Naturally, I felt obliged to put that to the test Frank. So No Lora, just the prompt with the placement of "3d Animation Style" as before.
Additional fingers aside, I think the middle one's my favourite.
Kawaii 🙂
I've had Flux for two weeks but I haven't used img2img yet. I should give it a spin.img2img can give your Daz renders a whole new lease of life. My Flux vampire dance images all stemmed from a Daz scene I couldn't get to work properly.
Beautiful images Evie! I like the pixar style.
BTW, there is not only img2img, but also RF Inversion, which is a style transfer (tries to "convert" the input image into a similar one with a new style).
It certainly can give some unexpected results. 🙂
"It certainly can give some unexpected results."
You should have added.".... in addition to the crazy randomness that usually comes out of AI." 🙂
@frank22 Totally true frank, however in the images here I tried to add as much noise as possible since it tends to give out more beautiful results. 🙂
@littleevie I am glad your interest got piqued 🙂 I had been looking to do style transfer for a few days, but the IP adapter thing was not doing it for me.Unfortunately, it doesn't seem to change style with all images (or maybe it's just the checkpoint I am using that is to blame), but I did get quite a few beautiful images.
Is this cheating?
Flux needs natural language prompts and while I could write them they're very time consuming. So I use a base image and a site to generate a prompt. Then change the prompt a lot to generate the image I want with respect to style, posing, clothing, expressions etc. In the end the image doesn't look anything like the base image but it does help with prompting for the basic layout/concept.
I guess for those who consider generative AI stealing, this is stealing on steroids 🙂 but it gets me what I want. Have I used enough of my own input to call the result mine? Should I even care?
You're using the tools available to get the image you want Frank, that's not cheating, it's fine-tuning your work process.
It's all cheating just in different ways.
Daz uses pre-made figures, AI uses pre-made art.
ATM there's a slight difference that AI is more random and decides some things for you.
But both are still cheating
- You must be logged in to reply to this topic.
