- This topic has 13 replies, 4 voices, and was last updated 1 year, 2 months ago by
.
-
Posts
-
I haven't posted in a while, but I thought this free tool is interesting. Currently, it is only available online and easy to use, plus I found a site where you can use it for free (and no registration needed either!). It doesn't always work, but I guess you should avoid picures that are too big (I am using Daz standard of 1000 x 1300 pixels here).
In the future it will be available for local use.
Here is what I tried out: I put image A and B in, put in a simple prompt and got as a result the last image on the right.
STEP 1: The girl with blue hair is wearing a white t-shirt, plaid skirt, and white boots.
STEP 2: The Girl is sitting on the bed and is smiling (I should have asked to zoom in first and then make her sit on the bed for a better result).
STEP 3: Zoom in. The girl with blue hair now has blonde hair.
I wouldn't say that it changes things by "a lot", just that it isn't trained on daz 3d models, hence the transformation of the 3d girl into a photographic one.
Other than that, there are some variations happening, but nothing major.
No worries, Frank. I think this is as close as possible to a consistent character with consistent clothing and environments as possible.
If you served an old Italian, noodle bolognaise in Italy, and told them it’s no different to spaghetti. You’d receive an ass whooping so legendary Satan himself would ask that it stop.
There isn’t a single thing that’s identical in the image labelled ‘Step 2’. Yes, both images have a bed in it, yes both have a carpet, and a ceiling.
Yes, both spaghetti and noodles are cooked in hot water.
But they are not the same!
It's impressive tech! Thanks for sharing. But to say it doesn't change things by "a lot" is a massive understatement. If you played spot the difference with that image, everything in that photo would have to be circled, because not a single thing is identical, even the ceiling is a different color!
This is interesting to play with Legolas. I used Florence-2 to get a detailed prompt for the base images, then changed some things like clothing, hair colour and background in the prompt. The model used for generation also appears to have an "understanding" of celebrities while Flux is not supposed to reproduce real celebrities. In fact FluxD was trained not to do so. No Loras were used for the images.
How it that Flux Kontext understands who Emma Watson is while prompting with her name in SD Flux has no idea who she is?
Flux Kontext
@geekaveli Thanks for the amusing story about the pasta 🙂 Yeah, italians can be quite passionate about it.
Now, I said "not a lot" because I am looking at things from the perspective of someone who has experience AI limitations for the last few years so, for me, this is a huge step forwards. And, in step 2, even the clothing has remained consistent, something that was previously impossible.
As for the environment, yes, it has changed, but the colors are easily fixable in photoshop and overall a great improvement from the last tech, so for me it's a win. Not a perfect one, but a really good one 🙂
@frank22 Overall, I am not too worried about faces, since those are very easily replaced with AI.
With text2img, you can just add a character LORA at the Adetailer stage (with face detection), and it substitutes the face automatically. It may require a 0.5+ denoise, but it can even maintain the same style of the image.
If you have an existing image, you can either inpaint the face you want, or use a workflow like this one, that allows you to substitute several faces in the same image in one go:
The cool thing for me is that Flux Kontext gives so much more control on how to make images, you just plug in the elements you want, and you create the scene. It makes my DAZ library more useful, I can now more easily add clothing, accessories, etc but I'll have to test it more, especially with strange poses.
I can't wait until it is open source so that I can create my own crazy workflows with it and with NSFW workarounds too 🙂
AI is progressing rapidly. I saw this video today and Hollywood must be very afraid of AI. I hope that in a few years I can create a comic with AI that consists of characters that look consistent.
@realteamkiller Actually, consistency can be done, it just requires work:
Watch the final result @24.36
Looks good, but within a year they will have invented something that makes training a lora even easier and faster.
@realteamkiller Well, if you have a 3d character, you don't have to train a LORA, since you already have a consistent character. You could even add in Kontext to change the style.
So I can use a warhammer character as input. Can more than one character be used to create a scene with the help of AI? And is an RTX3090 still sufficient?
@realteamkiller If you use Kontext, yes for style change. I typically use it only for that. Otherwise, you'll need to use each character as an input image.
As for a 3090, it is still a top GPU!
- You must be logged in to reply to this topic.
