- This topic has 25 replies, 3 voices, and was last updated 2 weeks ago by
.
-
Posts
-
-
I’ve probably clocked up a few hours of experimentation with AI video generation (I won’t use the term “creation” as it seems to cause offence), and I’ve learnt a great deal. Being detailed and precise with prompts is critical, but so too is avoiding leaving things we think of as obvious unsaid.
I’ve recently reached the point where about 70% of the time I get precisely what I wanted from my prompts. This is beneficial, not only for achieving results but also for avoiding the waste of precious credits.
Then there’s that other 30%. Many of them I can correct with just a tweaked prompt. Sometimes it takes a couple of goes. That leaves the small minority which drive me mad! These are the ones where the AI seems intent on finding any little chink in my prompt to ruin the output.
Take the example of the video clip of a naked woman turning her back to us. It’s a 3 second clip, so you wouldn’t think much could go wrong. I won’t quote the prompts as I didn’t store them all and it would make this post far too long.
The 1st attempt saw the woman turn in the correct way and end in the correct position (a standing pose with arms by sides) but I forgot to specify that she shouldn’t have hair-covered buttocks. After mentioning the lack of hair, I tried again.
That 2nd attempt matched the movement perfectly but added hair on her back instead. I adjusted the prompt to specify no body hair on her entire body and tried again.
The 3rd one featured no back or buttock hair, but instead added a large brown stain on the inward curve of her buttocks. Another prompt adjustment was called for.
The 4th go results in a clean hairless bottom, but no buttocks at all! Now I needed to add a detailed description of her backside to ensure it was included.
The 5th generation produced quite attractive buttocks but covered in some kind of coarse brown powder! I deleted the entire prompt and started again.
This resulted in no unwanted hair, no brown deposits but for reas
-
I know I’m not getting any feedback on this, but the potential is brilliant. I put together the clips I have done so far to make a 7 minute video with imperceptible image degradation. There are some minor glitches if you’re looking closely but I have to say for a first proper try, the results have stunned me.
Sound is the main problem now.
Definitely want to make many more clips, more ambitious clips and do something on an even bigger scale. There are all sorts of possibilities here. I might post this video on a age-verified only page tomorrow. Anyone interested?
-
I’m watching closely and very interested in your progress.
Behind the scenes I’ve installed ComfyUI on a new computer and am trying a few low resolution generations. At the moment I’m limiting it to CPU because the GPU on this machine is AMD. Comfy presently expects Nvidia. So generations are very slow. There are instructions around for using AMD but they’re not very clear and some of them are described as a development preview. So at the moment it’s useful to experiment with the models and see what they’re capable of without having to spend any credits.
-
The ComfyUI website says it requires an Nvidia GPU. They don’t mention that it runs without one. Do you need to do anything special to it to make it work on just the processor?
-
I’m installing it on my laptop. The “CPU-mode” is almost hidden but definitely there. My NVidia card is probably too old to support CUDA so, as my laptop is more powerful, I’ll install it there.
-
For CPU mode, download the “portable” version, extract the folders and look for run_cpu.bat.
CPU mode is slow. Some low resolution generations came out in about an hour for 5 seconds, but that’s on a fast desktop. It uses about half the memory and half the CPU time, so I can leave it running in the background without affecting my work too much. Some models take much longer than others.
I’m hoping they get the AMD GPU sorted.
-
I’m not sure what counts as fast these days. My laptop has an 8-core Ryzen 7 with 4.4GHz clock speed. It works out about 20% faster than my desktop PC according to the benchmark scores, and it’s got 32Gb instead of my desktop’s 16Gb.
I should invest in a new NVidia card for my desktop, that would blow both of them out of the water.
-
-
I find that if you give it a very detailed prompt that is a good idea most of the time but sometimes the more detailed you are the more likely it is to get confused. My main issue is I still often get a lot of images where they look good on the surface but if you look more carefully people have extra fingers and toes and maybe a couple of extra limbs here and there. Again every AI image creators different, which is why if you have a particular thing you are looking for it might pay to give one prompt to several different AI image makers. Like one that I tried making the other day that didn’t have any nudity involved was I just wanted a picture of an Asian woman getting her back scratched by a kitten, in three out of the four pictures it resulted in her scratching the kittens back when I really wanted it the other way around. So yeah AI can be unpredictable, sometimes it feels like AI has practically read my mind and other times it’s like where the hell did this go so horribly wrong?!
-
I find Qwen and GetImg don’t produce extra limbs/fingers and are very good at producing anatomically correct people.
Giving the right amount of detail is the key. Try not to suggest ideas that would not have cropped up otherwise. The other thing is repetition. I find stating the same thing 2-3 times in different words tends to lock it in.
When I first got it to generate a nude woman turning around, I didn’t expect this:
-
-
@Ed
“I find Qwen and GetImg don’t produce extra limbs/fingers and are very good at producing anatomically correct people.
Giving the right amount of detail is the key. Try not to suggest ideas that would not have cropped up otherwise. The other thing is repetition. I find stating the same thing 2-3 times in different words tends to lock it in.
When I first got it to generate a nude woman turning around, I didn’t expect this:”
Yes random shit stains are strange. I haven’t been able to get getimg to do nudity any longer but maybe I should try again.
-
GetImg and Qwen both block nudity, but they have other distinct strengths.
I use multiple different systems to get my results. I use GetImg to create characters, Qwen to create and edit scenes and HuggingFace to create the videos from them. When I have a lot more credits on Pollo, I’ll use that to produce the videos as it does much higher resolution videos.
-
For those of you who are age-verified, I’ve posted a first-draft of what I’m working on here:
-
I mostly use META AI, seaart, Perchance and getimgAI, I was using hugging for a while but then it said I used up all of my credits and although it says I’m supposed to get like a thousand a month usually when I go back to trying to use it I maybe use it a couple of times and then it says that I have exhausted my options more or less. I think for anything not explicit meta-a right tends to be really good except now that they have upgraded it like each picture saves like 2 MB on my hard drive and they give you for so I have tried using some of the other ones. Seaart and Perchance I think are the best if you are going to be doing nudity, but again a lot of these vary. If I have a really good concept that I like I tried putting it through all four of them!
-
I find Perchance is completely unpredictable, really inconsistent results. For my purposes, that renders it unusable. Huggingface gives credits daily now. Unless you pay, the free credits are only enough for 2-3 5 second clips per day.
I’ve never heard of Meta AI. I’ll give that a go. I know I tried SeaArt but it’s on the list of ones I gave up on (can’t remember why).
What I’d really like is one of the very high price accounts on Pollo.ai but I can’t afford that for now.
Qwen would be incredible if it allowed nudity. It’s the most consistent by far and allows fantastic editing of pictures to correct anything which isn’t what you want without changing the rest of the photo and without having to draw mats.
-
-
@Ed
“I find Perchance is completely unpredictable, really inconsistent results. For my purposes, that renders it unusable. Huggingface gives credits daily now. Unless you pay, the free credits are only enough for 2-3 5 second clips per day.”
I heard that right now per chance is giving itself an update so that’s part of the reason why it’s being inconsistent at the moment. Sometimes the pictures come out really distorted and bizarre but they don’t seem to give you any problem about nudity is what I was kind of getting at. Hugging face I have never actually tried making videos with any of these things so far I was just using it for pictures but I guess I will have to try now because maybe they have changed their policy because I know that now getimg AI also gives you like 140 daily credits which amounts to about 28 pictures a day if each picture costs five credits where it used to give you only that many per month.
“I’ve never heard of Meta AI. I’ll give that a go. I know I tried SeaArt but it’s on the list of ones I gave up on (can’t remember why).”
Meta AI goes along with Facebook it always gives you four pictures for every prompt and they recently upgraded it so now the pictures are very high quality and artistic looking but each picture is like 2 MB which if you’re going to a lot of these takes up more room on your hard drive. Up until recently that was the one I used for most things that didn’t but now I am mostly using Getimg and Seaart. Seaart It’s pretty decent about giving nudity but sometimes it just totally excludes that but that seems to be true of most of these AI image makers. Like I tried all of these image makers to get Russian dictator looking at a naked ballet, because I want to illustrate a story I plan to write about naked ballerinas assassinating a dictator, but every time that I tried it on every single one of these AI image creators the ballerinas always kept their clothing on, I couldn’t get it to do a naked
-
Qwen created an image:
And then with the prompt “Put them both in just bikinis” produced this:
-
-
@Ed
I think that those came out rather wellm and yes it is good that after you get initial picture you can refine it like that, but you said that it won’t do nudity so I guess she won’t lose the towel! I tried it with two pictures and they came out to be high quality but once again really large pictures that take up like a megabyte or two. But I will agree that Qwen is high quality even if it doesn’t do nudity.
-
HuggingFace can remove the towel. Setting up scenes in GetImg or Qwen saves my HuggingFace credits. I find it works best to use the strengths of each.
-
-
@Ed
“HuggingFace can remove the towel. Setting up scenes in GetImg or Qwen saves my HuggingFace credits. I find it works best to use the strengths of each.”
I have to admit I haven’t really used editing and refinement of these pictures or uploading a picture from one to another and asking it to edit the picture, I guess I’m not really especially great with this kind of stuff, generally speaking I will go to an AI image creator and give it a prompt and if it doesn’t give me what I want I will try another AI with the same prompt hoping for better results. I feel if I have a really great prompt though I will try it just in every single different AI image creators that I used just to see which results it will give me because it gives you a good variety. Although a lot of them do give very similar ones it must be said because I guess they are probably trained using similar methods.
-
There’s a lot to learn, and the technology keeps changing and advancing. What you knew yesterday may be different today. I don’t know if you’ve seen the video I posted yesterday (age-verified forum), but I’m sure I would not have been able to create that just a year ago.
-
One more Qwen picture. None of them seem to do reliable male nudity unless you start out with a naked photo as a video source. Qwen is pretty good at changing the text on signs and posters.
You can see in this photo that too many tweaks and changes and you clearly lose image quality. Looks like 5-6 is fine, but much beyond that can cause problems. If you want to make the effort you can use a photo editor to neatly clone just the changes into the original photo, thus prserving the original quality.It’s still very much the case that the more effort and planning, the better the results.
-
Generally I try to keep AI generation (or photo editing) down to 3-4 levels of change. There are techniques to get around that, which I’m slowly learning, but it definitely requires creative planning to get the best out of it.
Do you think it’s worth doing some sort of non-nude video so I can show it in the public area of the community? I know non-nude kind of avoids the topic of this community, but this age-verification thing is stifling discussion.
-
Anything that’s relevant to attempts to create images for the site is likely to be of interest even to those who haven’t yet been age verified. And maybe those people will choose to get age-verified, in order to see more.
-
I hope so!
-
-
It’s an annoying waste of credits when things go wrong, but some of the off results are crazy. I might post some of the videos which went completely off the rails if anyone’s interested.
-
- You must be logged in to reply to this topic.