AI Image Generation

Bandwagon

Bandwagon

Kolohe
<Silver Donator>
24,134
64,566
A picture implying that whatever you posted was going to be, uh, temporally ambiguous.

First time I've been suspended from bing
 
Kharzette

Kharzette

Watcher of Overs
5,307
4,053
I was going to just see what it did with one of my halloween spooky pictures :D I'm still downloading the smaller model. My internets are slow.
 
Kharzette

Kharzette

Watcher of Overs
5,307
4,053
Well I gave it a good go. Got down to a raw terminal and tweaked everything I could find down to minimum. No luck.

Reading back over that article I saw it mention it needs 80 gigs of gpu ram :emoji_laughing:

Hmm this post has some decent ideas

I'll try em later
 
  • 1WTF
Reactions: 1 user
Edaw

Edaw

Parody
<Gold Donor>
13,176
87,202
Kharzette said:
Well I gave it a good go. Got down to a raw terminal and tweaked everything I could find down to minimum. No luck.

Reading back over that article I saw it mention it needs 80 gigs of gpu ram :emoji_laughing:
Click to expand...
not sure if relevant



Maybe poke around on that reddit - likely there is something more recent.
 
Kharzette

Kharzette

Watcher of Overs
5,307
4,053
Haha you found the same thing. Yea I was searching for a pruned model and came across that.
 
  • 1Like
Reactions: 1 user
Kharzette

Kharzette

Watcher of Overs
5,307
4,053
Kharzette said:
I was going to just see what it did with one of my halloween spooky pictures :D I'm still downloading the smaller model. My internets are slow.
Click to expand...
Wow I'm stupid. The image I was using was 2048x2048

By the time I figured that out I was already half an hour into changing to that reddit requirements.txt. That was a big mistake as it completely trashed my conda env and now I'm starting over.
 
Last edited:
  • 1Like
Reactions: 1 user
Kharzette

Kharzette

Watcher of Overs
5,307
4,053
I got it going with the web thing. Sweet spot for 12 gig seems to be 768x448 video with a 512x288 starter image. Results are not great.

Forum doesn't do mp4 bleh
 
  • 1Like
Reactions: 1 user
Kharzette

Kharzette

Watcher of Overs
5,307
4,053
I'll find a way to post a few when I get a bit better results. The key is a low vram switch in the web helper stuff. I'm not sure how that makes it down into the guts, but I'd like to transfer that to the raw scripts, then get out of X altogether and have all my gpu mem free in a shell.

The number of frames seems to affect the vram needed and I'm not sure why it would.

Right now my results are just a character on a slightly panning background. I think I need to use an image that has the appearance of motion or action. I'll try later me smeepy. :emoji_zzz:
 
Kharzette

Kharzette

Watcher of Overs
5,307
4,053
These were generated from raw shell. The trick to low vram is to do model.model.half(); I'm guessing that is just chopping the model in half or converting to fp16 or whatever.

 
  • 1WTF
  • 1Like
Reactions: 1 users
Daidraco

Daidraco

Golden Baronet of the Realm
9,956
10,196
Kharzette said:
I wonder what they used for this:
Click to expand...

Shit makes me want to play it again, so I guess it accomplished that. What I love about AI, though, is even after forcefully changing code to be more DEI - all the training data still makes the most picturesque white people.
 
Kharzette

Kharzette

Watcher of Overs
5,307
4,053
That indeed gives me hope. The beauty of the elves lingers in scandi/germanic dna.

I've been thinking more about this, and I think maybe they might be just rendering an asston of frames with stable diffusion using controlnet to animate, then using some kind of morphy tweener to generate inbetween frames.

That would be super tedious but doable.

I need to try controlnet again. Last time I did I was still using my old hand-me-down 980gtx. Now I've got a hand-me-down oddball 1080 with more ram:
+---------------------------------------------------------------------------------------+ | NVIDIA-SMI 535.183.01 Driver Version: 535.183.01 CUDA Version: 12.2 | |-----------------------------------------+----------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+======================+======================| | 0 NVIDIA GeForce GTX 1080 Ti Off | 00000000:01:00.0 On | N/A | | 19% 55C P0 69W / 250W | 748MiB / 11264MiB | 0% Default | | | | N/A | +-----------------------------------------+----------------------+----------------------+
 
Kharzette

Kharzette

Watcher of Overs
5,307
4,053
This took 34 minutes to generate. It's about 100 frames of derpy lazy stable diffusion gens with the tweener generating 6 frames between each real frame.

 
  • 1Like
Reactions: 1 user
Edaw

Edaw

Parody
<Gold Donor>
13,176
87,202
My guess is the more professional videos are using runway or luma. There is kling too but I think it's not as good.
 
You must log in or register to reply here.