- Options
- Login
- Desktop Style
Options
AI-generated "Voynich manuscript pages"
- AI-generated "Voynich manuscript pages"
-
RE: AI-generated "Voynich manuscript pages"
Searcher > 15-07-2022, 10:19 PM
I tried DALL-E Mini, as well. it's interesting like a toy, but not for a masterpice, of course. It is funny how it imitates the Voynich style, but as much as I see, "he" is impressed mostly with circular diagrams
My nymphs and plants were always in circles, of strokes imitating text.
I signed up for the DALL-E and Midjourney waitlists. I'm still waiting for an invitation to try the DALL-E AI, but I've already received an invitation to Midjourney and have even used my free trial to experiment. Unfortunately, there were not so many attempts, besides, I immediately rejected several options. When I realized that this AI ignores numbers of objects (in this case, 6 towers) in the same way as the AI in Disco Diffusion, I finally moved on to other attempts. I must say that in other respects Midjourney clearly outperforms Disco Diffusion. Firstly, the generation is very fast, about a minute, first 4 options are offered, you can choose one of them, improve (finalize) or make options for the already selected picture, then you can upscale the desired picture; secondly, the picture is much more consistent with the text description than in Disco Diffusion, i.e., the AI generates the image more accurately; thirdly, there are almost no incomprehensible artifacts in the picture, only given objects. Well, I will add that Midjourney does an excellent job of generating human figures and faces, as well as animals. In general, the results impressed me, but I'm not ready to buy the paid version, so far.
Further, I also want to note the disadvantages, at least of the trial version. It takes place in the general chat of Midjourney newbies on Discord, you can only use text prompts, as well, you need to constantly look for your message (image) in the chat, as new messages from other newbies are constantly added. To be honest, I spent more time searching for my pictures than creating them, however, in the process of this, I saw many successful and interesting incarnations.
Nevertheless, I must confess, despite all the disadvantages of Disco Diffusion, it is quite good in landscapes, it is not so accurate in details, but, it seems, more creative and artistic. Little and far details are often blurred or deformed, but there is some artistic charm in this. And sometimes, it even seems that AI once works with inspiration, another time - without any enthusiasm.
My aim is a little differ, in general. I still try to make some Voynich scenes realistic. You can evaluate how Midjourney coped with my task:
-
RE: AI-generated "Voynich manuscript pages"
MarcoP > 20-07-2022, 03:26 PM
DiscoDiffusion 5.6 has a new "portrait generator" that you can select as diffusion_model. Probably useless to generate Voynich images, but it seems to work rather well. -
RE: AI-generated "Voynich manuscript pages"
ReneZ > 20-07-2022, 07:59 PM
This is extremely cool!!
I like the two centre ones (top row and bottom row) most, for different reasons.
I wonder what the tool would do with cleaned up images of the MS page(s), though the noise in the input image doesn't seem to bother the tool too much. -
RE: AI-generated "Voynich manuscript pages"
bi3mw > 20-07-2022, 11:02 PM
I have preset a path under "custom model" that is not created at all in my copy of Disco Diffusion v5.6. Does anyone else also have this problem ?
custom_path: "/content/drive/MyDrive/deep_learning/ddpm/ema_0.9999_058000.pt" -
RE: AI-generated "Voynich manuscript pages"
MarcoP > 21-07-2022, 06:48 AM
(20-07-2022, 07:59 PM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.This is extremely cool!!
I like the two centre ones (top row and bottom row) most, for different reasons.
I wonder what the tool would do with cleaned up images of the MS page(s), though the noise in the input image doesn't seem to bother the tool too much.
Hi Rene,
I am glad you enojoy these experiments!
The tool has a huge number of parameters, but I only tried a few of them. Here, I used different prompts, so that bottom left is "a drawing by Leonardo da Vinci", while top center is "a watercolor portrait". I also varied skip_steps, where higher values produce results which are closer to the original: top-center has a much higher skip_steps value than the other images.
As you say, noise does not have any visible effect. Actually, intermediate steps when running the AI and noise is totally removed at the start of the process, with different and less noisy high-frequency data added in the following steps.
(20-07-2022, 11:02 PM)bi3mw Wrote: You are not allowed to view links. Register or Login to view.I have preset a path under "custom model" that is not created at all in my copy of Disco Diffusion v5.6. Does anyone else also have this problem ?
custom_path: "/content/drive/MyDrive/deep_learning/ddpm/ema_0.9999_058000.pt"
Hi Matthias,
I see the same assignment and the file is not on my google drive, but for some reason this does not seem to cause any problem. I understand very little of how this all works, so I am afraid I cannot help. -
RE: AI-generated "Voynich manuscript pages"
MarcoP > 27-07-2022, 12:42 PM
Jar experiments with DiscoDiffusion. Prompt: "Detailed photograph of an ancient jar". Various skip_steps values (120/150). Initial images were pre-processed to highlight edges and/or separate the jars from the background. -
RE: AI-generated "Voynich manuscript pages"
MarcoP > 12-06-2023, 08:06 AM
DALL-E expansion of Voynich details (f82r and bottom-center Rosette).
-
RE: AI-generated "Voynich manuscript pages"
RenegadeHealer > 12-06-2023, 04:30 PM
MarcoP, very fitting that that second image symbolically cries out “protect me from the evil eye!” ? -
RE: AI-generated "Voynich manuscript pages"
Aga Tentakulus > 12-06-2023, 07:12 PM
You mean the Turkish eye. -
RE: AI-generated "Voynich manuscript pages"
MarcoP > 30-08-2023, 07:11 PM