Help
Duke4.net
DNF #1
Duke 3D #1
For comparison, here's the same sprite processed with the same query but at 0.85 strength. Seed is obviously different as I can't change that but as you can see there's a pretty radical change from the input image.
GFPGAN Hall of Horrors "To spare MrFlibble's sanity"
#91 Posted 29 October 2022 - 10:13 AM
The page I use has a setting that allows you to adjust strength. this was at 0.55 strength I think. At lower setting you get a picture closer to the original but obviously less refinement. At higher settings you get higher "quality" but more deviation from the input image.
For comparison, here's the same sprite processed with the same query but at 0.85 strength. Seed is obviously different as I can't change that but as you can see there's a pretty radical change from the input image.
For comparison, here's the same sprite processed with the same query but at 0.85 strength. Seed is obviously different as I can't change that but as you can see there's a pretty radical change from the input image.
#92 Posted 03 November 2022 - 12:41 PM
Good news, I found a similar web app that allows both supplying a different image and forcing a specific seed https://huggingface....tuned_diffusion
I’m too lazy to try it right now, but maybe someone else will have some fun with it
I’m too lazy to try it right now, but maybe someone else will have some fun with it
#95 Posted 19 November 2022 - 05:18 AM
Found another site which I think uses the same algorithm as Diffuse the Rest but 1. Allows you to set a seed before processing 2. Doesn’t have the NSFW filter
https://www.artbreed...om/beta/collage
https://www.artbreed...om/beta/collage
#97 Posted 26 November 2022 - 08:42 AM
Nice! Except isn't that's supposed to be an ukiyo-e, not a photorealistic image of a tiger? (I suppose someone already found the original image in higher res?)
#98 Posted 28 November 2022 - 07:34 PM
Golden Retriever
Firetruck
(complete with fireman sitting in the back)
Firetruck
(complete with fireman sitting in the back)
#99 Posted 29 November 2022 - 07:19 AM
The golden retriever painting downscaled to 128x128 and run through the Pixelsharpen model
Donkey Nukem run through the same model
Donkey Nukem run through the same model
#100 Posted 08 December 2022 - 04:43 AM
Not bad! I don't think you mentioned Pixelsharpen before (I've not touched anything ESRGAN-related in a while), it seems to produce a decent imitation of palettized pixel art, complete with dithering effects. The Donkey Nukem image still has noticeable ringing around the edges though.
#101 Posted 08 December 2022 - 05:53 AM
Here are a few more processed with Pixelsharpen
Edit: I noticed I had used the wrong scale on the last two, here are versions at the "correct" scale
Edit: I noticed I had used the wrong scale on the last two, here are versions at the "correct" scale
This post has been edited by Phredreeke: 08 December 2022 - 05:56 AM
#103 Posted 20 December 2022 - 10:32 AM
Looks crazy, but it seems to have potential. Can you make it show the same girl in all frames?
#104 Posted 09 February 2023 - 05:46 AM
As for the dancing girl, this is the most consistent I've gotten it
#106 Posted 09 February 2023 - 02:53 PM
Honestly there's not been much GFPGAN here since page 2 I think
#107 Posted 13 February 2023 - 05:39 PM
Upscaled with XBR, then threw on a LOT of picture noise before processing with stability diffusion
#111 Posted 19 February 2023 - 06:02 PM
I can downscale the existing sprite, but it wouldn't produce the stylized look that the AI does
#112 Posted 22 February 2023 - 06:14 PM
A little something I did...
https://cdn.discorda...ces_of_jojo.mp4
explanation of what was done: 1. the initial texture was upscaled with XBR 2. the resulting XBR texture was run through CLIP interrogator to generate a prompt 3. said prompt was entered along with the texture in stable diffusion and run with 30 different seeds to generate different results 4. the resulting images were faded between eachother
https://cdn.discorda...ces_of_jojo.mp4
explanation of what was done: 1. the initial texture was upscaled with XBR 2. the resulting XBR texture was run through CLIP interrogator to generate a prompt 3. said prompt was entered along with the texture in stable diffusion and run with 30 different seeds to generate different results 4. the resulting images were faded between eachother
#114 Posted 26 February 2023 - 09:38 AM
I used DeepDanbooru String instead of CLIP interrogator for this one. I used the Openjourney model with the mdjrny-v4 style tag and also used a long negative prompt that I copied from another stable diffusion space.
If you want to download 400 different renders of innocents here's a link
https://cdn.discorda...4_innocents.zip
#115 Posted 16 May 2023 - 02:29 PM
(used Allysmix 3 along with Pixel Lora for this, can post specifics tomorrow if there's interest in it)
#116 Posted 21 May 2023 - 05:57 AM
Phredreeke, on 16 May 2023 - 02:29 PM, said:
(used Allysmix 3 along with Pixel Lora for this, can post specifics tomorrow if there's interest in it)
Yes, please do. I've not followed AI image generation developments for a while, and it looks like there's some potentially interesting stuff going on.
BTW, what is currently the best option for online text2image generation?
#117 Posted 22 May 2023 - 04:15 AM
Your best shot, unfortunately, is to buy compute units on Google colab. It used to be possible to run Stable Diffusion on free instances but that has been blocked now.
There are free spaces on huggingface but those limit you to a particular model, and often have queues.
If you got an iPad or Mac with an M1 or M2 chip you can render images locally using Draw t
hings.
I'll post the parameters used for above image when I get home (it did use a set of frames from Duke3D as input image)
There are free spaces on huggingface but those limit you to a particular model, and often have queues.
If you got an iPad or Mac with an M1 or M2 chip you can render images locally using Draw t
hings.
I'll post the parameters used for above image when I get home (it did use a set of frames from Duke3D as input image)
This post has been edited by Phredreeke: 22 May 2023 - 04:15 AM
#118 Posted 23 May 2023 - 02:52 AM
parameters
realistic, a woman that is dancing in a leotard, from behind, a raytraced image, inspired by Hans Erni, sexy-s 100, girl in white dress dancing, closeup, character art, high quality, masterpiece, <lora:pixel:0.65>
Negative prompt: bad result, worst, random, invalid, inaccurate, imperfect, blurry, deformed, disfigured, mutation, mutated, ugly, out of focus, bad anatomy, text, error, extra digit, fewer digits, worst quality, low quality, normal quality, noise, jpeg artifact, compression artifact, signature, watermark, username, logo, low resolution, worst resolution, bad resolution, normal resolution, bad detail, bad details, bad lighting, bad shadow, bad shading, bad background, worst background
Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 9, Size: 512x512, Model hash: f427dd7a4f, Model: allysmixiii, Denoising strength: 0.5, Mask blur: 4
Input image is attached below
realistic, a woman that is dancing in a leotard, from behind, a raytraced image, inspired by Hans Erni, sexy-s 100, girl in white dress dancing, closeup, character art, high quality, masterpiece, <lora:pixel:0.65>
Negative prompt: bad result, worst, random, invalid, inaccurate, imperfect, blurry, deformed, disfigured, mutation, mutated, ugly, out of focus, bad anatomy, text, error, extra digit, fewer digits, worst quality, low quality, normal quality, noise, jpeg artifact, compression artifact, signature, watermark, username, logo, low resolution, worst resolution, bad resolution, normal resolution, bad detail, bad details, bad lighting, bad shadow, bad shading, bad background, worst background
Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 9, Size: 512x512, Model hash: f427dd7a4f, Model: allysmixiii, Denoising strength: 0.5, Mask blur: 4
Input image is attached below
