5 ms·

I was trying this recently with the Sierra Christmas Card from 1986![0] The images that I generated were[1], and I was trying to tweak the model parameters with

I was trying this recently with the Sierra Christmas Card from 1986![0] The images that I generated were[1], and I was trying to tweak the model parameters with different denoising and CFG scales. When you get the parameters just right you can preserve the composition of the input image very well while still adding a lot of detail. This isn't a completely automatic process though, with Stable Diffusion you have to provide the right prompt otherwise the generation process isn't guided correctly, so approach works better for aesthetics and style transfer than regular image super-resolution such as ESRGAN. [0] https://archive.org/details/sierra-christmas-card-1986 https://archive.org/details/sierra-christmas-card-1986 [1] https://i.imgur.com/WxD05gX.jpeg https://i.imgur.com/WxD05gX.jpeg

Agentlien 4y ago
If you're using stable diffusion 2.0 or later you can use its depth-to-image mode[0] to create variations of an image which respect its composition without having to keep your parameters within a narrow range. [0] https://github.com/Stability-AI/stablediffusion#image-modification-with-stable-diffusion https://github.com/Stability-AI/stablediffusion#image-modifi...
- siraben 4y ago
  I still struggle with SD 2.0 prompting because many of the tricks (greg, artstation) don't work anymore, have people had success with it or do I have to use custom models?
  - speedgoose 4y ago
    You need other tricks and put a bit more thoughts in the negative prompt. The quality is then higher but you really need a good prompt. It’s a bit frustrating compared to midjourney or dall.e. You should take a look at embeddings too. They are tiny files, no more than 128kB, that have a huge influence on the final output. You put the files in the embeddings folder and use the filename in your prompt. Ideally the filename is a unique word so it doesn’t interfere with the normal prompt logic. You can find the best embeddings in the stable diffusion discord.
    - Agentlien 4y ago
      > The quality is then higher but you really need a good prompt. I've seen this stated but it has not been my experience. Nor have I seen solid examples of it. For 2.0 (haven't tried 2.1) I find the model very finicky and unstable. Any prompt which works well also seems to work at least as well in 1.5.

Agentlien 4y ago

If you're using stable diffusion 2.0 or later you can use its depth-to-image mode[0] to create variations of an image which respect its composition without having to keep your parameters within a narrow range. [0] https://github.com/Stability-AI/stablediffusion#image-modification-with-stable-diffusion https://github.com/Stability-AI/stablediffusion#image-modifi...

siraben 4y ago
I still struggle with SD 2.0 prompting because many of the tricks (greg, artstation) don't work anymore, have people had success with it or do I have to use custom models?
- speedgoose 4y ago
  You need other tricks and put a bit more thoughts in the negative prompt. The quality is then higher but you really need a good prompt. It’s a bit frustrating compared to midjourney or dall.e. You should take a look at embeddings too. They are tiny files, no more than 128kB, that have a huge influence on the final output. You put the files in the embeddings folder and use the filename in your prompt. Ideally the filename is a unique word so it doesn’t interfere with the normal prompt logic. You can find the best embeddings in the stable diffusion discord.
  - Agentlien 4y ago
    > The quality is then higher but you really need a good prompt. I've seen this stated but it has not been my experience. Nor have I seen solid examples of it. For 2.0 (haven't tried 2.1) I find the model very finicky and unstable. Any prompt which works well also seems to work at least as well in 1.5.