Type | |
Stats | 714 |
Reviews | (90) |
Published | Mar 24, 2023 |
Base Model | |
Training | Epochs: 20 |
Trigger Words | josephstalin |
Hash | AutoV2 BB494F341C |
Stalin Diffusion
IMPORTANT NOTES:
>This model is NOT meant to share or glorify any political view.
>Trigger word: "JosephStalin". Recommended weight: 0.6 and higher.
>Trained on mostly b&w pictures, in order to get good results appropriate prompting and model is required.
>The best results achieved with realistic models (tested with Realistic Vision).
>Probably the easiest way it to just use inpaint.
Prompting
>Highly recommended to add this to your prompt:
Positive: in color
Negative: ((cartoon)), (comics), black and white, (blurry), (noise), (outline glow), monochrome, grain, oversaturated, simple background
>Trained mostly on upper body pictures, might be hard to get full body.
>Interesting images can be created with low CFG scale.
>I suggest using 768x768 resolution.
How to make a similar model
I know this model is far from being great, but if someone wants to know the workflow, here you go:
1. Find good images and crop them. 2. Upscale x2~x4 (in my case mostly R-ESRGAN). 3. Remove noise using any graphic software. 4. Scale down 5. Train LoRa (using Google Colab, e.g. Kohya Trainer).
I mostly use b&w photos, so following steps 2,3,4 leads to significant increase of quality and reduction of image noise.
Also:
>Why do you share such controversial models?
Mainly for fun, also because I can as it’s allowed and I’m not doing anything illegal.
>Some of example images were created using ControlNet and img2img.
>If you want to use my model anywhere feel free to do so, but it would be nice to put me in credits.
>Check out my another model: Adolf Diffusion
>For any question you can contact me rafikxd123#7133