(8/22) New 1.4 AI model released! Tested fully functioning, no adjustments needed!
The definitive Stable Diffusion experience ™
Special thanks to all anons who contributed
What does this add?
Gradio GUI: A retard-proof, fully featured frontend for both txt2img and img2img generation
No more manually typing parameters, now all you have to do is write your prompt and adjust sliders
K-sampling: Far greater quality outputs than the default sampler, less distortion and more accurate
Easy Img2Img: Drag and drop img2img with built-in cropping tool
CFG: Classifier free guidance scale, a previously unavailable feature for fine-tuning your output
Lighter on Vram: 512x512 img2img & txt2img tested working on 6gb
Randomized seed: No more getting the same results, seed is randomized by default
Step 2: Git clone or download the repo from https://github.com/harubaru/waifu-diffusion/ and extract
(Make sure you have Git beforehand anyway, it will be needed)
Step 3: Go into the repo you downloaded and go to waifu-diffusion-main/models/ldm.
Create a folder called "stable-diffusion-v1". Rename your .ckpt file to "model.ckpt", and put it into that folder you've made
Step 6: Download Miniconda HERE. Download Miniconda 3
Step 7: Install Miniconda. Install for all users. Uncheck "Register Miniconda as the system Python 3.9" unless you want to
Step 8: Open Anaconda Prompt (miniconda3).
Go to the waifu-diffusion-main folder wherever you downloaded using "cd" to jump folders.
(Or just type "cd" followed by a space and then drag the folder into the Anaconda prompt.)
Step 9: If you have existing folders named "clip" and "taming-transformers" in /src, delete them
Step 10: Run the following command: "conda env create -f environment.yaml" and wait
(Make sure you are in the waifu-diffusion-main folder)
Step 11: Run the following command: "conda activate ldw"
(You will need to type this each time you open Miniconda before running scripts!)
- Open Miniconda and navigate to waifu-diffusion
- Type "conda activate ldw"
- Type "python scripts/kdiff.py" and wait while it loads into ram and vram
- After finishing, it should give you a LAN ip with a port such as '192.0.1:3288'
- Open your browser and enter the address
- You should now be in an interface with a txt2img and img2img tab
- Have fun
--NOTES AND TIPS--
- Build great prompts using the prompt builder
- Check out the wiki https://wiki.installgentoo.com/wiki/Stable_Diffusion
- Sampling iterations = how many images are made in a batch
- Samples per iteration = how many images are rendered simultaneously. It shouldn't be greater than 1 or 2 unless you have very high vram
- (img2img) Adjust Denoising Strength accordingly. Higher = more guided toward prompt, Lower = more guided toward image
Anywhere between 0.3 and 0.9 is the sweet spot for prompts
- If your output is a jumbled rainbow mess your image resolution is set TOO LOW
- Feeding outputs back in using the same prompt with a weak strength multiple times can produce great results
- The more keywords, the better. Look up guides for prompt tagging
- It's recommended to have your prompts be at least 512 pixels in one dimension, or a 384x384 square at the smallest
Anything smaller will have heavy artifacting
- Try Low strength (0.3-0.4) + High CFG in img2img for interesting outputs
- The seed for each generated result is in the output filename if you want to revisit it
- You can use Japanese Unicode characters in prompts
- This guide is designed for NVIDIA GPUs only, as stable diffusion requires cuda cores.
AMD users should try https://rentry.org/kretard
- A good tool for upscaling your outputs is Real-ESRGAN: https://github.com/xinntao/Real-ESRGAN
- You can prune a v1.3 weight model using "python scripts/prune.py" in waifu-diffusion-main
Pruning shrinks the file size to 2gb instead of 7. Output remains largely equivalent
- (Prune.py does not work on the new model, and does not matter as v1.4 is less heavy than v1.3 )
- If your output is solid green, the half precision optimization may not be working for you:
- GREEN SCREEN FIX:
1- change the value of "default" to "full" in line 169 and 343 of kdiff.py
2- delete ".half()" in line 89 of kdiff.py
(Note: this will raise vram usage drastically)
The original v1.3 leaked model from July can be downloaded here:
Backup Download: https://download1980.mediafire.com/3nu6nlhy92ag/wnlyj8vikn2kpzn/sd-v1-3-full-ema.ckpt
Torrent Magnet: https://rentry.co/6gocs
8/22: renamed "gradio.py" to "kdiff.py"- previous name conflicted with Gradio package causing an AttributeError.
If you are having issues, please rename it
- added fix to green screen of death
- added official v1.4 model links
8/23: Installation process now simplified vastly using new environment.yaml, original guide available at https://rentry.org/kretardold if problems arise