Over the years, I've been changing and improving my generation settings in hope of getting closer to a 1 to 1 replication of what I believe to be the ideal 2d anime style. (in my head) I have come across a few settings I believe would be beneficial to many if they choose to use them. I will keep this article updated for any future changes I make.
If you have any tools/tricks you use to improve your image, please leave it in the comments. I'm always interested in trying new things.
Generation Tool
I currently use ReForge and will continue to use it for the foreseeable future. You can take what is written here and use it on whatever model you like.
Current Models Used
Generation Settings [Epsilon]
Sampling Method: Euler a
Schedule Type: Normal
Sampling Steps: 45
CFG Scale: 6
Width/Height: 1024x1472
ADetailer: face_yolov8n.pt [Used during MultiDiffusion upscaling, see below]
Generation Settings [V-Pred]
Sampling Method: Euler Ancestral CFG++
Schedule Type: DDIM
Sampling Steps: 45
CFG Scale: 1
Width/Height: 1024x1472
ADetailer: face_yolov8n.pt [Used during MultiDiffusion upscaling, see below]
-Additional V-pred Settings [Required]-
Advanced Model Sampling for reForge (Backported) - On
RescaleCFG for reForge - On
Extensions
This is the most important section and pretty much the reason why I made this article. I will separate the extension into two categories. One, image generation extensions and two, UI extensions.
-Extensions: Image Generation
APG's now your CFG - After many test generations, I feel this extension improves the overall image coherence. I have all the settings set to default except APG Norm Threshold which was changed to 25. [Not used for V-pred models]
Sliding Window Guidance - This improves the prompt comprehension and image coherence greatly. I'm really shocked how well this works. I leave all the settings to default. [Not needed for V-pred models]
Note: Sliding Window Guidance is part of a series of extensions within this add-on called Perturbed-Attention Guidance. The other features are also worth checking out.
MultiDiffusion - I wrote an entirely different article for this. You can read it here. Basically, it's the method I use to upscale all my images.
-Extensions: UI extensions
Stable Diffusion WebUI Aspect Ratio selector - Quick access to custom aspect ratios.
SD WebUI Tag Autocomplete - Auto complete Danbooru tags. Pretty much required if you're using an anime based model like illustrious.
sd-webui-prompt-all-in-one - This extension take prompting to another level. It works perfectly with autocomplete. It allows you to save tags, easily rearrange tags, quickly adjust weights and comes with a list of tags you can choose from if you're running low on ideas. I can't recommend this enough.
Stable Diffusion webui Infinite Image Browsing - A fast and beautiful image browser right in the UI. I use it to keep track of all my previous gens. You can quickly access the metadate of any image or just send it straight to where you need it. If you gen a lot of images, this will be very helpful for you.
Control Net
CN-anytest_v4-marged
diffusion_pytorch_model_promax
Note: I don't use control net when I generate images for my lora pages as a rule. I just use control net for personal images.
Conclusion
I've used many other methods and extensions but I've eventually dropped them as I fount they didn't impact the final image well enough to justify their use. I will continue to evolve my settings.
Updates
Mar. 8 - 2025 - Added settings for v-pred models. Made various changes.
Feb. 20 - 2025 - Changed Sampling Method from DPM++ 2Sa to Euler a. And Sampling Steps from 25 to 45. Sliding Window Guidance has made Euler a at 45 steps a much better choice than DPM++2Sa. The final image is cleaner and even more coherence. Also, added ADetailer to set up.