Update README.md

README.md

@@ -8,7 +8,7 @@ tags:
 - art
 ---
 
-This repository is for the diffusers port of 2dn by Advokat, mainly so InvokeAI users can connect to the diffuser through the repo, which is preferable for remote users like myself. While putting it together, I decided to play around with tokenizer and scheduler settings. The result is slightly better, but mostly different, allowing you to explore different places in an already gorgeous model.  
+This repository is for the diffusers port of 2dn by Advokat, mainly so InvokeAI users can connect to the diffuser through the repo, which is preferable for remote users like myself. While putting it together, I decided to play around with scheduler settings, and learned a few things about how diffusers behave that was really fun and cool. The result isn't so much better as it is *different*, allowing you to another perspective of this already gorgeous model.   
 
 **Original Author: Advokat**
 
@@ -16,10 +16,10 @@ For checkpoints and a more detailed description, visit the original at CivitAI:
 https://civitai.com/models/4807/2dn
 
 **Changes made by me:** 
+- Converted to diffuser format from the original checkpoint/safetensor on CivitAI.
+  - I've noticed that not all diffusers made from checkpoints function the same as a diffuser as they do a checkpoint. At low resolutions, the checkpoint and the diffuser make nearly the exact same image with the same settings. However, when you start moving into higher resolution territory with or without high res options, they start to differ. Not in a bad way, the model is still solid and still performs well, but in a different way, as if the diffuser format is exploring a slightly different space in the model.
 - Switched out the default scheduler for EulerAncestralDiscreteScheduler
-  - I find it to have the best tradeoff results between speed and accuracy/quality. 
-- Switched out the default tokenizer for laion/CLIP-convnext_xxlarge-laion2B-s34B-b82K-augreg-soup
-  - Provides greater accuracy when reading prompts. 79.4% over OpenAI's 63.2%. Of the OpenCLIP options, I found the soup to offer the best images in all the models I've tried it on.
+  - I find it to have the best tradeoff results between speed and accuracy/quality. If you set your sampler to Euler_A ( or k_euler_a in InvokeAI), it should *FLY* like the wind and create awesome images at just 20-30 steps. 
 
 **Recommended sampler:** Euler_A