Fixy - Exploring styles in Stable Diffusion

Hello, amazing Patreon!
This week I've generated 2000+ different images (some of them even NSFW) to test how the trained model would handle the different style exploration for Fixy. Check out the prompts and the denoising strength. It was not an immediate result. Out of the 2000+ generated images, 80% was not usable at all. Be patient, generate your grids (I used 4x6). But! Here's the big take away when using your own 3d render into the img2img tab:

"If you want to use "anime screencap" style prompt, make sure you are using a resolution 3:2 in ratio in SD (Stable Diffusion), otherwise you will get NOISY and JAGGY artifacted images"

I also found the following incidences:
1. After 8 (GENERATE) clicks, SD starts to use previous "noise" data, to accelerate the regeneration of new art. This is counter productive to "find" a new variation. 

2. To troubleshoot this problem (specific with anime screencap prompt), change the image resolution to something like 640x576 (make it 3:2 but slightly bigger). This forces the tensors to reevaluate UPscale instead of reusing lowres previous data.

3. Image size DOES affect creativity on SD. When you use the .93 Denoising strenght (deviate from the img2img artwork and use the ai own constructed promtps based off yours) the first batch will not be a very good one. Be patient, and after some "training" (push generate button x each time), you'll get better results. 

4. Once you get a very good result, instead of using the SEED number, USE THE ACTUAL IMAGE you liked from the batch, and regenerate that with a lower Denoising Strenght (to re-train) the GENERATE button. You'll see better results in the same session.

I'll be uploading a video about this exploration. You as a Patreon can get the images, so let me know if I should include all the material (including the NSFW) as reward this month. Thanks!

-Cheers!

Pierre.