Overview

The goal of StoryMix is where the boundaries between fantasy and reality merge. With a humble start by discovering its current capabilities and flaws. Future updates will be focusing on improve the model and make it easier to use.

Positive Prompt: best quality, masterpiece, ultra high res, (photorealistic:1.4), raw photo, sharp focus, HDR, detailed skin

Negative Prompt: Anime, cartoon, ng_deepnegative_v1_75t, (badhandv4:1.2), (worst quality:2), (low quality:2), (normal quality:2), lowres, bad ???????, bad hands, normal quality, (monochrome), (grayscale) watermark.

Merge Based On:

1. https://civitai.com/models/8124/a-zovya-rpg-artist-tools

2. https://civitai.com/models/47130?modelVersionId=70157

3. https://civitai.com/models/43331/majicmix-realistic

(Optional) Recommended VAE
to get a better high quality result, please use this VAE:

• https://civitai.com/models/76118/vae-ft-mse-840000-ema-pruned

(Optional) Recommended Textual-inversion

to prevent bad hands and ugly results please use these textual-inversions:

• https://civitai.com/models/16993/badhandv4-animeillustdiffusion

• https://civitai.com/models/4629/deep-negative-v1x

Steps used when I generate images using this model

1. Start prompt engineering

   1. Using 'txt-2-img',
      or 'img-2-img' using existing images & prompts if you want

   2. Start simple without After Detailer Extension

   3. Settings:

      1. Sampling method = Euler a / DPM++2MKarras

      2. Sampling steps = 20-60

      3. CFG = 5-7 for more creativity, 10-15 to be strict

      4. Size: 920 x 512 px (portrait)
         - Or 512x920 landscape mode that tend to generate more characters
         - Or 512x512 if you prefer to use the result for training

      5. Start with batch count = 1-4

   4. Edit the prompts until reaching the intended results

      • (Recommended) Ask chatGPT to generate text-to-image prompts suitable for Stable Diffusion according your requirements

2. Polish the Details (if needed)

   1. Keep using 'txt-2-img' with the config in step 1

      • or choose the image you like and 'Send to ImgToImg' if you want it to be very similar

   2. Reuse seed from the previous image
      (Useful when we want to keep some major details like faces)

   3. Use the After Detailer Extension

      1. ADetailer Model = yolov_8n

         1. For most cases, first detailer is for face

         2. Second detailer is for hands

         3. Set Confidence = above 65 to avoid false detection

   4. Settings:

      1. Sampling method = Euler a / DPM++2MKarras

      2. Sampling steps = 40-100 for a better quality

   5. (optional) Use DeepBoru to improve prompts but be careful to also check missing prompts not covered by DeepBoru

   6. (optional) Fix minor flaws with inpainting/sketch/control-net or other tools

      1. Inpainting:

         1. Erase area to improve

         2. Then generate with high resolution

      2. Sketch/Control-net: rarely used, but can be useful to improve minor flaws:

         1. Bad hands can be fixed with control-net depth map

         2. Bad pose with control-net OpenPose

         3. When prompt is hard, minor tweaks can be improved with sketch, e.g. adding missing legs or characters

3. Send to Extras

   1. Scale to = 1080 x 1920 / 1920x1080

   2. Upscaler1 = SwinIR_4x or other 4X scaler

   3. GFPGAN visibility = 0.95

   4. (If it needs cropping) Crop to fit = checked