Daru
Well-Known Member
Hello frens. I wanted to share some results from a Pehpeh trained image AI (Stable Diffusion) I made. It was trained using LoRA technique, where I gave it some images and descriptions of Pehpeh so it could "learn" the character and style. The model gives decent results, even though I haven't done any fine-tuning whatsoever, so this can definitely be improved.
Some images have a noise/fuzzy background (see examples below) if I don't specify a setting ("in Toronto streets", "in a green field", etc).Not sure if it's from my training images having transparent backgrounds (this was fixed). This is my first time training an image AI so I'm still learning. Also I'm a noob when it comes to prompting both positive and negative prompts correctly.
If anyone else has worked on something similar before, let me know. I'll share my favourites below and maybe work on improving the model.
Also I can try to generate images if anyone has ideas of what they want to see. I can't guarantee good results though. For example I couldn't get some prompts to work (pehpeh refusing vaccine, pehpeh fighting Justin Trudeau, pehpeh holding rope noose in hand).
Enjoy!
PS: I am by no means trying to replace our wonderful Pehpeh artists. This was just a personal project I wanted to try.
Some images have a noise/fuzzy background (see examples below) if I don't specify a setting ("in Toronto streets", "in a green field", etc).
(pehpeh hacker black hoodie)
(pehpeh lifting heavy weight in the gym, jacked, muscular, strong)
(pehpeh smoking weed bong stoner high)
(pehpeh lifting heavy weight in the gym, jacked, muscular, strong)
(pehpeh smoking weed bong stoner high)
If anyone else has worked on something similar before, let me know. I'll share my favourites below and maybe work on improving the model.
Training
-I followed this guide and trained on 51 images of Pehpeh, taken from this thread. I didn't use the female version or any other characters.
-To label each training image I just specified what was Pehpeh's expression along with what he was wearing or holding.
-I trained on the base Stable Diffusion 1.5 model, following this guide. Only took 25 min on my 3070.
Generating Images
-I used Automatic1111 WebUI to create the images.
-The main model I used when generating images was Dreamshaper.
-The Realistic Vision model was also decent.
Prompting
Each positive prompt had this, with ______filled in for what I wanted:
(masterpiece, top quality, best quality, cartoon, 4k) pehpeh ______ <lorahpeh_LoRA_Model:1>
Negative prompt:
(worst quality, low quality:2)
Settings
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: -1, Size: 512x512, Model hash: e9d3cedc4b, Model: dreamshaper_631BakedVae, Lora hashes: "Pehpeh_LoRA_Model: da1816f42dfa", Version: v1.3.2
-I followed this guide and trained on 51 images of Pehpeh, taken from this thread. I didn't use the female version or any other characters.
-To label each training image I just specified what was Pehpeh's expression along with what he was wearing or holding.
-I trained on the base Stable Diffusion 1.5 model, following this guide. Only took 25 min on my 3070.
Generating Images
-I used Automatic1111 WebUI to create the images.
-The main model I used when generating images was Dreamshaper.
-The Realistic Vision model was also decent.
Prompting
Each positive prompt had this, with ______filled in for what I wanted:
(masterpiece, top quality, best quality, cartoon, 4k) pehpeh ______ <lorahpeh_LoRA_Model:1>
Negative prompt:
(worst quality, low quality:2)
Settings
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: -1, Size: 512x512, Model hash: e9d3cedc4b, Model: dreamshaper_631BakedVae, Lora hashes: "Pehpeh_LoRA_Model: da1816f42dfa", Version: v1.3.2
Also I can try to generate images if anyone has ideas of what they want to see. I can't guarantee good results though. For example I couldn't get some prompts to work (pehpeh refusing vaccine, pehpeh fighting Justin Trudeau, pehpeh holding rope noose in hand).
Enjoy!
PS: I am by no means trying to replace our wonderful Pehpeh artists. This was just a personal project I wanted to try.
Last edited: