sclover13
Content Creator
- Joined
- Nov 29, 2012
Stable Diffusion SDT Content Generation
I have been experimenting lately with generating SDT content using stable diffusion with varying success!
Using a twin set of LORAs I have made a reliable means to generate accurate hair models (for either tracing reference or to edit directly for PNG static hairs) and to generate clothing references for SWF creation.
While generation is about 1-5 minutes on a RTX4070 per image, this is a passive thing and configuring the prompt to activate the LORA is realitively striaght forward.
I have run a few tests and come up with some optimal settings for generating OC hair/clothing or Character hair/clothing. Some characters are known quantities while others require a separate LORA to inform the generation.
Downloadable Mods Made from AI Generated Content:
First generations had pretty terrible results with hair and body elements not appearing 'in style'.
Tweaked the prompts and parameters for the hair generation LORA..
Right now, I'm able to toggle between "Upper Body" and "Full Body" generation using those activation prompts, respectively.
My Prompts
Positive Prompts
Negative Prompts
Parameters
Hires 2x, 16 steps, 0.33 to 0.5 denoise. 4x Ultrasharp
Model trained on ponyDiffusionV6XL_v6StartWithThisOne and will work with various Pony flavors.
Lora Download
Instructions
[ to be added ]
Using web automatic1111
The Loras were trained with 819w and 1024h. Using this size yields best results.
Using negatives like "(((out of frame:1.3)))", "cutoff", "cropped image" prevent the hair from being cropped out.
Other Uses
While these LORA are designed for hair creation specifically, I've also experimented with generated custom body parts for ExtraMod (noses, mouth, ears, better feet/hands, etc) as well as creating clothing templates.
Clothing generation examples:
I have been experimenting lately with generating SDT content using stable diffusion with varying success!
Using a twin set of LORAs I have made a reliable means to generate accurate hair models (for either tracing reference or to edit directly for PNG static hairs) and to generate clothing references for SWF creation.
While generation is about 1-5 minutes on a RTX4070 per image, this is a passive thing and configuring the prompt to activate the LORA is realitively striaght forward.
I have run a few tests and come up with some optimal settings for generating OC hair/clothing or Character hair/clothing. Some characters are known quantities while others require a separate LORA to inform the generation.
Downloadable Mods Made from AI Generated Content:
First generations had pretty terrible results with hair and body elements not appearing 'in style'.
Tweaked the prompts and parameters for the hair generation LORA..
Still had to work to get the framing, background, and style correct. However this first real set of generations proved that the LORA might work out for reference images.
Some examples of specific characters made with AI
Right now, I'm able to toggle between "Upper Body" and "Full Body" generation using those activation prompts, respectively.
My Prompts
Positive Prompts
Code:
score_9, score_8_up, score_7_up, [source_anime], <lora:sdt-hair1:0.9> sdchan, 1girl, solo, from side, mouth open, nude, neutral lighting, (full body, (white background)), feet, hands on thigh, <lora:Anime_Screencap:0.35> anime coloring, [[female focus]], breasts, kneeling, (detailed hair), <lora:add-detail-xl:1.1>,
BREAK <!--Break is not necessary -->
<!-- Insert all your hair prompts here. Things like long hair, short hair, twintails, low ponytail, wavy hair, ... or character lora or info will work well -->Negative Prompts
Code:
Negative prompt: score_5, source_realistic, watermark, censored, blur, (((out of frame:1.3))), expressive lighting, closed mouth, cutoff, cropped, (teeth, tongue), (bad anatomy), huge eyes, looking up, looking down, hair out of frame, washed out, dark, low quality, close-up, deformed, horror, awkward position, [custom nose, custom mouth], (background),Parameters
Code:
Steps: 24, Sampler: Euler a, Schedule type: Automatic, CFG scale: 7, Seed: 2504589038, Size: 819x1024, Model hash: ac006fdd7e, Model: autismmixSDXL_autismmixConfetti, VAE hash: b8821a5d58, VAE: sdxl_vae.safetensors, Denoising strength: 0.38, Clip skip: 2, Token merging ratio: 0.2, Token merging ratio hr: 0.2,Hires 2x, 16 steps, 0.33 to 0.5 denoise. 4x Ultrasharp
Model trained on ponyDiffusionV6XL_v6StartWithThisOne and will work with various Pony flavors.
Lora Download
Instructions
[ to be added ]
Using web automatic1111
The Loras were trained with 819w and 1024h. Using this size yields best results.
Using negatives like "(((out of frame:1.3)))", "cutoff", "cropped image" prevent the hair from being cropped out.
While these LORA are designed for hair creation specifically, I've also experimented with generated custom body parts for ExtraMod (noses, mouth, ears, better feet/hands, etc) as well as creating clothing templates.
Clothing generation examples:
Last edited:
