Release of Our Text-to-Image Model
We are proud to announce that Eight Buffalo Media Group has shared our first text-to-image model!
Both Checkpoint (.ckpt) and Safetensors(.safetensors) files types are available:
Huggingface - https://huggingface.co/EightBuff/8Buff_Gen
CivitAI - https://civitai.com/models/41928
Version 1 of our first text-to-image model still has a few issues. Please be patient as we continue to improve. Constructive feedback is always welcome.
This freely available model is a combination of many different models that have been mixed, merged, and specifically trained on a few things, such as people in glass jars. The goal is to create a strong general model that offers control similar to many anime models, while also providing a more realistic look and feel.
See image below for prompt matrix testing.
This model does require a good deal of prompt crafting, the trade off is you have a lot of control on the images you create. I would recommend finding a prompt that generates the style and quality to your liking and saving it as a style.
Current known issues/limitations:
More people you add, the more likely you will get blurry/distorted faces
Tends to make women young looking. Add negative prompts to help this.
Tends to make men old looking. Add negative prompts to help this.
When putting people in jars, faces tend to be blurry/distorted. The more jars and more people, the more blurry/distorted. Prompts and and restore faces can help this a little.
DPM++SDE Karras sampling method tends to get the best consistent results, but experiment.
The aim in training and merging this model was to create a general purpose model with a fine level of control. Merged with a lot of different models, but don’t remember all of them. Links below to those that were remembered or provided inspiration for direct training.
Meinahentia - https://civitai.com/models/12606/meinahentai
Girls in Glass Jars (inspiration) - https://civitai.com/models/10453/girls-in-glass-jars
Shampoo-mix - https://civitai.com/models/33918/shampoo-mix
Stable Diffusion - https://huggingface.co/stabilityai/stable-diffusion-2-1
Please review and respect any related licenses. See https://huggingface.co/stabilityai/stable-diffusion-2-1 for the post known license related to known used models.
Common Negative Prompts Used:
(((worst quality, low quality, bar censor, distorted face))), (((jpeg artifacts, signature, watermark, text, username))), blurry, bad faces, blurry faces, bad eyes, bad anatomy, bad hands, error, extra limbs, missing fingers, extra digit, fewer digits, cropped, normal quality, missing fingers, extra digit, fewer digits, censor, black and white, monochrome, NSFW,
Enjoy!
Comments
Post a Comment