inference: false
tags:
- stable-diffusion
- stable-diffusion-diffusers
- text-to-image
license: creativeml-openrail-m
Stable Diffusion TrinArt Derrida model (Characters v2)
Derrida (formerly TrinArt Characters v2) is a stable diffusion v1-based model that was further improved on the previous characters v1 model. While this is still a versatility and compositional variation anime/manga model like other TrinArt models, when compared to the v1 model, Derrida was focused on more anatomical stability and slightly less on variation due to further multi-epoch training and finetuning. The pre-rolled augmentation phase by generating slight variations w/ img2img is still applied right before the finetuning phase. This is the same model that was released in AI Novelist/TrinArt service from mid-Oct through early November.
Hardware
- 8xNVIDIA A100 40GB
Custom autoencoder
Note: The autoencoder uploaded here is the same checkpoint as v1.
We also provide a separate checkpoint for the custom KL autoencoder. As suggested by the Latent Diffusion paper, we found that training the autoencoder and the latent diffusion model separately improves the result. Since the official stable diffusion script does not support loading the other VAE, in order to run it in your script, you'll need to override state_dict for first_stage_model. The popular WebUI has the script to load separate first_stage_model parameters.
Safety Consideration
The dataset has been filtered to avoid extremely NSFW materials, but slightly less strict than v1. As with any other image generation model, we don't recommend deploying this model publicly without safety considerations and measures. Depends on prompting, one may still be able to extract highly questionable images from this model. This statement does not necessarily restrict third-party from training a derivative or mix of this model that includes NSFW.
Examples
Below images are directly generated by the native TrinArt service with its idiosyncratic upscaler, parser and processes. Your mileage may vary.
TrinArt 2022 Artstyle preset negative prompts: retro style, 1980s, 1990s, 2000s, 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
TrinArt More Details preset negative prompts: flat color, flat shading
We recommend to add known sets of negative prompts in order to stabilize the anatomy such as: bad hands, fewer digits, etc.
Credits
- Sta, AI Novelist Dev (https://ai-novel.com/) @ Bit192, Inc. Twitter https://twitter.com/naclbbr (Japanese) https://twitter.com/naclbbre (English)
- Stable Diffusion - Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bjorn
License
CreativeML OpenRAIL-M