This model's been created by merging some older models like Siren Forge (retired) with some newer experimental merges focused on photorealism, with the intention of using the images generated to train downstream models/lora's.
Usage is quite simple; keep the descriptions short, it's not meant to follow prompts to the letter. The idea is to make a simple image (eg. "Blonde 45yo woman, wearing a red dress, standing in the middle of a forest path", which will generate an image that will then be more accurately captioned using a tool like JoyCaption.)
This is primarily to generate synthetic data for training, so don't expect an all in one model that can be used for everything under the sun. Also, it's usually pretty NSFW, so be careful!
Generation parameters:
5-10 steps, with an upscaler like 4x_UltraSharp 4-8 steps. The Sampler/Scheduler is LCM Exponential/Polyexponential. Image size, 512x512, 512x768 with 2x upscaling.
Description
DMD2, VAE baked in.
FAQ
Comments (5)
Siren Forge is a great model that will live permanently on my hard drive. Why did it need to be retired?
Thanks for the kind words. The base was becoming a bit too unstable for merges. I'm not sure why but there seemed to be quite a bit of loss of basic 'understanding' of concepts when merging with newer models or lora's. I think it just had too many merges that fought with each other...I'm working on a new base that I can train a bit easier, though, and so far it's better than Siren Forge, although not quite ready yet.
@TrueToLife_Fauxto Ah, thanks for the explanation. I've only recently gotten into checkpoint merges but I've already noticed that there's a point where too much becomes too much.
i have no idea why you saying 512x768 but your model is good , keep tweaking more models push them more better. it just like making people arms shorter , maybe error from me also idk.
Best model for refiner.