Hello world
Let's get real
It's been a while since the last update, and this iteration is quite different than the last. That's not to say it is necessarily groundbreaking compared to anything else on this website (it's a merge after all), but I've tried to focus on a couple of very specific things:
A high amount of realism without any use of LoRAs or embeddings.
Having a good basis provides a lot of room to go crazy with those later.
Functional with a limited/small negative prompt.
Similarly, a small negative prompt often makes for the most realistic outputs. It is always possible to write your novel, or add embeddings in there along the way.
Consistency in poses / hands / other features within these constraints.
Decent control over light and darkness through prompt alone
There are great LoRAs out there for lighting / noise offset, but there is already some great functionality by default.
Responsiveness to the wildest lists of wildcards ChatGPT has to offer.
I like to have random lists of locations, time periods, clothing/materials, concepts, all generated (and embellished) by your local GPT.
Recommended VAE (not included): vae-ft-mse-840000-ema-pruned
-High base resolutions are advised: such as 640x896 for a portrait format
-If possible use hires.fix or a similar implementation: a denoise of 0.3~0.5 will do well in most situations various ESRGAN upscalers give good results. x1_ITF_SkinDiffDetail_Lite_v1 is the one used in all examples here, but there are other great options.
-Don't feel obligated to use time consuming samplers such as the one in most examples. I was suffering from a short bout of SDE Karras addiction when generating the examples, but I have since come back to 2M Karras @ 15~20 steps, which is usually my baseline and more than fine on its own (even with hires.fix).
*Note that some of the example prompts can generate nudity, prompt against this to prevent it. For actual examples of nudity you can check out the album uploaded separately.
**The model in the metadata is QGO-3002a, which is the same as QGO-PromptingReal-v1
Description
There is a Safetensor file that is "QGO-mk8" with the dreambooth added on top. Then there is a smaller Pickletensor file that is the raw dreambooth trained on 1.5
This dreambooth will most likely make your model unusable for anything else. The outputs will often be broken (also depending on what you add it to), but sometimes the outputs can be great. Is triggered by the word "cum", but will probably appear regardless.
Use "Add difference" to add it on top of another model, for example:
A: YourModel + ( B: cum dreambooth - C: sd1.5 )
which subtracts sd1.5 from the dreambooth, and adds that to a model.