Note: For a discussion of all things related to realistic AI image/music/video generation, join us and share your art with other model creators and enthusiasts at our new Discord server, Unstable AI.sland: https://discord.gg/B6BSCbdAJX.
DMD2 (Speed) Version release:
This version is a DMD2 4-step Lora merge of XXX-Ray V1.1. No need to download/activate the Lora separately.
DMD (Distribution Matched Distillation) allows you to generate fairly good quality images in as little as 4 steps with the LCM sampler. It cannot produce images equivalent in overall quality to a non-DMD2 checkpoint, so please keep this in mind. On the other hand, with 6-10 steps, this thing is fast as hell and outputs more than reasonable generations—which is great for anyone without a high-end graphics card.
No PAG needed for this one. Just ~8 steps, LCM w/Exponential at CFG ~1-2.0. A convenient workflow is provided that will get you off the ground running with text-to-image, an additional DMD2 fast upscale with a Tile ControlNet (to keep things intact while adding detail), and a bonus FaceDetailer node that also takes advantage of the speed of DMD2/lcm. A major weakness of DMD2 is that the low-cfg requirement eliminates attention to the negative at cfg=1.0. My workflow adjusts the noise tonemap and rescales cfg so that you can boost cfg a bit and still have that negative prompt perform somewhat functionally. Get the workflow here: https://civarchive.com/articles/13329/xxx-ray-dmd2-basic-workflow
DMD2 Pros:
Fast as fuck, boi (i.e. ~1s with a 4090 at t2i)
Good anatomy--e.g. You can generate at 1024x1440 at t2i without anatomy horror (mostly), which might eliminate the need for additional upscaling. If you decide to generate at this resolution, the upscale and facedetailer portion of the provided workflow will be overkill.
Sharp
DMD2 Bad:
LCM isn’t great. Background details are a bit blurrier/painterly
Low steps means less attention to prompt and less fine detail overall
Slightly less creativity in terms of pose and composition
Some options:
Download the XXX-Ray DMD2 checkpoint, download the workflow (https://civarchive.com/articles/13329/xxx-ray-dmd2-basic-workflow), get the custom nodes + CN models required for that flow (links in the workflow), generate. This flow is optimized for XXX-Ray DMD2.
Download the XXX-Ray DMD2 checkpoint, use it in your own flow with the following settings:
LCM
Exponential (or SGM-Uniform)
8-10 steps (6 also works acceptably)
CFG ~1.0-1.2
Notes: Disable PAG, SAG or any other modifiers that affect CFG when using DMD2. You’ll get a burnt mess if you leave them on.
What is XXX-Ray:
This is a photorealism model merge using (block) DARE ties w/model masking. It was tailored to produce slightly darker and grungier images, which is my own stylistic preference. I'm primarily releasing this for friends who've been asking for it, but also anyone else who is looking for a "moodier" realistic model that has good skin texture, lighting, and prompt adherence, but doesn't want that "Pony" look. No Pony derivatives were merged into this mix (v1.0). If you like the types of images that I generate, you'll probably like this model too. Good for both sfw and nsfw.
Many of your favorite realistic models are included in this merge to some degree, so expect similar results. It has been tested against other popular merges with X/Y plots and is able to generate unique output. Great for selfies, polaroid-style images, and analog shots in general. Good for upscaling too, though perhaps not as good as the incredible Lustify at the latter. It gets skin color/tone right most of the time. Its weaknesses are eyes (which is an easy fix with a detailer) and the occasional floating wang. Also not great for anything too fantastical. Overall its pretty good, and once in a while knocks it out of the park. Have fun with it. And remember, you get what you pay for.
The example images include a basic Comfyui workflow that I whipped up to help you get off the ground. XXX-Ray has been tested in ReForge as well and works as intended with the suggested settings below.
Some advice for its use (V1/V1.1):
Use PAG (Perturbed Attention Guidance) at ~1.5-2.0.
Asp tokens were kept intact as best as possible in the merging and work well: (https://gist.github.com/fpgaminer/0243de0d232a90dcae5e2f47d844f9bb).
The Detail Daemon sampler (~0.04), like PAG, will boost output quality.
Using score tags (e.g. score_7_up) will affect your generations slightly, but they are not strictly necessary.
Use the included comfyui workflow if you're having issues: https://civarchive.com/articles/9981
Upscaling is your friend. This is also included in the provided workflow.
Settings that have worked well for me with this merge (V1/V1.1):
DPM++2M sde Karras at 23+ steps, DPM++ SDE Karras at 18+ steps, or DPM++2M sde AYS at 14+ steps.
CFG ~4.0-4.5.
PAG (strongly recommended) at 1.5-2.0.
If you'd like to use DMD2, use the 4-step LoRA, LCM w/Exponential at 8-steps, CFG = 1.0-1.5.
Note: Credit to @nutbutter, @erotes_anteros, @coyotte and @ndimensional for their masterfully trained checkpoints. The community should be ever grateful for their efforts.
Big Asp: https://civarchive.com/models/502468/bigasp
Lustify: https://civarchive.com/models/573152?modelVersionId=1094291
NatVis: https://civarchive.com/models/617652/natvis-natural-vision
Anteros XXXL: https://civarchive.com/models/430454?modelVersionId=479579
Description
Keeps the same texture, lighting and moodiness from v1.0 but is now slightly better at generating consistent and coherent anatomy. Also has better eyes, more fine detail (e.g. skin), and better contrast and shadows. Does not improve all aspects of v1.0, so you'll have to experiment to see which you like best.
FAQ
Comments (11)
I am impressed with the quality, it provides excellent results without the need to use so many loras, it has compatibility problems with some loras, but they are almost not necessary
Anything you can use in the prompts to stop people getting their pussies or cocks out all the time. Just don't want this every image. Thanks
Prompt real careful and don't mention anything sexual in the prompt. No breasts, no ass, no nudity. You get it! Instead describe clothes.
YUP! that seems to be an effective strategy for tucking away the dangly bits. Thanks for everything Subtle you truly are the checkpoint FKN maaaastrerrrrrrrr
If you are using DMD, Hyper, other tubro loras, you can't use a negative prompt. But there is a method called "NegPiP". It allows you to use negative tokens in your positive prompt, like this: (penis:-2.0) (pussy:-2.0)
For ComfyUI you need to download a custom node (ppm). Don't know how to do it with other programs.
@zz5 Never seen negative weights mentioned before. Thanks, will try. But you can use negative prompts with speed loras if you use a cfg above 1. Even with DMD2 you can use cfg of 1.25 or 1.5, even more if you reduce the DMD2 strength or use the rescale cfg feature. I did not use any negative prompt in multiple months. Easy to adjust the positive prompt to achieve something similar.
@SubtleShader setting cfg slightly above 1.0 and using the negative prompt doubles the inference time. Also I found it to be less effective compared to NegPiP. But NegPiP has one drawback: the more negative tokens you use, the less positive tokens you can use (SDXL has a 77 tokens limit), but I never found it too restrictive.
@zz5 No double time here.
@SubtleShader If I only knew where "here" is :) I was talking about ComfyUI
@zz5 Forge
I love this model for realistic things. Not great at fantasy or distorting real things.