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
See suggested generation parameters in the model description.
FAQ
Comments (49)
Forget Lustify Endgame, Natvis or BigAsp. This is now the new frontier checkpoint for SDXL. It doesn't get any better than this (so far)! Well, maybe BigloveXl is on par and the only real contender (anatomy is a little better there at cost of some iq, and sometimes faces are better in BigloveXL)
Now you have better skin detail, better color rendition/palette, better contrast... All what you asked for and more!!!
Truly an state of the art checkpoint finetuned to excellence.
Can this get even better yet? can we go further down the hole? we don't know. How this was made? was this made by aliens? This is truly made by the stuff dreams are made of. Make no mistake! We're approaching an Event Horizon in true image fidelity. We crossed the Rubicon and we're reaching uncharted territories of magnificience never experienced before.
A legendary checkpoint that truly means a before and after in SDXL history.
btw, I recommend using "lightning lora" with this. Sampling: DPM ++SDE, Schedule: SGE Uniform, CFG. 1.5-2 and about 8-10 steps.
CONS: watermarks once in a while, nothing too serious, it doesn't bother me too much, tbh. A lot of excellent checkpoints have the same flaw (like Lustify or BigLoveXl). It's a small price to pay for this quality work.
Besides, there are checkpoints with -a little bit- stronger anatomy (BigloveXL) due that one is an hybrid checkpoint which uses Pony tech.
In conclusion, whoever merges those two checkpoints could destroy the entire galaxy with its sheer power. I hope nobody does that. It would be waaay too powerful to handle for humankind. It would be totally irresponsable. Fortunately, i don't think we have the technology to do so.
Have a nice day.
nice review, which is why i am downloading, thanks
Is there a specific "lightning lora" you recommend? Cheers
@robbiephotograph4850 The standard 8-step lightning lora can work with this model, yes. You might get better results with the 4-step dmd2 lora, however. With that one, set your steps to 8 using LCM and SGM-uniform or Exponential, CFG ~1.0-1.3 (no PAG)
@RaymondLuxuryYacht aaaah Raymondluxuryyactch!! i'm your biggest fan!!
@RaymondLuxuryYacht cheers!
Unfortunately, for whatever reason, Civitai will never allow us to use this and other amazing models on their site :(
Some of us simply don't have a powerful enough GPU
There will soon be a link to TensorArt available for using this model. 50-60 free images per day.
@SubtleShader oh hi! what a coincidence! Now i've just updated my review of this checkpoint comparing this to BigLoveXL. I hope you like the conclusion.
@ElectricDreams Just testing it myself. Got some nice images so far. It has a different style than Big Love, different faces, more raw output. It was developed for different kind of images. So a matter of taste. Will likely merge it into Big Love XL2 as it has its unique style.
@SubtleShader i've been testing it for hours. Those two really trade blows. My conclusion is BL has better anatomy but i like the richer color palette of this checkpoint a lot more (maybe on flux would be "meh" but on sdxl is top tier). Yeah i've noticed some faces, especially when are at a some distance are sort of hit and miss. Closeups are a lot better, though. I still don't have an opinion about it... But you're probably right. Anyways, i'm glad there are people developing better checkpoints at really fast pace. Keep up the great work!
@SubtleShader Tensor's implementation is trash. The model definitely does not work well with their setup.
@ElectricDreams The consistent anatomy in BL is due to Pony, yes. Adding it or any derivative to a realistic mix (even at .01%) instantly alters the uniqueness of a model and pulls it towards the Pony "look." BigLove is great model, but @SubtleShader and I are looking for different things in our merging. Faces are easily changed/adjusted with a detailer node, so it wasn't something I really considered in the process.
@RaymondLuxuryYacht the more, the merrier then! looking forward to see how evolve each chekpoint. Amazing times for SDXL.
@RaymondLuxuryYacht Set it up like Big Love XL1 on TA. Same default settings. Will send them.
Celeb and lora capable? Thats the achilles heel of every porn model.
Does celebs nicely.
@SubtleShader It does, but not nicely. Recognise the famous ones from SDXL base, who has many datas , rest with less data are weak. Thanks to BigAsp.
@azlanshivank171 Only does as well as any XL model—maybe slightly better because of some
choice merging, but, it’s a merge. Not new training data.
Loras work great.
@RaymondLuxuryYacht I disagree. Prompted Selena Gomez, one with the most datas, result is pure likeness.
Celeb loras nowhere near, like on an all arounder, but thast ok. Bigasp erase everything.
@azlanshivank171 Use the allrounder with the lora in txt2img. Then apply a bigASP-based model (like Ray's or mine) in img2img to improve image quality. Or do it vice versa. Problem solved.
@SubtleShader Sorry, what? This is a bigasp merge, and why are you angry? I didnt even talk to you.
@azlanshivank171 Just gave you a solution. Stop complaining & use it. 😉 Or create your own merge that does it better. In any case have fun with your celebs.
@SubtleShader I just replied for the celeb likeness, dont be a crybaby.
@azlanshivank171 Who is the crybaby here? You are asking the impossible & you stated that yourself. So what do you want?
@SubtleShader nothing, lol. Forget it, only hope is a newly trained model, which has nothing to do with bigasp.
@azlanshivank171 You don't need to wait for a new model if you use existing techniques to work around model limitations. Or train your own private lora.
@SubtleShader How many loras are for Bigasp, 3-4:)))))))))))
@azlanshivank171 Could be 😂. But you can use normal SDXL loras too. Not sure how much bigASP the XXX-Ray model has, but judging from its output, it has more than my Big Love model, which supports all SDXL loras and the more realistic Pony loras. But XXX-Ray seems to do celebs better than my model.
So these guys are right, this is great. But -- it is really important to be using the Perturbed Attention Guidance like in Ray's example workflow. Trying to integrate the changes into my own workflow didnt work out immediately, but it does appear worth it.
Maybe it's the sampler itself, honestly. Get really good results using the recommended DPMPP_2M_SDE, compared to DPMPP_3M_SDE_GPU which I normally use. I also think I learned that the GPU samplers keep either more multi-run memory, or something. Cause that's almost exactly what I've noticed in these runs. Man, I bet if I actually read on this shit I'd understand why. If you have a source Raymond, I'd appreciate, but I'm appreciative already of what you've helped me experience.
I've installed all your recommendation and your checkpoint creates stunning images with them being applied!
Glad you’re enjoying it!
Pag and detail demon are crippled on Forge.
Similar feeling, when Stable Yogi advice Adetailer. Of course, because all of his model has shitty, deformed eyes, and cant even look at viewer.
Interesting, thanks. PAG on ReForge works well for me and others--haven't heard of any issues yet. Not sure about Detail Daemon, sorry--though DD is not really needed for this model. In ComfyUI it will give you a slight boost in detail (obviously), but it's not a requirement.
@RaymondLuxuryYacht Thanks. Just thought, its REQUIREMENT, otherwise it has bad outputs.
How does PAG work on Reforge? I tought its same code as Forge. I tried it on Forge, and absolutly not worth the speed penalty. I wouldnt say even the outputs are better, just different. A tech guy said, it was crippled on Forge. Demon lacks the cond and uncond mode on Forge, without it, it loses the whole usability.
@cadenaldren789 I'd have to see what your generation parameters are to help you troubleshoot it, but even without PAG you shouldn't be having these issues. PAG does help a lot, and I've done literally thousands of comparisons to confirm. Are you following the reqs in the description? Can you share your gen params?
@RaymondLuxuryYacht I am just testing it, no trouble yet. Thanx.
@cadenaldren789 The built-in PAG of Forge is useless, but the sd-perturbed-attention extension is good. I only ever used it to get reasonable results out of bigASP.
@SubtleShader There's a built-in PAG on forge? I've only ever used the sd extension. Hilarious.
@SubtleShader I used that extension, and im not impressed. What are your settings?
@emorydamarkus653 I only used it once with bigASP. With my own model I don't see any real effect as I'm using Lightning or DMD2 anyway.
I really want to get into running models like this locally since I have powerful enough hardware but don't know where to start. Anyone know of a good guide?
Search for a beginner guide here on Articles, download Foooocus or ForgeUi. Fooocus is for SDXL and Pony only, but very beginner friendly and good.
@graecynjhon747 As @graecynjhon747 suggested, Fooocus is a great, easy way to start. Forge or ReForge is a bit more complicated, but still pretty simple, and ComfyUI is the most complicated, but also the most versatile and customizable. In terms of tutorials, Youtube has a ton. I really enjoy Sebastian Kamph.
@RaymondLuxuryYacht Raymond, I have seen your stuff everywhere on here. I have been trying to take as much inspiration from your Loras, etc. but still struggling. Do you have any tips on being able to get consistent characters for NSFW? I use ComfyUI, and I am able to get consistent characters on Fooocus pretty easily, but not in Comfy yet.
I don't suggest Fooocus anymore since it will not be updated for new models. So when you'll want to use new models other than SDXL/Pony (and you will) you'll have to restart from beginning with Forge or Comfy.
That's what happened to me. Go for Forge.
For hardware, min 32G RAM and the best price/value is the RTX4060Ti 16G (but a 3060-12G is more than enough for SDXL).
@Brewce Yes, if you have good hardware, you want try other models. If you have 8gb, like me XL and Pony is the only option.
I used 1.5 with Easydiffusion, helluva good Gui, nothing can beat it since then. And it came with windows installer, haha, more than 1,5 years ago.