@ElectricDreams No worries, I appreciate the checkpoint! From what you have posted, this just might be one of the best photorealistic Illustrious models I have ever seen.

I experimented with 3D/photorealism in Illustrious before and it was so time-consuming to get it even remotely close to accurate. So from what I have seen in the gallery thus far, this looks very promising!

@ElectricDreams Some immediate feedback I can give is that it is (in my opinion) very good. As far as my earlier comment, I reaffirm that this is one of the best photorealistic models out there for Illustrious. It will be my go-to in the future!

In my experience, Illustrious LoRAs were able to shine through the photorealism at CFG 1.5 with acceptable losses to visual clarity. Some of them had noticeable impact even at CFG 1, with no loss to visual clarity.

All of this is to say, good work!

@ElectricDreams  EH Nexus Illu actually works fairly well with several different LoRAs I've tried, even SDXL and Pony versions. Not everything, but a lot!

@ElectricDreams I m using Forge to generate pictures. So lowering the cfg score to 1 means disabling the negatiuve prompt. This isn't suitable as you will not generate exactly what you expect the first time most of the time. You will often have a finger, toe, or something else to deal with using your negative prompt. Also i don't see any info regarding which tool has been used to generate pictures that you ve posted but i presume that most people ll be using ComfyUI.

On the LoRA side, as most people would expect with a hybrid model, you will need to ensure that this checkpoint works with the specific LoRA you want to use, as this is a case-by-case compatibility issue.

@ElectricDreams I m going to post a couple images without any Lora usage at cfg 1 / 12 steps and at cfg 1,5 / 12 steps. So you ll see what i mean by too saturated to look realistic. But as stated above if you can get rid of this issue then this would be one of the most photorealistic checkpoint for sure. Also, if possible, try to solve the strong tendency that your checkpoint has to turn vaginal sex into anal sex.

@ElectricDreams Glad to help, thanks for your work on this checkpoint. From my tests, using more than 8 steps results in images that are way too saturated. But even at 6, it's still a bit too much for a realistic rendering. And this is the parameter that has the most impact on image saturation. Of course, the higher the CFG score, the more saturated the generated image, but this is the same for all checkpoints. So it's advisable not to exceed 8 steps and stay at 1 CFG for less saturated images, in my opinion. One last comment though. After playing around with this checkpoint for a bit, I have to say that it seems to have trouble following your prompt. So for complex images that don't depend on luck, this might not be a good choice at the moment. But that's the case for most DMD2-type checkpoints.

@ElectricDreams Indeed. Anyway, as said above, thank you very much for your work on this checkpoint which has a lot of potential, that's for sure.

Any chance of putting out a "non-accelerated" version? As an RTX 5070ti owner, speed isn't really an issue, so I'd rather have stronger prompt adherence via higher CFG than worry about a few seconds of render time. Please and thank you.

@ElectricDreams  Illustrious adherence is usually the best ime. Personally, I'd like to see a non-turbo version as well, but just cause my software doesn't support the LCM sampler.

@ElectricDreams Definitely understandable. I downloaded Nexus Illustrious and plan to give it a quick try with Euler A just to see what it does, but I'm thinking it probably won't work out too well. I'll still enjoy seeing the results that you and @6__6 make with your models. Take care!

@ElectricDreams  I'm using Easy Diffusion on my main rig and LCM isn't an available sampler choice (guessing it might be included if/when the new version comes out of beta). My new laptop is running Intel's Playground AI, which I think supports LCM. I'll have to give it a try with that.

@ElectricDreams @Slaughterford  So I just figured out the silliest thing. Of course Nexus Illu looked a bit nasty using Euler A, so I loaded up the FP32 DMD2 LoRA from @6tZ at -0.8 to -1 strength and now it behaves exactly like a non-turbo checkpoint even using Euler ~5 CFG and 40 to 60 steps! Surprisingly, the image quality and realism isn't bad at all. Maybe not as good as using LCM, but at least now I can mess around with all the recent turbo checkpoints I've been skipping lately. Think I'm gonna try the smaller FP16 DMD2 LoRA to maybe cut down on load times.

@AFD_0 I just tested it with the DMD2 FP32 at -1 and it works pretty well. I'm using Eular A and tried it with DDIM and Karras, and both look great. Prompt adherence still needs some work even at CFG 7, but it still looks great regardless.

@ElectricDreams  I absolutely love using your model this way, but I'm really not sure exactly how faithful it is to your original vision. Everything I've tried gives relatively comparable results to most other modern realistic illu models I've been using (and in many ways much better), but I can't seem to accurately mimic the results you got with your sample prompts, so not sure if I'm actually able to get the intended experience, if that makes any sense. Whatever it is, it's incredibly fun and I'm really looking forward to trying it with your SDXL Event Horizon! Unfortunately, I'm usually far too verbose to get good results from most of those these days :D

@ElectricDreams That's awesome news! Just be sure that doing so doesn't sacrifice the vision you've been achieving with the past EH and EHN models. The results that both you and @6__6 have been posting are literally some of the best-looking photo realistic pictures I've ever seen on this site! As an aside, I did try using the smaller FP16 DMD2 LoRA at negative values (around -0.8) and the resulting image quality maybe wasn't quite as good as when using the FP32 version negatively (at -1 things start to wash out for me), but either works pretty well. I'm trying both stacked at -0.4 right now just to see what it does (edit, it gives me double subjects!), but the larger FP32 version alone at -0.8 with ~4.9 CFG and 45 to 55 steps (needed for my longer prompts) in Eula A has been outstanding imo. Anyway, best of luck and thank you for all the work you've done - I'm really enjoying it! <3

Thanks for your work/models !!!!
but NO : it doesn't sounds "really good" at all (imao...) to be "invited" to go to another site then to register and then pay for downloading the model. But i understand your strategy and TensorArt's one... do as you wish

If its strategy, i dont really understand. The tensor exclusive looks like V1 or even more realistic, why your xl models here become more and more artsy.

What's not cool is a freeloader not appreciating that people spend money on training models.

@ElectricDreams Great anatomy btw. Fingers and toes always perfect. A lot of doubling of the subject and ignoring parts of the prompt unfortunately. I really liked the look i dragged out of the normal EH model, but in that one I can't make the limbs work. All the samplers/schedulers that give me those slightly desaturated images cause body horror :(

https://i.postimg.cc/vmWq9NXn/EH30.png

@ElectricDreams  That's really sad to hear! They're both excellent models and each excel in their own ways, but I personally get better adherence, photorealism and just have a lot more fun with Nexus Illu than with EH SDXL. They're both really good models, just different.

Was really hoping we'd get a non-DMD2 Nexus Illu or maybe a v2 :(

It's understandable, though. I always had the same experience with the Pony-heavy versions of @SubtleShader 's amazing Big Love model. Everyone else got better results with the SDXL version, so it was eventually dropped as well. It's a bummer tbh, but I completely respect you both for deciding to stick to your own visions of what you're each wanting to creatively achieve. Think that's really important.

Do appreciate the one version of Nexus we got, so thank you! <3

@ElectricDreams Prompt adherence of anime Illustrious is good, but when SDXL was merged in it declined a lot. Was the same with Pony. Pure SDXL still works best for photorralistic. My Big Love Photo3 even does Illustrious prompt more realistic than so-called Illustrious realistic models. Tried it yesterday with Illustrious prompts. But had to remove the useless fill words first.

@AFD_0 https://civitai.com/models/2037441?modelVersionId=2385148
You can try this version with non-LCM samplers.
It is the DMD2 offspring allowing to use various other samplers.
Working combinations in model description.

@homoludens Your link does not work!

@SubtleShader Hmmm It seems correct. Checked it - it opens the CinEro NoiceAI

@homoludens Works now. Did you reduce the DMD2 component when merging? Or what did you do?

@SubtleShader 
TL;DR; I did not reduce the DMD2 weight, but I used the training and MBW adjustments to make CFG range wider and allow other samplers.

I just created a usual DMD2 based checkpoint (with DMD2 weight == 1.0) from my CinEro ILL R4e then trained that derivative in multiple stages with few dataset re-creations. As a result, ability to get a viable results from 6 steps still there but LCM is no longer the only sampler that works. During training I used some old LORAs to adjust CFG sensitivity (to make the CFG range wider).

@homoludens Interesting experiment. Looks like the DMD2 effect was reduced by the training resulting in a HDR-like effect. But not really my taste.

@homoludens  I'll give it a try without DMD2 LoRA, but rarely have good experiences with models that require ultra-low CFG and low steps. Think I'd have better luck using DMD2 at negative weights with normal CFG/steps, but we'll see what it does!

@AFD_0 Firstly, try the LCM + CFG 2.0..2.8. Add a HiRes Fix with Euler A + Exponential with CFG 3.0 and Denoise like 0.3 or lower.
Also, you can go with Euler A + Exponential for Text 2 Image and any other sampler for HiRes Fix.
If you can use CFG 1.5, then you can start with Euler A + SGM_Uniform / DDIM_Uniform / Karras.

@SubtleShader HDR is a side effect of sup-optimal CFG adjustment. Will fix it soon. Right now I compensate it with my LUT advertized in recommended resources. New dataset almost ready and one of 3 parts already used in running training on Civit.
"Looks like the DMD2 effect was reduced by the training" - boring composition and poor narrative diversity were washed out by rich and diverse datasets during training, yes. But main positive part of DMD2 looks intact - the speed. Creativity OK. Adherence not yet deeply tested. Textures already good, and improvements ongoing.

@homoludens  Yeah, my software doesn't support LCM, so I haven't been going out of my way to mess with most accelerated models, other than a few I've used with negative DMD2 LoRA weights. CFG 1.5 with Euler A will probably look bad with my prompting, but I'll give it a try with and without negative DMD2. Think I'm already regretting this, but you're a really good dude and have surprised me before.. ;)

@AFD_0 I tested Euler A at CFG up to 3.0 in Text-to-Image. Should work with minor issues like ground texture too grainy or some rough edges in background. Test it and compare with others if you can. That will help to improve.
Use a mix of Pony tags and natural language