I believe this model to be very good at photorealism although it does have some artistic capabilities as well. Here are a few tricks to using the SDXL versions of GonzaLomo to its fullest potential:
Clip Skip is your friend. I recommend trying a prompt with both 1 and 2 values. 1 can make images that are more creative, artistic, retro or grungy whereas 2 is better at photorealism and the, um, physical attractiveness of the subject(s).
Generate at 1024x1536 or higher using the PatchModelAddDownscale node in ComfyUI. This will create a much more detailed image that can then be upscaled further, resulting in incredible photorealistic detail. Using that node, I can do an initial generation as high as 1280x1920.
You can play around with different sampler settings but I find that the absolute best results are with LCM/Karras, 12 steps and CFG 1.0.
Selective usage of the Self Attention Guidance node can sometimes give amazing results. I always have it in the pipeline and usually bypassed but turn it on when I think an image has potential but just needs a little more oomph like color saturation or contrast.
I get really interesting/creative results by generating with the V1.1 Semi-Real model and then resampling that image with the V2.0 Unity model at around 0.60 denoise. You really need to play with the denoise to find the right value that tips the image from cartoonish to photorealism but you'll be amazed at what's possible when you do. Note that V1.1 Semi-Real requires Clip Skip 2+ and will not work with 1 at all.
And lastly, always be upscaling! More pixels equals more detail equals more photorealism.
Note: I only use ComfyUI so I apologize in advance if you're having trouble with my model in some other tool and I can't help.
Current Models
V2.0 Unity
Essentially, this is V1.1 Real with a bit of Siren Forge, Cyberrealistic and Lustify added to the mix. The base is still Gonzales and LomoXL.
I called this version Unity because I originally wanted to have separate models for photorealism and artistic styles. However, I realized that I could get both in one model and created V1.1 Artsy for that purpose but it was a failure. This is my second attempt at that combination and I think/hope it's a success.
At the very least, I don't want to degrade the photorealism capabilities and just want to add more surreal, artistic or fantastical styles as an option.
The recommended settings haven't changed since V1.1 Real:
Sampler: LCM (also dpmpp2a and ddim can provide interesting results)
Scheduler: Karras, Exponential or Beta
Steps: 8-12
CFG: 1.0-1.3
Clip Skip: 1-5
I normally keep my settings at LCM/Karras with a CFG of 1.0 but I tend to vary the clip skip between 1 and 5 because those can change the image pretty drastically. Sometimes an image at clip skip 2 will be terrible but might become amazing with clip skip 1, and vice-versa.
V1.1 Semi-Real
I'm doing a little experimenting with non-photorealistic models. This is a merge of GonzaLomo v1.1 Real and ToonComix which is a toon model that does a really good job with detailed backgrounds. The whole Comix family of models is very good so check them out.
Sampler: LCM
Scheduler: Karras, Exponential or Beta
Steps: 8-12
CFG: 1.0-1.3
Clip Skip: 2-5 (1 is not supported)
Retired Models
V1.1 Real
This is a simple 60/40 merge between two of my favorite photorealistic models - Gonzales-NSFW-PonyV1/V2-DMD v2.0 and LomoXL (with a little bit of Lustify thrown in). Many, many thanks to the makers of those excellent models.
Gonzales is a Pony model so GonzaLomo inherits all of those capabilties. LomoXL has a very strong analog feel to it which helps give an edge to Gonzales' "pony-ness".
Sampler: LCM
Scheduler: Karras, Exponential or Beta
Steps: 8-12
CFG: 1.0-1.3
Clip Skip: 1 allows greater analog effect but lessens NSFW capabilities
Clip Skip: 2-5 lessens/removes the analog effect but allows more NSFW
Try this workflow to get started, if you need it.
To be clear, I created this merge for my own use cases. It's certainly not better than Gonzales for most things. I'm just hoping to inject Gonzales with some of the gritty, analog feel that I was looking for in some cases. LomoXL also allows for a wider variation of faces, especially with clip skip 1.
Enjoy!
P.S. This is my first model upload so I would appreciate any constructive feedback or comments.
V1.1 Artsy alpha (Not very good imo, use V2.0 Unity instead)
This was an attempted merge between V1.1 Real and a few art-focused models. The main purpose of this model is to be able to produce artistic styles but it has the tendency to try to force photorealism when prompted clearly for an artistic style. This model has been abandoned in favor of V2.0 Unity which does a much better job of combining the photorealism of V1.1 Real with artistic styles.
Description
I created this version of GonzaLomo v4.0 without the integrated DMD2 LoRA. There have been a few people that have asked for it and I recently came across a technique for removing DMD2 and resaving the checkpoint, so I figured why not.
Personally, I much prefer the DMD version of my SDXL models. I just think the realism and detail is superior. However, I will admit that the non-DMD version can come up with some pretty nice compositions, even if the detail is lacking. Therefore, I kind of think of this as similar to my Schnell model - better composition but needs to be refined by my SDXL DMD model to fix the detail issue.
Recommended settings:
Sampler: dpmpp_3m_sde or dpmpp_2m_sde
Scheduler: Exponential or Karras
Steps: 20-30
CFG: 3.0-5.0
Clip Skip: 1-5 (I like 3 best)
FAQ
Comments (31)
Noob question: My images are coming out with bad colours/posterized, like there's a VAE problem (using automatic1111). Is there a specific VAE I should be using with this somewhere?
I'm using the default VAE of Automatic1111, it works fine on my end.
Set you sampling method to LCM, Schedule type to automatic, sampling steps to 10, and CFG scale to 1. it should work as intended.
I'm a ComfyUI user, not A1111, so I can't say exactly what the problem could be. The checkpoint contains the standard SDXL VAE. @portaloopawesomesauce878 is right about the sampling settings.
Even on LCM / Karras / CFG=1 / Steps=10 / Clip skip 2 images are oversaturated and far from photorealism. May be there are another settings for this model (Gonzalomo XL V3)? And how to fix blurred background like on tele lenses?
@40inD No, those are the right settings and the ones that I use most of the time. Are you using Comfy or A1111 or something else?
@GBRX I use Comfy
@40inD I'm not sure what the issue could be for you. I don't see any images you've posted so I can't look at the workflow.
V2 Unity as a refiner of ZaxiousXL is ...wow....a blast
Great, thanks for the feedback! V3 and V4 should work just as well, if not better, if you want to give those a try.
Hands down the best SDXL model so far. Thank you <3
I love the way V4 Non DMD understands the complicated or vague prompts. But in Forge it never make a perfect image in one shot as the DMD versions do many times. I'm not a ComfyUi guy, I hate it. A few of my latest images were started with the V4 Non DMD model, but my way of tweaking the images in the end in Forge is with inpainting and I find V3 or V4 DMD do a better job at that. So it's not because I don't post images finished with V4 non DMD that I don't use it. Thank you for that model. The only models I don't wanna use for pictures with people are all the Flux models.
Hey, it's not for me to judge how you use the models. I'm just happy you're finding such a good use for them! That said, I'm a Comfy guy and never used Forge. Does Forge allow you to do img2img by model switching? ie pass the first image generation to another sampler with a different model. If so, then you might want to give Flux a try creating the base image and then pass it off to the DMD model to add all the realistic detail. That's how I normally use Flux because it creates images with great color, composition and backgrounds but the people detail is so plastic. SDXL fixes that.
@GBRX Sure Flux in Forge can do img2img and then you can use an SDXL model on the resulting images. I agree that Flux can make good compositions and colors. However if I give Flux a prompt of a XXX scene, it will be unable to understand the scene. It renders 2 people kissing as if there is censorship or that it wasn't trained to do that XXX stuff.
@allxander99339 Very true, Flux is horrible at NSFW except for very basic nudity.
hey! what do you use for your deepfake ? instant id? pulid?
Awesome, thanks!
The best for NSFW and really fast! Thanks
Thanks!
Simply unmatched. The best SDXL model by miles.
Still learning, but trying to understand the acronyms here. Is AIO "All in one"? What's the S in Flux S, is it just a variant name? What does DMD stand for, Distribution Matching Distillation?
Sorry, AIO does mean All In One - referring to the fact that it contains the CLIP and VAE in addition to the UNET. Flux S means Flux Schnell. I don't really know what DMD stands for but it's the integrated LoRA that makes it much faster.
GBRX Ah gotcha, pretty convenient having it all in one. Well if it does mean distribution matching distillation, that's when they distill the models into smaller ones which makes steps go from like 20-50 to 1-5, so that would make sense why it's much faster!
goongoo You're right, it does stand for Distribution Matching Distillation. Interesting reading about it in the repo, https://github.com/tianweiy/DMD2
Noob question, why "failed to load model, restoring previous"? I put this on the Stable Diffusion model folder along with other models, I can't load it tho, am I doing something wrong? I use auto1111. Thanks.
Sorry, I wish I could help but I'm a noob when it comes to A1111. I'm not sure what the problem could be but I do know that people are successfully using the model with A1111.
GBRX Are people just putting it on the models folder like I did? Maybe I don't have enough memory? Thanks anyway, this model looks great, anatomy-wise it seems like the best one yet in here.
AIsupermodelsDev2975 Sorry again, but I don't know how they're using it. I would assume it can be used just like any other checkpoint. It has the same memory requirements as other SDXL models so if you can load those then mine should work just the same. I hope you're able to get it working.
FLUX is not supported by A1111 yet (See: https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/16314), as the last update to the webui is now almost a year ago. Try SwarmUI, ForgeUI or, if you are comfortable with nodes and like sorting spaghetti, ComfyUI.
Blupp +1 on the ComfyUI recommendation. It's so much easier to use than it was a year or two ago and it's really not hard to understand at all if you begin with a basic workflow and build from there.