If you like what I do, leave me a ❤️ to receive notifications of future updates and follow me here on Civitai. Remember that it is a hobby for me and it costs a lot of time and everything is shared entirely for free, I appreciate that you make a review and share your creations on this wall.
Introduction:
R.O.A.R. is a very special checkpoint for me. In its version for SD 1.5 I really enjoyed mixing models and LoRas to give it shape, and in its version for SDXL 1.0 I wanted to do something different. My first idea was to train it with high resolution images taken directly from the R.O.A.R. original, but... I wanted to experiment and follow a process as close as possible to R.O.A.R. from SD 1.5. So I trained several LoRas specifically to mix them with Stable Diffusion XL 1.0 and the result was... chaos. Until... after many attempts and reviewing all the parameters, R.O.A.R. XL v0.1 was born. . The following weeks I dedicated myself exclusively to improving it by adding content and making it more versatile. The result is this... I hope you like it.
R.O.A.R. (Realm Of Awesome Realism) XL v2.2
Recommended Settings:
Width/Height: 1024x1408
Sampler: DPM++ 2M SDE Karras
Steps: 20-30
CFG: 4-7
VAE: baked
Negative Prompt: No negative prompt needed, add afterwards the things you don´t want to see in the image.
HiRes: 4x_NMKD-Siax_200k with half the regular steps and 0.15 Denoise + 1.5 Upscale
R.O.A.R. (Realm Of Awesome Realism) XL v1.3 + CFG Fix
Several Bugs Fixed, CFG Bug fixed which now allows more flexibility and more different samplers to choose from. Changed VAE that improves image consistency and prompt fidelity. Better Hands and better Faces.
Recommended Settings:
Width/Height: 1024x1408
Sampler: DPM++ 3M SDE Karras
Steps: 30 - 40
CFG: 3-12
VAE: New VAE, Already baked in.
Negative Prompt: No negative prompt needed, add afterwards the things you don´t want to see in the image.
HiRes: 4x_NMKD-Siax_200k with half the regular steps and 0.15 Denoise + 1.5 Upscale
R.O.A.R. (Realm Of Awesome Realism) XL v1.2
This first version focuses exclusively on quality and versatility. It contains some errors, but in general, it returns very good results. It is designed to work with high resolutions, although it also works well at low resolutions.
Recommended Settings:
Width/Height: 1024x1408
Sampler: DPM++ 2M SDE Karras
Steps: 30 - 40
CFG: 3-12 (Works great with other samplers, but at low CFG)
VAE: Already baked in.
Negative Prompt: No negative prompt needed, add afterwards the things you don´t want to see in the image.
HiRes: 4x_NMKD-Siax_200k with half the regular steps and 0.15 Denoise + 1.5 Upscale
Adetailer is advisable in images with many faces and hands.
I hope you enjoy this release and give it all your love. Share your creations in this wall and help me improve the checkpoint. See you in the next update
Description
V1.2
FAQ
Comments (17)
not a real XL model. strong block-shaped artifacts
Maybe if you read the instructions, you wouldn't have those results. In any case, I am at your disposal to answer your questions.
In most cases the cfg stale is set to high. This could also be influenced by used loras.
@paskmoe Indeed, that is a bug that I am aware of and that using any sampler outside of DPM++ 2M SDE Karras or DPM++ 2M SDE, forces the use of low CFGs. I'll probably release a fix soon or fix it directly in the next update. Thanks for highlighting it, I will continue working on it.
@DeViLDoNia Thanks, appreciate the work you put into this :)
@DeViLDoNia What qualifies as 'low CFG'? 4 and under? I stay within the 3-5 range for CFG but I also stick to Restart and DDPM as samplers. Plan on checking out your checkpoint soon ^_^
@DeViLDoNia is that a bug? I'm experiencing lately that lowering the CFG to between 3 and 6 REALLY helps with the coherency of images, at least as far as LoRAs are concernec
@Up_punchington and general quality / photorealism. My CFG is currently at 3.75 and it works well with the majority of checkpoints that emphasize photoreal. My range is 3-5. I imagine it will be higher when I switch to stylized/toony stuff.
@lucidzachary473 That's what I noticed. Using some LoRAs that are trained on real photos but with Pony, everything looked terrible until I turned the CFG down and now it's crazy good quality
@Up_punchington If you have to lower the cfg a lot with a LoRa, it is because it is overtrained. As a general rule, a LoRa should work more or less well between ranges 3-10, so it also depends a lot on the checkpoint. Likewise, a well-trained checkpoint should work more or less well between those same ranges. You can also get a LoRa or a Checkpoint to work better with low or high CFG (LCMs, for example, are designed to work better at ultra-low CFG).
But returning to the case of version 1.2 of R.O.A.R. XL, for some unknown reason, even sometimes setting a low CFG (for example at 4), in some graphs it seems to continue working at high CFG values. Searching the forums I found a way to solve it, and that didn't happen to me personally, but thanks to this incredible community, I found out about the problem.
I tried your prompts on another model for comparison and got almost the same images.)
If there's anything special about your model, your prompts don't show it...
ditto. its very sharp. But also very the same.
@WTFusion All the prompts I use for my sample images are generic AI-generated prompts. As for what R.O.A.R. It aims to contribute something new or special, it does not have so many pretensions. Image quality and sharpness without having to resort to third-party extensions. For now it is what it is, you may like it or not, that is no longer in my power. It will improve in future updates, like all my checkpoints, I do this as a hobby and because I like it, I don't make a living from this and that's why I share it for free.
I appreciate your opinion anyway and I hope you will test my work again in future updates.
@DeViLDoNia I was more disappointed in one of my models than in yours. You’re right about the Juggernaut; it spread like a plague. That same freckled face is everywhere, making me sick
based on my findings, this is not trained, it is very very sharp however. I use the same words to test models and the results are identical to 70% of the models here. But there Is a lot of detail. Wish there was more creativity though.
First of all: thank you for your opinions and for having tested this checkpoint.
A couple of notes that I would like to contribute:
This checkpoint is 100% based on own LoRas merges with Stable Diffusion XL v1.0, it has absolutely nothing else. About the LoRa that has been injected, two of them are only image quality and sharpness improvements and are published in Civitai. The point is: All or most of my XL LoRas were trained in Tensor.art, and the only option they give you if you want to train something realistic is to use the Juggernaut XL checkpoint, it doesn't give you the option to use Stable Diffusion XL 1.0 for example. The vast majority of Civitai checkpoints carry some Juggernaut XL, this is true and we all know it, either from using it to train LoRas for merges, etc...
It is inevitable that there will be no similarities, nor will the results of simple prompts return the same images in the first version of a checkpoint, unless you have trained it completely regardless of all these factors. So, in later versions and with their relevant updates, the checkpoint will move further and further away and it will return different results, the challenge is to maintain the quality while distinguishing itself from the others. But it's a long road.
Once again, I thank you very much for your opinion and I hope you continue to help me with your comments to improve this checkpoint.