Check my exclusive models on Mage: ParagonXL / NovaXL / NovaXL Lightning / NovaXL V2 / NovaXL Pony / NovaXL Pony Lightning / RealDreamXL / RealDreamXL Lightning
If you are using Hires.Fix with V5 Lightning, then use my recommended settings for Hires.Fix (3 Sampling Steps, Denoising strength: 0.5 and CFG Scale 1.0 - 2.0) or other settings you find better for you.
Use Turbo models with DPM++ SDE Karras sampler, 4-10 steps and CFG Scale 1-2.5
Use Lightning models with DPM++ SDE Karras / DPM++ SDE sampler, 4-6 steps and CFG Scale 1-2
Please pay attention to the model file name, the part of the name after the underscore is the true version of the model.
The model is already available on Mage.Space (main sponsor)
You can also support me directly on Boosty.
RealVisXL Hugging Face Full Collection
The model is aimed at photorealism. Can produce sfw and nsfw images of decent quality.
ᅠ
Recommended Negative Prompt:
(worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth
or another negative prompt
ᅠ
Recommended Generation Parameters:
Sampling Method: DPM++ SDE Karras (30+ Sampling Steps) or DPM++ 2M Karras (50+ Sampling Steps)
ᅠ
Hires Fix Parameters:
Upscaler: 4x-NMKD-Superscale-SP_178000_G / 4x-UltraSharp upscaler / or another
Denoising strength: 0.1-0.3
Upscale by: 1.1-1.5
ᅠ
Optional Parameters:
ENSD: 31337
ᅠ
This model is:
ᅠ
Huge thanks to the creators of these great models that were used in the merge.
SDVN6-RealXL by StableDiffusionVN
TalmendoXL - SDXL Uncensored Full Model by talmendo
ProtoVision XL and DynaVision XL by socalguitarist
Cinemax Alpha | SDXL | Cinema | Filmic | NSFW by viakole
Description
Use with DPM++ SDE Karras sampler, 4-10 steps and CFG Scale 1-2.5
FAQ
Comments (26)
UPD: My nerves were destroyed by CivitAI because you can't add example images after uploading a model (or I'm stupid and didn't figure it out). V2.01 Turbo will be uploaded again soon.
A fix for all Turbo versions will be released soon:
- fixed problems with normal SDXL resolution that could produce stretches and mutations
15-30 steps with turbo or is lower steps possible with it?
Hi, this model can generate images on the DPM++ SDE Karras sampler at 4+ steps.
In my very limited testing of 7, 10, 12, 20, 30, 40, and 50 steps, I found that anything over 12 didn't show a noticeable improvement. You could always produce a series of them at a lower step count and then experiment with your favorite(s).
If you click on the 2.0 download button, 2.0 Turbo is downloaded incorrectly.
it might just be incorrect naming but the correct model. Also experienced this with Yamer's Unstable Diffusers. As soon as the page title added the word turbo, all the filenames inherited the word too, even if they weren't turbo models.
Is there a way to use this model for inpainting?
Hi! I think some basic operations like face generation on distant plans in good quality are possible on the regular version of the model. As soon as it will be possible to make inpainting model I will do it.
can you explain somewhere the workflow with RealVisXL V2.0 + Turbo to obtain these images? do you work from the beginning with the final resolution of the image? do you start with a low resolution image and then upscale it with the same model? please, someone give these details!
Hi, first I generate an image without Hires.Fix, if I like it I use Hires.Fix to increase the resolution and details of the image.
Recommended Hires Fix Parameters:
Upscaler: 4x-UltraSharp upscaler / or another
Denoising strength: 0.5
Upscale by: 1.1-2.0
Now I usually use DPM++ SDE Karras sampler with 15 steps, for Hires.Fix I use 10 steps.
I see people using higher steps with this turbo model vs others. So you recommend 15 or more??
@SG_161222 Thank you for answering and for your magnificent and generous work.
UPD: links have been updated!
Hi everyone! I made a small update for V3.0 and have already uploaded it on Google Drive MEGA (links below). This update should add more detail and improve the skin texture a bit (possibly improved hands).
RealVisXL V3.0 ---> RealVisXL V3.0 U1
To anyone who decides to help me test this model I will be very grateful. Your feedback will help me with further improvement of this model.
noticed it accepts a high CFG. good results at 15-17. really pops the colour and skin detail
@lytan Thank you! I think V3.0 U1 is already well ahead of V2.0 at some points.
your V2 was easily my favourite model. best skin textures :) thanks for that! high hopes for V3
Subjects are ok, but some photos are oversaturated a bit, I think there is a little too much light and a little too much contrast between shadows and lights, drawing of the crowd is improved even with few steps. If you put the same seed and the same prompt in v1 and v2 checkpoint, the scenario is similar, but with v3 checkpoint sometimes it is different
I will definitely test the Turbo version of V3.
@OzzyOsman Thank you! Currently testing a non-Turbo version :)
@OzzyOsman Or use this Turbo lora, 6-8 steps: https://civitai.com/models/216190?modelVersionId=247778
V3.0 U1 is much better than 3.0 U0. V2.0 had great quality detail and texture, but was often too dark. V3.0 U0 fixed the brightness issue, but overall a step backwards in quality. V3.0 U1 is brighter (MAYBE too bright) and retains the quality and detail from V2.0. For hand artifacts, I don't really see any difference. I see hand artifacts with any SDXL checkpoint I try.
@SG_161222 you may want to merge your base SDXL checkpoint with the DPO checkpoint. All I gotta say is wow! I merged your models (regular and Turbo) with DPO and it's gangbusters.
just made a quick comparison with some community posts avoiding any LORA in the generations for clean comparison. Some hires options are changed because of my GPU but it mostly remains the same.
https://imgur.com/a/0CKD7DU
Does anyone manage to use any Controlnet option with this XL + Turbo model to have some control of the pose? I have tried with several models (canny, depth...) and they all destroy the image quality even though I use very low weight values in Controlnet.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.