Like the work I do and want to say thanks? Buy me a coffee or Support me on Patreon for exclusive early access to my models and more!
Meet ProtoVision XL! This model is my "Bob Ross" model, a happy accident that happened when I goofed up some merge settings on my DynaVision XL model and stumbled upon some magic.
So what is ProtoVisionXL? At its core it's a custom built version of NightVision XL, which I then pile multiple fabulous LORAs on top of to create a unique look.
Like all of my other models, merges, embeds and tools, ProtoVisionXL is designed to be easy to use and excels with simple prompts, but is very coherent and can build fabulous scenes with exquisite detail.
ProtoVisionXL is a portrait model, it favors showing human subjects. That's not to say you can't get other art styles, creatures, landscapes and objects out of it, as it's still SDXL at its core and is very capable.
WARNING - DO NOT USE SDXL REFINER WITH PROTOVISION XL
The SDXL refiner is incompatible and you will have reduced quality output if you try to use the base model refiner with ProtoVision XL.
Follow me here by clicking the heart ❤️ and liking the model 👍, and you will be notified of any future versions I release. I also need your help with feedback, please please please post your images and your honest feedback below, I will use your feedback and your output to help guide future revisions!
Patreon coming soon
LORAs used for the initial Alpha version of ProtoVisionXL
add-detail-x
AdvancedEnhancerXL
DetailedEyes_xl_V2
MovieStillXLV7
Neonify
offset_example-lora_1.0
photorealistic_slider_v1-0
Samaritan 3d Cartoon SDXL
dwzstyle-000018
wowifierXL
Description
Release Notes v0.6.3.0
With version 6.3.0 I have trained and merged in several detailers that increases overall fidelity, image quality, scene reproduction and gives scenes a more realistic "lived in" feel. The detailer has not changed the flexibility of ProtoVision, it still very much sits somewhere in the middle between DynaVision and it's cartoony aesthetics and NightVision with its focus on realism and photography.
A word of note on sizzle - this model operates best in the 3.5 to 5 CFG range. While you can generate images at 7.5 CFG (the defaults for both A1111 and Comfy for some dumb reason), they're gonna look fried and overbaked and sizzled 🥓 with crushed colors and blown out whites. Lower that CFG down to 5 if you want your colors to pop, or to 3.5 if you want nice neutral color temps that you can toy with to your liking in post.
Changelog 10/27/23
merged 2 new detailer LoRAs
SCG-Detailer-v1
SCG-Detailer-v2
Known Issues
Hands
Male Genitalia
eyes and facial features at a distance can look bad (that's an SDXL thing, hoping to fix this over time with more trainings)
FAQ
Comments (17)
Release Notes v0.6.3.0
With version 6.3.0 I have trained and merged in several detailers that increases overall fidelity, image quality, scene reproduction and gives scenes a more realistic "lived in" feel. The detailer has not changed the flexibility of ProtoVision, it still very much sits somewhere in the middle between DynaVision and it's cartoony aesthetics and NightVision with its focus on realism and photography.
A word of note on sizzle - this model operates best in the 3.5 to 5 CFG range. While you can generate images at 7.5 CFG (the defaults for both A1111 and Comfy for some dumb reason), they're gonna look fried and overbaked and sizzled 🥓 with crushed colors and blown out whites. Lower that CFG down to 5 if you want your colors to pop, or to 3.5 if you want nice neutral color temps that you can toy with to your liking in post.
Changelog 10/27/23
merged 2 new detailer LoRAs
SCG-Detailer-v1
SCG-Detailer-v2
Known Issues
Hands
Male Genitalia
eyes and facial features at a distance can look bad (that's an SDXL thing, hoping to fix this over time with more trainings)
For what it's worth, I think 6.2 is better than 6.3. After testing dozens of samples side by side, 6.2 seems to me to be almost always better than 6.3: better lighting, color, detail, natural, prompt following, etc. That is taking into account lowering the CFG to 5 and 3.5 for 6.3 too. Version 6.3 does a good job, but I like the overall look of 6.2 better.
I am currently running a private instance of Auto1111 (not commercial or an AI gen service) on a service runpod.io. I am trying to wget through the linux CLI, however there appears to be some form of block that results in a tiny html file being downloaded instead. I have tried numerous methods including using the curl command without being able to onboard the 6.5 gb model file onto my private stash. I had not run into this previously, with other models on CIVITAI so I thought I'd drop a little comment and see if there is another link that would accept the wget/curl commands? The reason I would like to do it this way is that for my local comp (nearly a decade old lolz) my download speed is quite slow, and then having to then turn around and upload the file onto my private storage unit is quite time consuming. Thoughts? Solutions? This is NOT intended to be a bad review or angry post. I am happy and content, just a curious soul who would like to probe this issue, even if the answer is that I must download then reupload.
You could always try putting it on a google drive share, then using that link into the run pod? Not sure why it's failing, nothing I can really do about it, sorry!
@socalguitarist Not sure if this is related but your model requires signin to download unlike other models. This will cause a download failure with tools that use the api that aren't signed in. You check by signing out and then going to the model link: https://civitai.com/api/download/models/201514
@K18V Bingo! Thanks for the reply. Now that you bring that up I do remember seeing that within the html file. Just out of curiosity, and a desire to learn, how did you arrive at that conclusion yourself?
@3VOLUTION I saw a similar issue at https://github.com/BlafKing/sd-civitai-browser-plus/issues/86 This extension shows the model download link and after clicking on it when I wasn't signed in I realized this was likely the issue.
try. lynx. which is a text web browser and can handle redirects, etc. better then curl
Noticed the same thing myself with this, while also using Runpod, and came to the conclusion that the Author now wants you to login to download the model. So Curl/Wget won't work properly.
This is the first model I write a comment for – because I'm frankly blown away! It beats any other model in terns of resolution and detail. Cant wait for the next release!
I can't download this and a couple others using wget onto a runpod instance anymore. It's claiming I need to be logged into CivitAI, which would require me to install something like Lynx. Why is that? I used to love this model, but if I can't directly pull it to runpod from CivitAI, I can't really use it... downloading it, then uploading it somewhere else I can easily get to is about the only option, and that's a big pain, so I've just stopped using it.
Great model. Too bad I can't access it anymore. :(
It's taking a month of Sundays to d/load to my PC in London UK - others of comparable size took 45 minutes ... it now says 2 Days!!!
There is a temp workaround for this. You need to first start a download in your browser then go to your "full download history" page (in case of google chrome) and right click on the download to copy it's link, it will be a really long link containing words like "civitai delivery worker prod". When passing wget command on runpod paste this link and replace the civit ai link. This will download the model on runpod. But you might have to replace this link next day because it expires after some time.
where can find the lora of add-detail-x,can you give the link？
It is built in - or you can add the SDXL version if you want more control over it.
Draw Things is asking me all sorts of questions during the import process:
- I assume the fine tuned image size is 1024x1024?
- Should I use a custom Variational Auto Encoder?
- Does Protovision use V-Prediction?
- Higher Precision?
What happens if I use the wrong settings with these?
1. 1024 x 1024
2. VAE is built in, no need for custom
3. no, that's a SD2.X setting
4. no, shouldn't need higher precision, it's an fp16 model.
I've asked Liu Liu to add my models to Draw Things a couple times now. Feel free to bug him on the Draw Things discord to get Proto (and my other models!) to Draw Things, because it's a pain to add them manually.