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!
NightVisionXL started as a lightly trained, photography focused SDXL model that specialized in photographic output. Over the past few months, NightVision has received dozens of trainings and tunings chasing the dragon that is pleasing aesthetic output. With the release of version 8.X, I feel like I'm much closer to my goal of effortless beautiful output that is aesthetically pleasing without requiring confusing negatives or comma separated nonsense.
NightVisionXL speaks natural language. Prompt it like you would describe a scene to a human being, and you'll be surprised by how well it can understand and comprehend your prompts.
I'm trying something new with V8 of NightVision, and I hope the results show in the output. I continue to scrape beautiful imagery that I think will train well, and this particular update is no exception with hundreds of beautiful license free AI images that I've sourced from MJ, Dalle3, Civit, Playground and my own outputs. On top of that, I'm also utilizing HF datasets for license free photography and the LAION-POP dataset, filtered to remove low quality images prior to training. Captioning is done using my own Spicy Burrito captioning tool (available to my Patreon subscribers on discord!) that creates incredibly verbose and effective captions that are essential for good coherent output.
On a final note, I've had multiple requests to "turbo-ize" this model. I have no plans to mix turbo into NightVision, as the licensing on NightVision is nice and open right now with the SDXL licensing, and I'd like to keep it that way. NightVision isn't fast, but I promise you the longer gen times are worth the wait!
This is NightVision XL, a lightly trained base SDXL model that is then further refined with community LORAs to get it to where it is now. NightVision XL has been refined and biased to produce touched-up photorealistic portrait output that is ready-stylized for Social media posting! NightVision XL has nice coherency and is avoiding some of the weird body issues and biases that are starting to plague some of the other photorealistic models. Further, NightVision XL produces rich deep blacks and great evening/night time scenes. It can also produce ridiculously bright output as well!
NightVision XL is capable of both SFW and NSFW output. As with all of my other models, tools and embeddings, NightVision XL is easy to use, preferring simple prompts and letting the model do the heavy lifting for scene building.
NOTE - This version includes a baked VAE, no need to download or use the "suggested" external VAE.
WARNING - DO NOT USE SDXL REFINER WITH NIGHTVISION XL
The SDXL refiner is incompatible and you will have reduced quality output if you try to use the base model refiner with NightVision 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!
Description
NightVisionXL V9.0.0 Release Notes
This is NightVisionXL 9.0 - This model is a fine tune of NightVision v8.1.1 with approx. 10,000 images comprising a rough mixture of 40% from a photography dataset, 40% from Laion Pop (filtered and recaptioned with GPTV), 10% from a cinematic dataset (same one used to train CineVisionXL), and 10% synthetic mixture of favorites picked from Civit and MJ, all hand chosen and captioned with GPTV. All input images were of high quality and large size, bucketed on training with a wide variety of sizes and aspect ratios, leading to a model that is very good at producing coherent output even in very extreme wide or tall aspect ratios, and anywhere in between. As with 8.1.1, NightVision v9.0 is a general purpose model that can output pretty much any style you like. NV continues to prefer natural language prompting over 1.5 style (though that does still work), however it has no Pony training or mixing, so those pony specific prompt rules won't work here (sorry).
I have not mixed in any speedups or improvements, nor any other alternate license models or LoRAs. NightVisionXL remains a pure SDXL model, and will stay that way moving forward.
Rather than my typical changelog (see above), I want to touch briefly on the release of SD3 Medium, the 2B model version. As a member of this community, I am very upset at the state of the model that was released. The hamfisted censorship of the model was uncalled for and effectively broke the model making training it almost impossible (from a technical standpoint it's just impossible to apply any type of decent learning rate without the loss skyrocketing and the train coming out completely terrible if not entirely broken). I know I've disappeared from the scene for awhile, I like many other trainers decided to "wait for SD3" - Now I regret that decision. This version of NightVision, this was my "private stock" version I trained for myself and my close friends. V8.1.1 was the swan song, v9 was my special gift to my friends, and it was those same friends who convinced me that v9 should be released, hopefully this can help bandage the wound that SD3 inflicted on the community. V9 isn't perfect, it has the same flaws that all SDXL models do (mostly related to the SDXL VAE, that I can't fix), but it's the best I've made, and now it's yours.
We'll see what happens with SD3. I'm going to say right now, unless there's a major training breakthrough with the 2B model, I'm likely just going to skip it and hold out for the 8B model (if they don't mangle that) or maybe look into training a different architecture, just not sure which one yet.
Known Issues
male genitalia is better, but still wierd. use a lora if you need perfect penises
hands are probably 85/90%, depending on the pose, but some things like holding objects can fight you.
Upside down or sideways faces (this is an SDXL VAE issue)
faces and detail at medium distance can be mushy in initial generations (another VAE issue)
FAQ
Comments (28)
Hi Folks! I snipped the below from the patch notes (see the box on the right if you want those).
I want to touch briefly on the release of SD3 Medium, the 2B model version. As a long time member of this community, I am very upset at the state of the model that was released. The hamfisted censorship of the model was uncalled for and effectively broke the model making training it almost impossible (from a technical standpoint it's just impossible to apply any type of decent learning rate without the loss skyrocketing and the train coming out completely terrible if not entirely broken). I know I've disappeared from the scene for awhile, I like many other trainers decided to "wait for SD3" - Now I regret that decision. This version of NightVision, this was my "private stock" version I trained for myself and my close friends. V8.1.1 was the swan song, v9 was my special gift to my friends, and it was those same friends who convinced me that v9 should be released, hopefully this can help bandage the wound that SD3 inflicted on the community. V9 isn't perfect, it has the same flaws that all SDXL models do (mostly related to the SDXL VAE, that I can't fix), but it's the best I've made, and now it's yours.
We'll see what happens with SD3. I'm going to say right now, unless there's a major training breakthrough with the 2B model, I'm likely just going to skip it and hold out for the 8B model (if they don't mangle that) or maybe look into training a different architecture, just not sure which one yet.
Love ya'll, here's to many more great breakthroughs and tunes! ❤️
Nathan aka SoCalGuitarist
thanks , i hope you work on human body pose on your next sdxl model , trust me the whole sdxl is good at standing human body and simple poses , but for real the reality not like only those simple poses , try push a model 'sdxl' for realistic human body shapes poses , that already no model can do..drop much much more of human pictures of not regular poses , not all human pictures is lying or standing , trust me again this is a big weakness point in whole models, despite struggle doing any other than 'one object' picture, all models can't do more than one object ..keep in mind while you train one more great model , thanks again!
I'm with you on this one - 25+ year software dev pro - there may very well be improves in underlying architecture, but output is going very backwards.
8B will probably be subscription based only access and will require you to be connected to their servers to use. I don't have faith in Stability AI any longer. I'm just waiting for them to issue a cease and desist on all distribution of their model. They'll pull a Nintendo soon and start suing individuals who are using their models for any purpose if it didn't come directly from them.
thank you for writing this. makes me feel less gaslit by all of these new flagship models taking even more effort to produce outputs that are both unexpected and ugly. Not sure whether it's related to safety or because they're having to remove copyrighted material from their datasets in order to commercialize, but it's not working.
Love this series/model. You're clearly very talented and dedicated. Much appreciated!
I feel like Nightvision is now the best general model out there. Thank you for all of your work and for continuing to iterate on the fundamentals.
One question, what are the recommendations for using this model?
- VAE
- CFG
- STEPS
- CLIP SKIP
Thanks !
- VAE: madebyollin_sdxl_vae_fp16_fix.safetensors is baked in
- CFG: I run 2.5 with PAG. If you're not using PAG, then 3 - 5 range is good.
- STEPS: depends on the sampler. been using DPM++ 2M SDE SGMUniform and DPM++ 2M SDE AYS a lot lately, and run both at 30 steps
- CLIP SKIP: Nope
@socalguitarist PAG ?
@socalguitarist Hello, noob here. What is PAG? I've only heard of PAWG. 🤭
Comfy user here. These are the settings that seem to work best for Comfy specifically:
- VAE: standard SDXL or whatever's baked in
- CFG: 5-10 (at least 6 if your prompts consist mostly of sentences)
- STEPS: 25-45
- CLIP SKIP: -2
@Mister_Kaos , Thank you very much.
@socalguitarist ,One last question, does this model work with controlnet? I have some concepts that require the use of controlnet, such as lineart and depth.
@socalguitarist , hi friend, are you ok? I got the settings I needed thanks to your model, look at the result:
https://civitai.com/images/20546350
Great model. I really like it.
This is a great model.
nightvisionXLPhotorealisticPortrait_v0791Bakedvae the best of all your models, beautiful and aesthetically pleasing, develop new models based on it.
不知道问什么，模型一直加载不进去，有人遇到同样的问题吗？
怎么会事呢
Can the images generated with this model be used commercially?
Yes, this model permits users to:
Sell images they generate
@olgav86569 This is great. Most people that ban commercial use also ban any use with the true Open Sources licences, like Apache. Wich is sad, honestly.
Does not work with Gourieff Reactor....and many upscalers....
You're like an older brother you always have to learn from.
]Ihate to say this, i love the images, but lookint at them and reading the prompts it doesnt seem to be much prompt adherence, half of the description in the prompt is not present in the images. I think you are too ptimistic trying to use natural language in a way the original sdxl base was not trained to. In my experience, poetic or literary style natural language does not work that great with prompts or especially with sd models. A lot of times i decontruct the prompts by removing words piece by piece without changing the result and most times i can remove a lot of words without affecting the result much which means they are not really having a guiding influence. But anyway that is my feedback, images are great, thought, i love the rcrushed blacks together with the vibrant colors, just I wish it was more controllable as far as prompt adherence. Any effort is commendable though.
By the way i meant the demo images, but looking at other user created images i see that itt does behave as expected with tag based prompts! and I do love the results.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.