ModelsLab/NightVisionXL - NightVisionXL V9.0.0

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)