CineVision 2.3.0 Public Release Notes

Hello folks, this one has been in the barrel for about a month and a half now, with multiple test revisions and multiple layers of training. All told, CineVision received about 120k total training steps over close to 9k training images. CineVision is a movie-focused model that will default to cinematic realism styles as it's normal output. It is very responsive to many different types of cinematic styles, terminology and scene setup. It is very coherent and responsive to your prompt, and works best with natural language prompting like in my example images.

CineVision, despite its name, is actually quite dynamic as a model - It can create photographs, cartoons, movies, weird scribbles, coloring books - you name it, it will probably work.

With SD3 drop coming right around the corner, this will probably be my last big SDXL update for this model (and my other models, which have "final XL versions" coming soon). I hope you enjoy it, it's my daily driver and remains so even as other new models come out. I'll be dropping a lightning version as well, and it also works great with the new Hyper loras too (several of the samples I've added it, though I think Hyper is a bit too hot on CFG)

Hands and medium distance objects are quite good now, especially if you use SAG or PAG (or both).

Changelog 4/25/24

5 trainings for about 120k total steps on several thousand training images.

3000 cinema stills handpicked and upscaled with ultrasharp to > 1024 and < 2048 on it's longest side

2000 photos from a photography dataset

2000 images from LAiON-POP dataset

1600 MJ aesthetic art dataset (my own library, hand selected)

350 anatomy training images

All captions done with GPTV/LLaVA-1.6

Known Issues

Men's genitalia has come a long way (no pun intended) but still struggles with generating convincing penises. Best to use a lora and inpainting if you're really wanting accurate male physiology.

smudgy effects with some movie styles - this can show up sometimes due to training on 24FPS content early on. It's mostly been quashed now, but can still show up if you're really pushing the old fashioned analog movie look (think 50's style technicolor)

Note - All of my sample images for this version are using Perturbed Attention Guidance (PAG), FreeU, self attention guidance, latent modifier, DynamicThresholding. All but PAG are built into ForgeUI, which I heartily recommend over A1111.