Chroma / zImageBase - Cinematic Atmosphere

zImageBase - 1.0

Mehh, most likely have to change all capitations for zImage to be more precise. Also, maybe lower strength for now. Seems to ping-pong a lot while training from hardcore hallucinations, to very clean, to actually what i want but undercooked and so forth.

V2.0

Added like a 100 more images, tried to balance it with various themes, lighting, composition and perspectives. Since i wanted it to be more cinematic, it will produce better images in a more movie like aspect ratio ( depends )

Also, because it's a factor that changes the result by a lot, i'm using a modified T5-Encoder, which unless other trained text encoders that claim to be less restrictive and do pretty much nothing, this one actually changes an image quite substantially ( same seed and settings ). A bit of a double edged sword, because it seems to hit not as many roadbumbs while generating as the normal T5 does. Follows positive and negative prompt a lot closer, which makes it quite a bit more flexible, but requires more careful prompting to not introduce some weird artifacts. Just saying because you might not get the same result.

Another thing, at least to my understanding, which is also the case for other peoples Loras. If not trained on a set that has a certain theme, artstyle, class token etc, a Lora has no effect on the outcome if you don't hit at least a word in your prompt that invokes the Lora to have an effect. Since i have no class token in most of my Loras you may need to use things like "cinematic" "movie still" and so on somewhere in the prompt or the prompt itself should "kinda" sound like it leans toward a more movie like setting. Models like SDXL wouldn't care so much about that since most of the time the text encoder get's trained alongside the images, but with Flux based things and just training UNET, tokens getting embedded on a very surface level.

Still some things to fix and add here and there.

Some of the training images where part of the professional photo lora, but i wanted them seperate. It's not there yet and to much playing with CFG and steps involved. Didn't plan to release it at the current state, but it's a base i can build up upon. Might use it here and there to see where it's lacking, which are quite a few things.

Depending on prompt and settings has tendencies to produce low resolution and blurry stuff.

90% of the images aren't from movies, just images i thought that have a cinematic "feel" to them, like perspective and composition. That also includes a few illustrations and paintings which fit the bill imho.