The training data and captions are included.
I advise a negative stronger than this, as it still gets artifacts. Let me know what you come up with.
vhs filter, blurry, pixelated, lowres, bad quality, disgusting, upscaled, ai generated, fisheye lens, depth of field, blur, pixelated, voxel, 3d,Combining the two models at around 0.6 strength seems to provide some good outcomes. Both at 1.0 works too. Either at 2x kinda messes the generations up.
Generates and improves a large array of liminal depictions in a kind of shortcut fashion as to what is already present in Chroma.
The dataset used is a curated set of 950 images, many of which were considerably larger than 1024x1024; some upward of 4096x4096. The size and forced downscale is likely one of the reasons the model is confused.
The images all have plain English captioning stemmed from the starting phrase "liminal," which I'm uncertain if Chroma likes or not. More than likely it didn't like this.
Chroma is already pretty good at liminal, so this is a shot at training a lora using the Civitai service, and the Lora itself didn't turn out THE BEST but it's pretty good.