zImage Base - v1.0
To much to explain which is most likely just wild theories anyway. Was a battlefield of several different training settings with a few complete restarts, still don't really know when it decides to work or not, sometime it does, sometimes it doesn't. Also save to go over 1.0 strength, 1.2 in some cases fixed hallucinations and bad anatomy ... as far as you can fix bad anatomy because that is pretty much the whole idea of the Lora.
Training mixed media in zImage with a bigger dataset is ... tricky. A focused and smaller set, less problems.
Some of the prompts are missing what the image actually shows by a lot ( still old captations ). Running the images through the LLM again with better descriptions seems to fix it for the most part, even after training, but there a quite a few with completely wrong capitations.
Not really to sad about it doing its own thing sometimes, means more variation i didn't expect. A bit to tame, but that's just zImage i guess, needs a nice little finetune, small push in the right direction, cough chroma cough
V3.1
Cleaning house, leftover images trained.
v2.0
More images added ( more environments and atmospheric stuff) , much, much more training, merging and a simple change in training strategy.
Anything movie related will still produce blurry b-movie like quality, so lower strength, but the overall quality enhanced a lot. Given the resolution of a few training images that was kinda unexpected. Sometimes the change is subtle but more than enough i think. In some cases, it can turn simple tags into a good looking image, even illustrations.
If the same works with other loras, i may have reached my personal peak... for now... with a lora that was just made for shenanigans sake.
After, i think, 40.000 steps i call it a day for now, till i collect more pictures of fitting environments and actually more unsettling atmospheric stuff instead of just monsters.
The pictures are pretty much cherry-picked for quality sake and because Civit's idiotic oversensitive filter will most likely run to mommy and putt everything under XXX just for a single harmless tentacle or some shit. It's not trained on anything NSFW, but most creatures or monsters have in general no clothes, so yeah, you may need to specify that. Should be no problem, because text-encoder doesn't get trained ( unfortunately to much RAM needed ) and Chroma doesn't use CLIP as far as i know, so only U-Net should be flexible enough.
It can and will produce much more fucked up things if asked to. I don't really know what people would call creepy, uncanny or whatever since i have no limit / barometer when it comes to things like that. Quite a few pictures where from old horror movies and some from new ones, a lot of them pretty niche. Also artsy images, abstract, biopunk and environments.
If you wanna really go for it use horror, movie, slimy, translucent, movie, cinematic etc. In a lot of cases the quality will suffer though because of the base material. Lower strength is maybe better or if you know how to prompt it can be negated for the most part.
For now i just wanna make some "base" loras which i can mix and match later or for merging with other loras to enhance certain aspects.
Description
FAQ
Comments (8)
Did you intend to spell Creppy in the title, or did you meant to spell Creepy?
Nah, fixed it.
@TijuanaSlumlord Cool.
Awesome,it's finally here, i will test it against your original Creepy Lora to see if it better to make that weird scary stuff
uptdate: comparing to the original, this somehow feels more family friendly creepy stuff
On general training terms... what do you think of Chroma against Qwen, FLUX or SDXL?
No idea about Qwen to be honest. SDXL and all its variants are the most common and i think it's easy to train and the quality is good too, just the coherence sucks. With Flux it's kinda strange sometimes to train something. The base model is extremely restricted and it's tricky to train concepts or styles without Flux fighting like crazy. Pretty much like "Oh, i don't know this style, but here is something that is similar, take this." and it will do it's damndest not to deviated from it. Sometimes that's actually a good thing, because you can take the most terrible quality images to train but get something better looking. Chroma is very easy to train, will go for pretty much everything, even after just a few steps. So far, in a lot of cases after seeing a preview image while training, i get the feeling Chroma is already trained with the same picture, even the most weird stuff. If you really go for it and know what you're doing, Chroma doesn't need any fine-tuning or lora imho. But for the sake of curiosity, random mix-matching and to "throw it off course" i'll do it anyway.
@TijuanaSlumlord Thanks!
Chroma will need some finetunes for very specific content or major themes. It is capable of doing much of what even Flux 2 can do out of the box if you spend time on the prompt. I have had mixed results with low rank LoRA training for Chroma, some work at rank 4 the first time and some require rank 16 (70 MB) along with masked training and customized text input. You don't have to beat it with a hammer as much as you do with Flux but the size of the base model needs way more processing than SDXL.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.