V.4.0*insert long-winded description here* Also, pineapples are now oversized resins or whatever those things are, bad economy.<hr />V.3.0By far the most convoluted training i have ever done. From to artificial to low resolution to illustration ( don't even know how that happened ) and the list goes on. Everything that matters is that i have my little toy to offset my other loras if they don't behave. The only thing that is still open are more nature photos, landscapes and most important stupid loras.More time to play with other peoples models.<hr />V2.0A slightly different approach. Merged smaller lora batches into the checkpoint and extracted a lora from this. A lot better to manage, since with so many pictures and different themes it takes to long to convervege. With less pictures it's faster and easier to determine when it's enough training.Very happy with the result. It takes the edge of pictures that in some cases still look to artificial or where it looks like a character got photoshopped in.I would keep the strength low ( around 0.7 ish ) for anything people or animal related, because it can fuck up anatomy and chroma is good enough to handle the rest.<hr />Even though Chroma is more than capable of producing this stuff by itself, needed to balance the fuckups i usually train and for combining loras with each other. Not finished yet, but will most likely pick it up in between to continue training till it converges, which by my calculation will take around 5 million years, roughly.Kinda works with normal Flux models but reduced influence, at least in Comfy, no guarantee for your personal setup and version.Trained on version HD rev 0.1And still no category for Chroma...

Chroma - Professional Photos

Different approach and more pictures.

- Chroma - profphotos_high_res_concatextr_2.0.safetensors

v2.0

v4.0

v3.0

v1.0

Even though Chroma is more than capable of producing this stuff by itself, needed to balance the fuckups i usually train and for combining loras with each other. Not finished yet, but will most likely pick it up in between to continue training till it converges, which by my calculation will take around 5 million years, roughly.Kinda works with normal Flux models but reduced influence, at least in Comfy, no guarantee for your personal setup and version.Trained on version HD rev 0.1

V2.0

A slightly different approach. Merged smaller lora batches into the checkpoint and extracted a lora from this. A lot better to manage, since with so many pictures and different themes it takes to long to convervege. With less pictures it's faster and easier to determine when it's enough training.

Very happy with the result. It takes the edge of pictures that in some cases still look to artificial or where it looks like a character got photoshopped in.

I would keep the strength low, because it can fuck up anatomy and chroma is good enough to handle the rest.

Guess it's finished. Adding more pictures won't do much i think.

Files with SHA256: 27a0117d8ca660daffe236e3f73a61b9ac3771073749f996f79da8bbf6187bd7