This LoRA was trained on my photographs taken with the Canon 1Ds. It is mainly designed to reproduce its rich colors, distinctive rendering, and beautiful bokeh that makes old flagmanship full-frame camera. The dataset includes images shot with 28mm, 50mm, 100mm, and 500mm lenses, so the resulting look may vary depending on the subject, composition, and focal length
Prompt should be detailed (trigger word is not enough), try to describe about out-of-focus objects on foreground or background
Description
FAQ
Comments (14)
thanks, I'm checking it out, can you give an insights on training parameters? Like how many images in the dataset, batch count , step count, learning rate? I'm interested in trying a project like this
Awesome! When your fave creator drops a new LoRA, it's going to be a good day.
That textfusion lora was driving me crazy yesterday lol. It works so well for so many things, but when it doesn't it took me forever to find the culprit. I had already cut my lora stacks way down, this lora might just convince me to keep it there.
I use it at 0.8 strength as suggested at ~9 steps on raw using [email protected] (it's a little crisp, but good at variety). I might not even need analogcore/vhs, all tests are coming back really good so far.
Surprisingly good, especially for a modern realism lora when so many others seem to break the outputs at most strengths. I suspect most of the modern loras aren't using as consistent of a dataset, or you're just a wizard, 'risi.
edit: the tag "available light" is amazing.
Hold up, what did you use with 0.8 strength?, this Lora or the text fusion?
@Neon_signs this. TF I was only using at 1.0. It was just a few sticky prompts that were driving me crazy yesterday.
Lora slows down generation at least 5 times
it's with turbo?
Not for me. I'm using it on the Turbo version of Krea 2, INT8 ConvRot quantization. Are you using a GGUF version of Krea 2 by any chance? GGUF-quantized models are known for being slow when LoRAs are used.
@mmdd2543 i use raw convrot and speed is the same as without lora
Slowdown with just this LoRA, or with any LoRA? If you are using GGUF or int8convrot, LoRAs may cause problem (not just this LoRA but all LoRAs).
Try updating your ComfyUI to see if the problem goes away (save your current comfyUI before updating in case the update breaks something).
@Danrisi What are your generation times?
Sick LoRA! This brings Krea 2's realism to a whole new level! I can't recommend it enough for realistic DSLR-like images.
ok this looks fantastic i somehow missed it earlier but i really enjoyed the zit version and now will be trying out the krea 2 version :D
So a weird thing happened. I was testing this LoRA and asked for some prompts with an LLM. And I was suddenly getting all cartoonish results. Seems the TextFusion_Refusal_Reduction LoRA was creating some cartoon style because of "c2n0n style" in the prompt.