What is v34-detail-calibrated?
"the detail calibrated is basically v34 + high res run. 2:1 merge ratio"
I've (kind of) answered myself
It's the chef's kiss, that's what.
I'm not entirely sure what it is, but it's insanely good. It's significantly better and more coherent than any other version I've tried, including generated text
I'm not an expert, but it could be related to the introduction of high-quality, high-resolution images into the training dataset. Early in training, lower-resolution images are typically used to help the model learn fundamental concepts like composition, poses, and structure, etc, without the burden of extremely long training times from higher resolution images. This approach gives the model a solid foundation, but it may lack fine details (maybe that's why it was bad at fingers not a long ago). Once it reaches a certain level of capability, high-resolution images are introduced to help it learn and reproduce those finer details.
@LHC88 when a new checkpoint comes out, Do i pick the detail-calibrated checkpoint or the other one if i am doing photorealism?
I only ask for clarification because often the one without the prefix is commit'ed afterwards, just in case it was an intermediate checkpoint for that version.
It doesn't matter, it never does since it's the same model
I found that the version of chroma-unlock-v42-detail-calibrated with detail-calibrated does not recognize Caucasian/white races. Let it generate Caucasian/white races, and it will generate various other races.Is't a bug?Please check!
It's a feature nowadays...
@Hello-tomorrow-yifei ...Are you serious? The prompt words: "Woman" or "Man" or "Girl" or "Boy" return White people exclusively unless you specifically ask for another ethnicity or "race".
@Hello-tomorrow-yifei ...Are you serious? The prompt words: "Woman" or "Man" or "Girl" or "Boy" return White people exclusively unless you specifically ask for another ethnicity or "race".
Yes.The model is doing great, and this version has a significant improvement in resolution. However, the following two photos don't look like Caucasian, and they look like Indian and Latino, respectively. What do you think? The parameters I set are: cfg = 4.5,num_inference_steps = 20,s5 means seed = 5.prompt:A photographic style image:A 20 years old whole body woman has fit body,large breast and large butt,wearing a beautiful light yellow skirt,and has a charming smile on her face,is standing on the beach.
Adding 'Caucasian' to the prompt is also inevitable. How do you think I can improve this?
@Hello-tomorrow-yifei ...Are you serious? The prompt words: "Woman" or "Man" or "Girl" or "Boy" return White people exclusively unless you specifically ask for another ethnicity or "race".
Yes.The model is doing great, and this version has a significant improvement in resolution. However, the following two photos don't look like Caucasian, and they look like Indian and Latino, respectively. What do you think? The parameters I set are: cfg = 4.5,num_inference_steps = 20,s5 means seed = 5.prompt:A photographic style image:A 20 years old whole body woman has fit body,large breast and large butt,wearing a beautiful light yellow skirt,and has a charming smile on her face,is standing on the beach.
Adding 'Caucasian' to the prompt is also inevitable. How do you think I can improve this?
Sorry, I said it wrong, I should have said they are like Indians and Arabs respectively.The seed is set to 5 and 6,respectively.
Sorry for very slow response (work/sleep/work lol).
This first image is Chroma 42 with a "smoothed out" version of the prompt you showed.:
I kind of see what you might be referring to, The generations have a tendency towards brown hair, and a "suntanned", or "olive" if you will complexion. They all still look "Caucasian" to me, but that's not my focus here right now.
This next image is the same "smoothed out" prompt with the addition of a few descriptive words.:
If you're wanting a specific type of woman then just ask for it...This model will give you exactly want you want as long as you tell it exactly what you want.
Hopefully this helps!
Thanks for your reply! The images you generated are much better quality. Perhaps the main difference is that I used diffusers instead of ComfyUI and didn't use LoRa. What kind of LoRa did you use? I'd like to see how to add LoRa.
I do use Comfy UI, but not for these images. I made them on a site called Tensor.art, if you're trying to crank out a bunch of images quickly and you don't own a monster graphics card you really can't beat it for value. The LoRa I'm using there is called "Hyper-Chroma-low-step-LoRA", and you can download it here:https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/blob/main/Hyper-Chroma-low-step-LoRA.safetensors. I only use it on tensor.art because I'm cheap lol! Tensor.art charges a certain amount of "credits" (that you have to buy) for each picture you generate, and that amount goes up depending on how many steps that you want to use for your picture. With this LoRa I can generate a high quality image there with just 14 steps so it saves me money :). It's very easy to add a LoRa to a generation, usually there will be a folder for them if you're generating locally. What are you using to create your images?
This site sounds like a good choice! I use diffusers in program to call the model.
I browsed the LoRA address:https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/blob/main/Hyper-Chroma-low-step-LoRA.safetensors, but I didn't see the "trigger" words.Do you know what these Lora "trigger" words are?
There are no "trigger" words for this LoRa, just load with your model at default strength (1) and it will greatly reduce the number of steps needed to get a quality image.
Thank you! I'll try it too.