Info about this style:
The Harlem Renaissance was an intellectual and cultural revival of African-American music, dance, art, fashion, literature, theater, politics, and scholarship centered in Harlem, Manhattan, New York City, spanning the 1920s and 1930s. Source : Wikipedia
KREA 2 version: Works on both long narrative prompts and short prompts with tags
Z-Image version: Works best with short prompts only
SDXL version: Prompts with tags
Illustrious version: Creates a more comic styling, not really in sync with the other models.
Flux version: Prompts with tags
Description
Trained on less images and trigger word is "harlem_renaissance" and triggers afro style. However if you want a painting you need to add "painting" in prompt aswell otherwise it might generate more realistic images that look less like the images it was trained on. Possible cause is that I put "painting" as word in training texts aswell.
FAQ
Comments (8)
Really useful for having a more generic African American look in um... more sexually explict work.
I will post more ... on-site generated content that fits the PG to PG13 content of the movement. ... and maybe figure out if there is a good equilivent to ((((trait)))) in caption based prompting vs. tag based prompting.
Thanks for all the posts! Z-Image understands many racial differences when used in prompts but somehow long prompts will weaken the Lora. Also the on-site configuration is slightly off from my offline ComfyUI setup. Some claim if a Lora is trained with atleast 4400 steps it will produce better quality. I might train a new version of this Lora this week and test if it will hold on extensive long prompts.
@egpieper
Honestly I think it is me attempting to fight a seperate Lora to generate black ladies, when it may have not had that in the training data.
The Lora is pretty good at producing mimics of the Harlem Renaissance photorealistic artto a bit more abstart art.
@sawbladex818 I was unable to mix Lora's myself. I just found a patch lora that alows mixing diffrent Z-Image Lora's not sure it works on the online generator, but it looks promissing. https://civitai.com/models/2233614/z-image-turbo-distill-patch-for-comfyui
It seems this patcher LORA work online aswell, it will patch disfigured results when using mixed Lora's and should be on strength 1, but it seems the other LORAs need to be set to 0.6 max to get better results.
Would love to see a Krea 2 variant.
Hi Hamaram888, My local setup is not configured to train Krea 2 Lora's yet. So I tried the CIVITAI Online trainer with my old dataset. Seems it did the job!
@egpieper That's wonderful! I'll give it a go!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.