Recommended Settings (v1 - Z-Image Turbo):
WORKFLOW IS IN THE SHOWCASE IMAGES. You have to use the workflow. If you use the LoRA with a simple workflow you:
Either get shitty overtrained plastic images
Or you barely change the style of turbo model and they might still look shitty
If you use it with my workflow, the seed variety is phenomenal. There is a very very less chance that you get the same-ish composition and lighting twice
Prompt adherence took a little hit (maybe its the workflow to blame). It is okay for me because it still generates the images I want. I loose a bit of adherence to get variety. If you dont like it then you can try other workflow's or reduce the weight of the LoRA or use another LoRA
Some prompts change the image way too much when you compare with Turbo but some prompts not so much. You can start with weight 0.7-0.9 for "Model with LoRA High Weight" node and 0.4-0.5 for "Model with LoRA Low Weight" node. The last stage (stage 3), disable the LoRA so you can bring back the skin texture - the image barely changes here because we use a low denoise. You can use the upscaler that I am using but keep in mind that its slow compared to others
If you want to use a character LoRA along with this LoRA, you have to enable the character LoRA at stages 2 and 3. In stage 2, you can unbypass the character lora loader node. In stage 3, you can use "Model with Character LoRA Only" node
Recommended Settings (v1 - Qwen Image):
Workflow is in the showcase images
My recommendation is to use the "EasyCache" node and let it skip 2 steps. It good with very little quality loss
I never tested it without lightning lora so I don't know how many steps to use or the CFG. You have to experiment yourself
880x1184 or 1328x1776 works best
Recommended Settings (v6 - Flux Dev):
Distilled CFG Scale: 3.5
Sampling method and Schedule type: DEIS with DDIM
Steps: 20
Resolution: 896x1152
Hires fix model: 4x_NMKD-Superscale-SP_178000_G
Steps: 10
Denoise: 0.3
Upscale by: 1.5
Lora Weight: 0.8. You have to experiment based on your prompts
Description
GPT4o Prompt:
I am planning to train a LoRA for the Stable Diffusion text-to-image model, which uses the T5XXL transformer in its architecture. The prompts should be in natural language and follow a specific format. I will upload images and need you to help me create detailed prompts based on those images. The prompts should start with "Amateur photography of" and end with "on flickr in 2007, 2005 blog, 2007 blog." Always give me the prompt in a single paragraph.The format should be:Subject Description: Start by describing all the people in the image in detail. It is very important to include their race and ethnicity, physical attributes (such as height, build, skin tone, and hair color), facial features, attire, and any expressions or poses they are making. Be as specific as possible. Make sure to always include the build of the subjects (e.g., plus size, slim, petite) without missing it.Scene Description: Accurately convey what exactly the people are doing in the picture. Describe the setting, background elements, any objects they are interacting with, and the overall environment (urban, rural, indoor, outdoor, etc.).Image Quality Tags: Include descriptive tags that highlight the quality of the image. Use terms like slight motion blur, cluttered background, warm tones, bright natural light, high contrast, vivid colors, etc. These tags should reflect the mood and feel of the image as well.The final output should combine all these elements into a cohesive, detailed prompt that accurately reflects the image.