Statement: In order to prevent malicious theft on other platforms, the model will be provided for irregular downloads, and the map running remains the same.

Make friends, join the AI mutual aid group, Q Group number:482570609

→→→→→→Get more high-quality models←←←←←←

SDXLrender_v1.0 : Simulates the real rendering effect of SDXL, improves the output quality of SD1.5

It is well known that SDXL is powerful and cumbersome. While improving the realism to a higher level, it brings slower rendering speed and higher configuration requirements (not to mention the rumored price jump to 30,000 for the 4090). This makes us small partners with limited funds very embarrassed. Therefore, I want to explore the limits of SD1.5 again. After all, the quality of many realistic large models in the community is not inferior to XL, so why not give it a try?

Advantages of SDXLrender:

1. Carefully selected hundreds of high-resolution SDXL character images for training to restore the realistic lighting and delicate skin texture of SDXL, saving a whopping 30,000!

Test model: majicMIX. Lora weight 0.6, with significant improvements in lighting, contrast, and especially skin texture.

2. You can get good results almost without any quality prompts, reducing the burden of inputting prompt words

Test model: SD1.5 basic model. From the weight test image, it can be seen (without adding any quality prompt words in the prompt, only using 1gir, portrait, outdoors+SDXLrender) that 0.0 has no effect, 0.1~0.2 starts to increase details, and from 0.3 onwards, optimization rendering begins for larger areas.

3. More sensitive to composition prompt words.

Test model: RealisticVision_v5.1. I selected the most commonly used nine sets of composition prompt words for testing, and the results are better than the original model's output. Lora weight is 1.

4. This is indeed a pleasant surprise. When SD1.5 is outputting large images, it often gets messy, even with multiple hands, feet, or people. But through this lora test, I found that it significantly reduces such occurrences, which may be related to the dataset mostly consisting of single-person images.

Test model: epiCPhotoGasm_x. After 0.3, there are no more instances of multiple people.

Directions for future version iterations (hopefully):

1. Sensitivity to hairstyle and clothing prompt words. I once tried to have a girl style a man's hair, to no avail.

2. If you want to draw a portrait, but the prompt words include pants, it will most likely turn into a full-body portrait. I currently don't know how to solve this.

3. Color contamination seems to be a persistent problem with SD1.5, a cliché issue (am I being difficult with myself...)

If you have any suggestions for improvement, please let me know, and yes, this lora seems to lean towards NSFW. Please do not use it in the company of others

Example of prompts/parameters

parameters

editorial fashion photo,
a woman,portrait,from below,crazy,(yellow glow eyes:1.2),black choker,yellow background,golden and black colors,<lora:SDXLrender_v1.0:1>

Negative prompt: NSFW

Steps: 36, Sampler: DPM++ 2M SDE Karras, CFG scale: 6, Seed: 4172998084, Size: 640x960, Model hash: ef76aa2332, Model: Realistic Vision V5.1, Clip skip: 2, Lora hashes: "SDXLrender_v1.0: 88ee78180c38", Version: v1.6.0

Respect copyrights, respect individual labor achievements, when using this model elsewhere, please credit the author. This model is for learning and communication purposes only, do not use it for any commercial activities.