santa hat
deerdeer nosedeer glow
Sign In

Silhouette Synthesis 逆光幻境

628
3.4k
6.7k
75
Updated: Oct 5, 2024
styleblade runnerwoman
Verified:
SafeTensor
Type
LoRA
Stats
3,422
6,699
Reviews
Published
Aug 19, 2023
Base Model
SD 1.5
Training
Steps: 10
Epochs: 15
Usage Tips
Clip Skip: 2
Hash
AutoV2
9D83F9AD63

Goodbye to the fucking internet-famous face shit!

The biggest issue with the current large models in terms of lighting is the separation of "people" and "objects". It's as if a person was pasted into a scene, missing the depth that lighting brings, which is a significant reason why it feels so AI-esque.

So, I began trying to improve this. If one were to delve deep to address this, the top priority would be the excessive or unreasonable lighting on the characters. The best way to solve this would be to create ample backlighting, allowing the character and the environment to blend more seamlessly.

Inspired by 'Blade Runner 2049', I decided to showcase my model with a subdued color tone, a style with distinct dark and bright contrasts. I named it 'Silhouette Synthesis'.

目前的大模型在光影层面最大的问题是“人”与“物”的分离,像是在一个场景里拼贴了一个人进去,光影带来的层次感是缺失的,这也是ai感十足的重要原因。

于是我开始试着去改善这一问题。这个问题如果想要去细分解决,那最优先应该解决的就是人物身上过度或者不合理的光照,那么解决这个问题最好的方式就是制造足够的背光面积,这样可以让人物和景物更好的融为一体。

同时,基于bladerunner 2049的启示,我决定用一种色调压抑、阴暗对比明显的光影风格来展示我的模型,同时起名为 Silhouette Synthesis。