ComfyUI_GridSwapper

ComfyUI_GridSwapper
★ 28

图像生成潜变量处理网格交换去噪增强
将一批潜变量分配到多个网格并在每步随机交换图像，使各图像在去噪过程中互相影响，提高多图生成一致性与多样性。
💡 用于让同批图像在去噪过程中相互影响以提高多图质量。
🍴 2 Forks💻 Python🔄 2024-10-27
📦
网盘下载
复制链接后前往夸克网盘下载
https://pan.quark.cn/s/51137d50651f
📄 README

Gridswapper

Gridswapper takes a batch of latents and spreads them over the necessary amount of grids. It then automatically shuffles the images in the grids for each step.

So, a batch of 12 latents for a 2×2 grid will generate 3 grid images in each step. It will then shuffle around the images for the next step. This makes it possible for all images to influence the others during the denoising process.

This approach works well for generating 2-4 grids.

To improve convergence, especially when generating many grids, consider:

  • Increasing the number of steps, (for example 4*batch size) – this makes it more likely that each image can influence all others during the shuffling process
  • Using ancestral samplers (like euler_a) – the added noise at each step seems to help with consistency
  • Train a lora on 2×2 grids of the type of pictures you want to generate.