# Character Swap using SAM3 to inverted color mask + MiniMax H3 — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/67a874c6ccccf50b

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-13 01:50 ingested 2026-10-14 02:40

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1x4k7sq/character_swap_using_sam3_to_inverted_color_mask/)

## retained text



```
  
I followed the idea from this:

[https://www.reddit.com/r/comfyui/comments/1wvff4i/super\_simple\_minimax\_character\_swap\_workflow/](https://www.reddit.com/r/comfyui/comments/1wvff4i/super_simple_minimax_character_swap_workflow/)

To take SAM 3 to create the mask:

[https://i.imgur.com/bUYkWk0.jpeg](https://i.imgur.com/bUYkWk0.jpeg)

Then using the ref2video model, I used the masked video as input reference and the reference images from my dog:

[https://i.imgur.com/FNX4wsu.png](https://i.imgur.com/FNX4wsu.png)

Prompt is something like this:

[https://i.imgur.com/JUTwJIm.png](https://i.imgur.com/JUTwJIm.png)

Took some guidelines from here as well:

[https://patricksheedy.com/posts/minimax-h3-character-swap-part-2/](https://patricksheedy.com/posts/minimax-h3-character-swap-part-2/)

That other guide is interesting, but it creates a black mask that loses all the facial information. Negative color is way better.

Here is my workflow:

[https://drive.google.com/file/d/1DK0n2fguT8O4JlgDTth1-JPy-R3bbY3N/view?usp=drive\_link](https://drive.google.com/file/d/1DK0n2fguT8O4JlgDTth1-JPy-R3bbY3N/view?usp=drive_link)

It has a node for Character Lora, but it was disabled for this flow.
```



document fetched 2026-10-14 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1x4k7sq/character_swap_using_sam3_to_inverted_color_mask
