This section is a incomplete. Phantom and MAGREF are relatively old finetunes based on Wan 2.1, so called “reference” models.
| Model | Type | Notes |
|---|---|---|
| Phantom | T2V | Could theoretically work with VACE, also exists as T2V LoRA |
| MAGREF | I2V | - |
Phantom can take 4 inputs; Phantom [likely] tries to use 1 reference per subject
If you’re using the HN/LN split, keep in mind that Phantom will only work with the LN part, so you have to either skip the HN stage totally (which is perfectly fine to do), or you have to make the HN stage output look close enough to your target references that Phantom can pick up where it left off and finish the details properly
It has been shown that Phantom generation is a somewhat influenced by Lynx.
with MAGREF you can sort of adjust the strength by padding your image with white; so for example put your full image without any padding as main image, then a version of that with padding to this and it sticks to it less
Q: how do u use magref?
A: Exactly like normal I2V, except you either remove background or pad the input with white
Q: any recommended sampler to test against it? to see realism, because when I use more steps in dpm++_sde can end up with a burned image, over coocked one? phantom-magref.md
A: With distil Loras, try lcm instead for less burn
Can be used with WanVideo Context Options.