[Feature]: codeformer_fidelity (done) and codeformer_scale support

[Arcitec]

Feature description

codeformer_fidelity

After seeing what https://github.com/mav-rik/facerestore_cf can do with its codeformer_fidelity parameter, I realize that it's an incredibly important aspect of face restoration.

• GFPGan: Doesn't support it/doesn't use that parameter.
• CodeFormer: Requires that parameter. At 0, the face looks nothing like the input but becomes very aesthetic. At 0.5 it's a mix. At 1.0 it very closely matches the input face but becomes high resolution. This usually produces the best faces.

Without that parameter, face restore in ReActor node is almost useless.

I could set ReActor to face restore "none" and route the output to the separate "facerestore_cf" node, yes, but this is slower because that means it has to do face-detection twice. So it's better if codeformer_fidelity parameter is supported in ReActor.

It's gives a HUGE quality increase! :)

codeformer_scale

This 2nd parameter is something I just discovered now and haven't seen before. You can read about it in this article:

https://blog.segmind.com/best-settings-for-codeformer/

Basically, scale is how detailed and what resolution the output face will be. It's another huge quality increase. Look:

Why?

Together, fidelity and scale makes codeformer face restore quality massively improved!

Would be a huge improvement for ReActor node. The current face restoration in reactor-node is operating at like 30% of the quality that can be achieved if those parameters are added. :)

[Gourieff]

Thanks for the info 👍 I will take a look

[ardha27]

Feature description

codeformer_fidelity

After seeing what https://github.com/mav-rik/facerestore_cf can do with its codeformer_fidelity parameter, I realize that it's an incredibly important aspect of face restoration.

• GFPGan: Doesn't support it/doesn't use that parameter.
• CodeFormer: Requires that parameter. At 0, the face looks nothing like the input but becomes very aesthetic. At 0.5 it's a mix. At 1.0 it very closely matches the input face but becomes high resolution. This usually produces the best faces.

Without that parameter, face restore in ReActor node is almost useless.

I could set ReActor to face restore "none" and route the output to the separate "facerestore_cf" node, yes, but this is slower because that means it has to do face-detection twice. So it's better if codeformer_fidelity parameter is supported in ReActor.

It's gives a HUGE quality increase! :)

codeformer_scale

This 2nd parameter is something I just discovered now and haven't seen before. You can read about it in this article:

https://blog.segmind.com/best-settings-for-codeformer/

Basically, scale is how detailed and what resolution the output face will be. It's another huge quality increase. Look:

Why?

Together, fidelity and scale makes codeformer face restore quality massively improved!

Would be a huge improvement for ReActor node. The current face restoration in reactor-node is operating at like 30% of the quality that can be achieved if those parameters are added. :)

can i find "codeformer_scale" on comfyui?

[Gourieff]

codeformer_fidelity is implemented https://github.com/Gourieff/comfyui-reactor-node/commit/ea536a43c8527341f1426d19ca438b4a0329d9e1

[KINGLIFER]

May I at least inquire how to implement codeformer_scale

[Gourieff]

But why do you need this option? You can scale with any model you want via Upscaler's Nodes or via Ultimate Upscaler

[KINGLIFER]

No this is the face because of the limitations. I made something that upscales a face but why have all those nodes when one can do it itself.

[Gourieff]

Because it's not the best way to upscale, there are a lot of models that can provide better quality

[KINGLIFER]

for the details of a faceswap? Can you be more specific?