higher resolution

[Yijunmaverick]

Thanks for sharing.

Is it possible to run this method on higher-resolution input? or 256 only?

[thaoshibe]

Appreciate your question! Resolution is not a fundamental constraint of CPM. We can use a higher resolution for sharper images. For example, by using 512x512 UV maps, we can get higher resolution images, as shown in the following figure.

(From left to right: 256x256 UV map (as shown in the paper), 512x512 UV map, and reference image)

[Yijunmaverick]

Thanks for the quick reply. Is 512 UV map obtained by another PRNet model or some upsampling technique? I may want to try ever larger size :)

[thaoshibe]

For the 512x512 UV map, I used the same PRNet model (no need to retrain). There's will be another face_ind_512.txt, uv_kpt_ind_512.txt files. (As your request, I'll upload it later).

---

But it worth noting about this method:

• [01] Basic (As shown in the paper): Input -> UV (256) -> Color (256) | Pattern (256) -> Output Image (256)
• [02] Higher resolution UV (As shown in previous figure): Input -> UV (512) -> Color (256) | Pattern (256)-> Output Image (512) (Upsampling used in Pattern Mask & Color transferred TsmC)
• [3] "Wholesome" Solution: Input -> UV (512) -> Color (512) | Pattern (512) -> Output Image (512)

I'm using [2] because that requires no re-train step (PRNet, Color, Pattern). To be specific, I do use upsampling for intermediate output (Color Transferred (bicubic resize), Pattern Mask (nearest neighbor resize)). (See attached image) (Getting higher resolution, based on [2])

---

You might wonder why didn't I use [3]? Isn't it the best solution? To do [3], we'll need to re-train Color Branch & Pattern Branch. It's feasible, but I haven't done the job due to: time-efficiency (it'll be super slow), dataset (most of the makeup datasets are 256x256 only), etc. ugh!

Hope this helps!

[Yijunmaverick]

Got it, thanks for explanation. The upsampling makes sense.