GPEN is better ?

[papagodzilla]

thank you for work you have done with roop unleased and regularly updating too. i want to ask everyone who are using roop, it is for me or GPEN giving better result than codeformer specially with eyes and skin tone ?

[couldnt-find-good-name]

yep, GPEN is giving better results with eyes.

[tikendraw]

GEPN > GFPGAN > DMTNet > CodeFormer for me. I tested on some videos.

[DarthKX]

What about Restorformer now that it's there? Is it even better? Thx in advance

[papagodzilla]

Restorformer produce blury eyes like opposite of GPEN. honestly GPEN is still winning almost everytime (according to my prefrence) .

[C0untFloyd]

I personally like DMDNet the most cause it took the most effort to port 🥲
Honestly I also like it the the best in terms of face identity and general appeal. You need to tone down the blending a lot though to prevent the glossy/rubbery look.

[GrungeWerX]

I'd like to try it out but it's so slow. What settings do you recommend?

[lysxelapsed]

Are you talking about DMDNet or GPEN now? DMDNet was by far the fastest enhancer in my benchmark: #461
Besides: You can try out everything from enhancers to settings on a short video. Cut some sequence(s) out your prefered target video with Losslesscut, fast & easy 😃 Or use the 10-seconds-clip from the benchmark thread 😉

I'm always torn between Codeformer at 0.9 blend ratio and GPEN at 0.5 - 0.6.. these are my go-to's for a while now.

[C0untFloyd]

What do you guys think about RestoreformerPlusPlus? I think it's ahead of the others now and seems to be like a mix of codeformer and gpen.

[lysxelapsed]

Restoreformer++ is a hard no for me. This is compared to Codeformer - direct original-size-comparison looks even worse than in the previews here.

Restoreformer++:

Codeformer:

[dastommygun]

About DMDNet , I'm gettin a OOM error when trying to use it, quite weird since I've a RTX 4070 with 12GB of VRAM:
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 400.00 MiB. GPU 0 has a total capacty of 11.99 GiB of which 0 bytes is free. Of the allocated memory 22.76 GiB is allocated by PyTorch, and 56.46 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

No issues with GPEN and Codeformer.

[dastommygun]

Will do... thanks for trying anyway!

[C0untFloyd]

Just out of curiosity, can you try running DMDNet with only 1 thread after a clean start? It could be a thread race condition, as this is the only enhancer still working with Torch and can't be easily ported over to the ONNX Architecture like the others. There's no way it's using that much VRAM, I can use it without problems on my 8 Gb GPU.

[dastommygun]

Nope... even on one thread, on a clean start:
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 1024.00 MiB. GPU 0 has a total capacty of 11.99 GiB of which 0 bytes is free. Of the allocated memory 23.96 GiB is allocated by PyTorch, and 48.15 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[dastommygun]

I've tried to perform a clean installation of roop-unleashed and during the pip install -r requirements.txt step I've got this error:

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
tensorflow-intel 2.13.0 requires numpy<=1.24.3,>=1.22, but you have numpy 1.26.2 which is incompatible.
tensorflow-intel 2.13.0 requires typing-extensions<4.6.0,>=3.6.6, but you have typing-extensions 4.8.0 which is incompatible.
numba 0.57.1 requires numpy<1.25,>=1.21, but you have numpy 1.26.2 which is incompatible.
Successfully installed aiofiles-23.2.1 altair-5.2.0 anyio-4.3.0 attrs-23.2.0 click-8.1.7 exceptiongroup-1.2.0 fastapi-0.110.0 ffmpy-0.3.2 ftfy-6.2.0 gradio-3.44.2 gradio-client-0.5.0 h11-0.14.0 httpcore-1.0.4 httpx-0.27.0 importlib-resources-6.3.1 jsonschema-4.21.1 jsonschema-specifications-2023.12.1 numpy-1.26.2 onnx-1.15.0 onnxruntime-gpu-1.16.3 opencv-python-4.8.1.78 orjson-3.9.15 pandas-2.2.1 psutil-5.9.6 pydub-0.25.1 python-multipart-0.0.9 pytz-2024.1 pyvirtualcam-0.11.1 referencing-0.34.0 rpds-py-0.18.0 semantic-version-2.10.0 sniffio-1.3.1 starlette-0.36.3 toolz-0.12.1 torch-2.1.2+cu118 torchvision-0.16.2+cu118 tqdm-4.66.1 tzdata-2024.1 uvicorn-0.28.0 wcwidth-0.2.13 websockets-11.0.3

May this be related to the problem I have?

Anyway, something may be different: after the clean install, by reducing the threads to 1 and trying several times, I was able to run DMDnet occasionally, but when it does it's suuuper slow:

[lysxelapsed]

May this be related to the problem I have?

Not sure, but could be. You can adjust the requirements.txt file to the versions specified (numpy==1.24.3 for example), save the file and run pip install -r requirements.txt again. If you already used this install, do a clean one (again) or delete every __pycache__ -folder before you run it again.

[p1ski]

Im having the same issues.. I have a 2080ti with 11gb ram and the only post processing algos that work are restoreformer++ and codeformer.. All the rest throw out of memory errors.. Doesnt matter what I do , It even bricks out the app and I have force close it down to work again..

[dastommygun]

I'm feeling your pain... at least GPEN is working for me.
I wasn't able to fix the issue on DMDnet, should you find any workaround let us know!

[p1ski]

Yeah its really frustrating, the op said hes running them all fine on 8gb card, The fact that not many others complain about it puzzles me.. Wonder if those with radeon cards are having none of these issues.. I can get them to work somtimes on cpu, but that also somtimes throws errors, and if it does work its chronic slow...