nihui / rife-ncnn-vulkan

RIFE, Real-Time Intermediate Flow Estimation for Video Frame Interpolation implemented with ncnn library
MIT License
817 stars 68 forks source link

[Feature Request] Add new 4.7 & 4.8 models #66

Open xzuyn opened 1 year ago

xzuyn commented 1 year ago

2023.10 - We recently release new v4.7-4.8 optimized for anime scenes! 🎉

Or at least a little guide on how users can convert future models to NCNN themselves?

TNTwise commented 1 year ago

If you recompile ncnn with the memory data parameter set as yes, you can use these models converted by styler00dollar: https://github.com/styler00dollar/VapourSynth-RIFE-ncnn-Vulkan/tree/master/models

If you want a tutorial, the best luck you will have is with this: https://github.com/styler00dollar/VSGAN-tensorrt-docker/issues/37, although, I have tried and had little success with model conversion

Edit: I recompiled a version with the models:

https://github.com/TNTwise/rife-ncnn-vulkan/releases/latest

Edit 2: A guide: https://github.com/TNTwise/REAL-Video-Enhancer/wiki/Convert-Rife-Models

noobtoob4lyfe commented 1 year ago

Just chiming in with a +1 for this request. 4.8 is specifically what I'd love to see here.

RackYican commented 11 months ago

If you recompile ncnn with the memory data parameter set as yes, you can use these models converted by styler00dollar: https://github.com/styler00dollar/VapourSynth-RIFE-ncnn-Vulkan/tree/master/models

If you want a tutorial, the best luck you will have is with this: styler00dollar/VSGAN-tensorrt-docker#37, although, I have tried and had little success with model conversion

What is resemble in the models?

TNTwise commented 11 months ago

If you recompile ncnn with the memory data parameter set as yes, you can use these models converted by styler00dollar: https://github.com/styler00dollar/VapourSynth-RIFE-ncnn-Vulkan/tree/master/models

If you want a tutorial, the best luck you will have is with this: styler00dollar/VSGAN-tensorrt-docker#37, although, I have tried and had little success with model conversion

What is resemble in the models?

I'm assuming you mean ensemble, and an explanation is here https://github.com/nihui/rife-ncnn-vulkan/issues/50#issuecomment-1229203013

2blackbar commented 11 months ago

Hey doods i have a bat file gui for this, just create in and out folders where the exe is but its not hardcoded you can set any folder. It future proof so it detects rife model names by scanning folders beginning from rife word

image

RIFEGUI.zip

RackYican commented 11 months ago

Hey doods i have a bat file gui for this, just create in and out folders where the exe is but its not hardcoded you can set any folder. It future proof so it detects rife model names by scanning folders beginning from rife word

image

RIFEGUI.zip

In my opinion that's more harder than using the terminal version

kikoferrer commented 11 months ago

If you recompile ncnn with the memory data parameter set as yes, you can use these models converted by styler00dollar: https://github.com/styler00dollar/VapourSynth-RIFE-ncnn-Vulkan/tree/master/models

If you want a tutorial, the best luck you will have is with this: styler00dollar/VSGAN-tensorrt-docker#37, although, I have tried and had little success with model conversion

Edit: I recompiled a version with the models:

https://github.com/TNTwise/rife-ncnn-vulkan/releases/tag/20231122

Sorry to bump this thread after awhile. Just wanting to ask if I download just the model folders from your repo, will they work out of the box? Or would you recommend I download yous release version instead? Thanks!

TNTwise commented 11 months ago

If you recompile ncnn with the memory data parameter set as yes, you can use these models converted by styler00dollar: https://github.com/styler00dollar/VapourSynth-RIFE-ncnn-Vulkan/tree/master/models

If you want a tutorial, the best luck you will have is with this: styler00dollar/VSGAN-tensorrt-docker#37, although, I have tried and had little success with model conversion

Edit: I recompiled a version with the models:

https://github.com/TNTwise/rife-ncnn-vulkan/releases/tag/20231122

Sorry to bump this thread after awhile. Just wanting to ask if I download just the model folders from your repo, will they work out of the box? Or would you recommend I download yous release version instead? Thanks!

Download the version from my releases, the models won't work with nihuis binaries. The release will include all the models.

RackYican commented 11 months ago

If you recompile ncnn with the memory data parameter set as yes, you can use these models converted by styler00dollar: https://github.com/styler00dollar/VapourSynth-RIFE-ncnn-Vulkan/tree/master/models If you want a tutorial, the best luck you will have is with this: styler00dollar/VSGAN-tensorrt-docker#37, although, I have tried and had little success with model conversion

What is resemble in the models?

I'm assuming you mean ensemble, and an explanation is here #50 (comment)

Are your models in resemble?

TNTwise commented 11 months ago

If you recompile ncnn with the memory data parameter set as yes, you can use these models converted by styler00dollar: https://github.com/styler00dollar/VapourSynth-RIFE-ncnn-Vulkan/tree/master/models If you want a tutorial, the best luck you will have is with this: styler00dollar/VSGAN-tensorrt-docker#37, although, I have tried and had little success with model conversion

What is resemble in the models?

I'm assuming you mean ensemble, and an explanation is here #50 (comment)

Are your models in resemble?

No, but you can download ensemble models from https://github.com/styler00dollar/VapourSynth-RIFE-ncnn-Vulkan/tree/master/models

TheFramo commented 8 months ago

No, but you can download ensemble models from https://github.com/styler00dollar/VapourSynth-RIFE-ncnn-Vulkan/tree/master/models

Hi TNTwise, i wanted to message you but i did not found a way, im new to github lol.

So im using MPC-HC with a RX 6900XT graphics card and im using SVP+RIFE. I wanted to checkl, if theres new models. So i found your "rife-ncnn-vulkan" repository. Inside the 4.14 model, theres only .param and .bin files. As far as i know, Rife needs those .onnx files to work. Currently im using the official rife_v4.9.onnx model.

How can i use your models? Did i get sthg wrong? I hope you can help me, im kinda new into this stuff! <3

TNTwise commented 8 months ago

Ncnn and onnx are different things. My repo only contains ncnn versions of the models. If you are looking for a .onnx model, they can be found here: https://github.com/styler00dollar/VSGAN-tensorrt-docker/releases/tag/models

TheFramo commented 8 months ago

thanks for making clear!! so would which one would you suggest out of all these 4.14's? lol :) and do i need anything else to run it in svp?

TNTwise commented 8 months ago

Fp16 clamp sim is usually the best. I don't have any experience with svp.

TheFramo commented 8 months ago

okay :) and last question, whats the difference between ensembletrue and ensemblefalse?

TNTwise commented 8 months ago

Ensemblefalse is the default, ensembletrue merges 2 different flows together, it takes longer but can produce a better result.