This AI makes blurry faces look 8 times sharper! SwinIR: Photo Upsampling

Поділитися
Вставка
  • Опубліковано 7 січ 2025

КОМЕНТАРІ •

  • @WhatsAI
    @WhatsAI  3 роки тому +2

    References:
    ►Read the full article: www.louisbouchard.ai/swinir/
    ►Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L. and Timofte, R., 2021. SwinIR: Image restoration using swin transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 1833-1844).
    ►Code: github.com/JingyunLiang/SwinIR
    ►Demo: replicate.ai/jingyunliang/swinir
    ►My Newsletter (A new AI application explained weekly to your emails!): www.louisbouchard.ai/newsletter/

  • @hexenkingTV
    @hexenkingTV 3 роки тому +3

    It seems like classic cnns combined the self attention layer mechanism from transformers are the next big key in image ai use cases.

    • @WhatsAI
      @WhatsAI  3 роки тому

      Yep it seems like! Both have complementary strengths and limitations!

  • @funcuriosity8739
    @funcuriosity8739 10 місяців тому

    6:12 pic right , it like a blurry, i see like it on tv Samsung s95c on youtube , and different in xbox serie x in youtube it like more detailed or grain ,for that i want test the upscaled , what you recommend, use usb and put image 1080p converted in video hevc for see how work the upscaling in tv? How can i know upscaling work ? I guess if it fit in full screen. --- edit some time it better on UA-cam app on Samsung and sometimes in xbox app 🤔, and the xbox serie x dont get full screen 4k pic in 4k tv while if i try in tv have the full screen in 4k pic

  • @giallo_o
    @giallo_o 3 роки тому +1

    Man I love your channel :D

    • @WhatsAI
      @WhatsAI  3 роки тому

      So glad you do! Thank you 🙏 ☺️

  • @jojosakaro6669
    @jojosakaro6669 Рік тому

    Can you make video on this whole code by explaining it from the begining.....?

  • @Seanweekhaizhen
    @Seanweekhaizhen 3 роки тому +1

    Can you make a tutorial on how to use it? I'm not familiar with running code

    • @WhatsAI
      @WhatsAI  3 роки тому

      I can help out but do you need to run the code absolutely or do you simply want to try it out? Because if you just want to try it there’s a demo where you don’t have to code, it’s linked in the description in the references :)

    • @Seanweekhaizhen
      @Seanweekhaizhen 3 роки тому +1

      @@WhatsAI I'll give it a try
      Thanks!

  • @M_Jema4703
    @M_Jema4703 3 роки тому +3

    Man this will be game changer once it overcomes limitations. Still results are amazing👌👏
    CNN + Transformer = Mind blowing🤯😍

  • @slashkeystudio
    @slashkeystudio 2 роки тому +2

    hold on to your papers

  • @hypersonicmonkeybrains3418
    @hypersonicmonkeybrains3418 2 роки тому

    Im a total noob here, but quick question. Is the upscaling used in DALLE-2 (diffusion or whatever) superior to this? And have Nvidia got an even better one as well?

    • @WhatsAI
      @WhatsAI  2 роки тому

      Diffusion models are quite good for upscaling, and I think some recent models are indeed better than this one, but are most certainly comparable for now!

    • @WhatsAI
      @WhatsAI  2 роки тому +1

      So basically use the one that you have the code for and is easily adaptable to your pipeline haha

  • @RaviPatel-vo4hn
    @RaviPatel-vo4hn 3 роки тому +6

    zoom and enhance😂

    • @Kinos141
      @Kinos141 3 роки тому +1

      That tv tropes is becoming a reality.

  • @lunarripper835
    @lunarripper835 3 роки тому +2

    Yay now I can make ppl's csgo vidoes more better... wait it's only for images???? wtf.....well no worries I will make every frame better one by one and make a video!

    • @WhatsAI
      @WhatsAI  3 роки тому

      There are better methods for videos! It is much more easy to enhance a video than an image and requires different networks :)

  • @etnaflixmedia353
    @etnaflixmedia353 3 роки тому +1

    too much artifacts for me there is something better repos

    • @WhatsAI
      @WhatsAI  3 роки тому +1

      Which repos would you suggest? :)

    • @etnaflixmedia353
      @etnaflixmedia353 3 роки тому +2

      @@WhatsAI 1st pass Gpfgan+ 2nd Real ESRgan or GPEN+Real ESRgan in both case better results

  • @3d-illusions
    @3d-illusions 2 роки тому

    seems to lose a lot of detail, becoming overly smooth. For example the grey haired fellow's shirt has no pattern afterwards, and his skin is smoother than a baby's arse. I can see the correct skin and shirt details in the low res image, so I think the model is lacking something here.

  • @vodkaru5907
    @vodkaru5907 3 роки тому +3

    Tried a few images. Not impressed at all

    • @WhatsAI
      @WhatsAI  3 роки тому

      Really?! I was really impressed with images around 512*512 to 2048! What did you try?

    • @vodkaru5907
      @vodkaru5907 3 роки тому +2

      @@WhatsAI 300 - 400 px. I'm not saying it does a poor job. It's just not that impressive

    • @vodkaru5907
      @vodkaru5907 3 роки тому

      @@WhatsAI You think it's the best one around yet? I haven't tried all the papers.

    • @WhatsAI
      @WhatsAI  3 роки тому +1

      Oh yeah like I said in the video it doesn’t work that well for smaller images! I didn’t test with 300-400 pixels but tried with 100-200 and it was quite bad. You should try with images of 500 or larger pixels and let me know what you think! I believe the x4 on 512 is amazing compared to other approaches I’ve seen, and it seems to stick to the real picture much better as well

    • @WhatsAI
      @WhatsAI  3 роки тому +1

      At least, I can assure you it is the best one that gives pretrained models and a free live demo haha!
      Other papers might be better, but we have to trust the results in the paper or retrain/reimplement them. This one is not corner cases or hand picked results and still it works well. This is what’s most impressive!