# Ore wa Subete wo "Parry" suru: Gyaku Kanchigai no Sekai Saikyou wa Boukensha ni Naritai (I Parry Everything) Mk. II --- > Updates: This minor version update replaces the EN subtitles from [KOTEX][source-kotex] with a better source, [SubsPlus+][source-plus]. Thanks to Animorphs for their [suggestion](https://nyaa.iss.one/view/1900545#com-1 "why did you get the English subs from KOTEX instead of SubsPlus+?") and informing us. [![Cover][cover]](https://wall.alphacoders.com/big.php?i=1377433) ## Series `Sav1or` | Episodes | ![MyAnimeList][mal-logo] MyAnimeList | ![TVDB][tvdb-logo] TVDB | ![Nyaa][nyaa-logo] Source --- | --- | --- | --- | --- **Season 1** | 12 *(S01E01-S01E12)* | [**Ore wa Subete wo "Parry" suru: Gyaku Kanchigai no Sekai Saikyou wa Boukensha ni Naritai**](https://myanimelist.net/anime/57058 "Ore wa Subete wo \"Parry\" suru: Gyaku Kanchigai no Sekai Saikyou wa Boukensha ni Naritai") | [**I Parry Everything**](https://www.thetvdb.com/series/i-parry-everything/seasons/official/1 "I Parry Everything") | [Lupin the Nerd][source-lupin], [VARYG][source-varyg], [SubsPlus+][source-plus], [Erai-raws][source-erai] --- ## Release A Dual Audio release encoded in AV1 and OPUS for optimum size and quality. Files are named to support media servers like Plex using TVDB metadata. `Sav1or` | Codec | Quality | Language | Notes --- | --- | --- | --- | --- **Video** | AV1 | 1080p 10-bit || SWAN **Audio** | OPUS | Stereo (2.0) 128K, Stereo (2.0) 64K | Japanese, English | Japanese Commentary 64K **Subtitle** | ASS, SRT | *marginally more decent* | English, French, German | Signs & Songs, Dialogue, SDH **Chapter** || *yes* **Encoding Details**: Encoded with our SWAN system so the preset and CRF values vary from scene-to-scene in order to target a ~85 SSIMULACRA2 score. SVT-AV1-Psy 2.2.1-B with Av1an (`--photon-noise 2`) ``` --preset 4 --crf 28 --aq-mode 2 --tune 2 --keyint 0 --scd 0 --enable-tf 0 --enable-qm 1 --qm-min 0 --qm-max 15 variance-boost-strength 2 --variance-octile 6 --sharpness 2 --qp-scale-compress-strength 3 --noise-norm-strength 0 ``` > This project is a production with the goal of the smallest video size with significant subjective compromises to image quality. This is not the hallmark of AV1, but close enough. This release should only contain changes to the video stream and audio tracks. The subtitle and attachment tracks are usually copied from the noted sources. Timings are not adjusted except in cases of extreme delay as a result of mixing sources. - `Sav1or` ![AV1 Logo][av1-logo] ### Source Thanks to the following sources: * ![Nyaa][nyaa-logo] [Lupin the Nerd][source-lupin]: Video, Audio, Chapter * ![Nyaa][nyaa-logo] [VARYG][source-varyg] - EN Audio, Subtitle * ![Nyaa][nyaa-logo] [SubsPlus+][source-plus] - EN Subtitle * ![Nyaa][nyaa-logo] [Erai-raws][source-erai] - DE, FR Subtitle ### [**Media Info**](https://rentry.co/vubpv49t "S01E10 Media Info") --- ## Comment Commentary @kaiserstar1: We have no plans to cover Helck as [Valenciano](https://nyaa.iss.one/user/Valenciano) has a release [here](https://nyaa.iss.one/view/1760212 "[Valenciano] Helck [1080p][AV1 10bit][AAC][Eng Sub]") and there is no new BD source compelling enough to create a new release to supersede that. If you are looking for EN audio, we recommend muxing that with the [VARYG](https://nyaa.iss.one/user/varyg1001) release [here](https://nyaa.iss.one/view/1871833 "Helck S01 1080p AMZN WEB-DL DDP2.0 H 264 DUAL-VARYG (Dual-Audio, Multi-Subs)") using a tool such as FFmpeg or MKVToolNix. Have a nice day! - `Sav1or` @Krazy0taku: Apologies for the issue. We'll take this opportunity to describe how SWAN works, its strengths, its pitfalls, and how we plan to overcome those pitfalls. Before that, a **TLDR**: Up to ~0.5% of scenes within the past 3-4 releases may or may not have similar issues. We have no plans to redo them unless we find multiple episodes with significantly worse results than what our metrics reported. We expect future releases to be vastly improved in both efficiency and minimum quality; two of the largest issues with the SWAN implementation used up till now. ### Scene-Weighted Av1an Normalizer SWAN is a method to predict or estimate optimal CRF values on a scene-to-scene basis by sampling or essentially encoding small portions of the input between a minimum and a maximum CRF until a target SSIMULACRA2 score is reached within some tolerance. For our fellow scientists and engineers: *Yes, this is indeed a crude implementation of a binary search algorithm! With a hypothetical range of 70 CRF values, the maximum number of sampling "passes" needed would be the ceiling of Log2(70)*. [](https://medium.com/@jacolam/binary-searches-9f0e90e52eff "Binary Search") The Normal-Boost and (Triple) Av1an Bypass methods, which require knowledge of the type of content in order to choose an initial CRF and manually predict (typically assume) the average CRF and SSIMULACRA2 score, also fall prey to the same pitfalls as the early SWAN implementations (see sections below). SWAN is a *vast* improvement over NB and TAB by requiring a minimum of just two parameters from the user: a target SSIMULACRA2 score and the tolerance or how "close" the user wants the estimation to be to the target. Optionally, the CRF minimum and maximum can be limited as well. No more guessing or having knowledge of the type of content being encoded. **Sampling** The sampling method used is the fastest we have devised, but also the least accurate. Despite the minimal data from the sampling method, it's not entirely worthless. The average size of Sakura Quest episodes was nearly *halved* despite scoring just as well in our preferred metric, SSIMULACRA2, as episodes encoded using our previous method, Normal-Boost and Av1an Bypass. Such amazing efficiency, yet the solution wasn't perfect and so we ended up with some pitfalls. **Sampling a Poor Representation of the Scene by Chance** One of those pitfalls that affected S01E08 in that particular scene is that the sampling method measured a black frame in the middle of the scene and considered the scene to be "easy" to encode and used a higher-than-appropriate CRF value. Ideally, that scene should not have contained a few seconds/frames of content in the first place, but the scene detection used is not a perfect method either, so the rare mistake is understandable. Regardless, we cannot entirely rely on scene detection for various reasons. Without going too far into the research, we expanded how many frames are in each sample as well as *where* in that sample we measure for our scene's score. Essentially, these changes alone should make the estimates/predictions far more accurate to their final, optimized counterpart. Big thanks to Lumen for their amazing (and ongoing) research into improving sample accuracy without significantly slowing the process. They are also working on their own implementation of SWAN using their own unique algorithms for CRF prediction. ***I don't like sand. It's coarse and rough and irritating and it gets everywhere.*** - Anakin Skywalker (*Star Wars: Episode II - Attack of the Clones*) Scenes with high amounts of noise or grain will *never* score well in metrics available today due to the fundamental nature of entropy and randomness. The SWAN implementation used will naively attempt to reach the target SSIMULACRA2 score, lowering the CRF until the minimum, ballooning the bitrate to absurd amounts, and ultimately failing to reach the target - assuming the target wasn't unreasonably low. So how do we achieve a good result given this conundrum? Bluntly, we cannot and we do not. As an alternative, we "discount" scenes with poor efficiency, then if the next pass contains discounted scenes with even worse efficiency, we stop estimations for these scenes and apply SVT Film Grain Synthesis with denoising enabled (*experimental*). The general idea is that encoded noisy scenes will never be faithful to the source and should instead be optimized to the extreme in order to curtail bitrate. We are running experiments testing this theory as of this writing, and early metrics look promising. It also provides more headroom to improve the overall quality of the video by targeting a higher SSIMULACRA2 score. **Refining the Optimized Video** We mentioned a final "refinement" pass before as one of the enhancements necessary to make SWAN more reliable. This essentially is a final measurement of all frames in the final optimized video and a re-evaluation of scenes that scored poorly and had a higher CRF than the average. These scenes alone are encoded with the average CRF via our Av1an Bypass method and replaced in the final video. Our current implementation is crude, but effective. A more ideal solution would be to repeat the sampling/prediction process with larger samples. > *Cheers, Enjoy!* - `Sav1or` PS: We somehow hit the character limit? *Absurd!* [mal-logo]: https://i.imgur.com/ZCtSawZ.png "My Anime List" [tvdb-logo]: https://i.imgur.com/dDhSn59.jpeg "The TVDB" [nyaa-logo]: https://i.imgur.com/7o97aB0.png "Nyaa" [av1-logo]: https://aomedia.org/assets/images/home-av1-500x200.png [cover]: https://images.alphacoders.com/137/1377433.png "Parry this you filthy casual!" [source-lupin]: https://nyaa.iss.one/view/1895638 "[BDMV] I Parry Everything!" [source-varyg]: https://nyaa.iss.one/view/1866229 "I Parry Everything S01E01 I Parry a Cow 1080p AMZN WEB-DL DDP2.0 H 264 DUAL-VARYG (Ore wa Subete wo \"Parry\" suru: Gyaku Kanchigai no Sekai Saikyou wa Boukensha ni Naritai, Dual-Audio)" [source-kotex]: https://nyaa.iss.one/view/1897423 "[KOTEX] Ore wa Subete wo Parry suru Scripts+Fonts for ReinForce [BD].zip" [source-erai]: https://nyaa.iss.one/view/1841170 "[Erai-raws] I Parry Everything - 01 [1080p][Multiple Subtitle] [ENG][FRE][GER]" [source-plus]: https://nyaa.iss.one/view/1880075 "[SubsPlus+] I Parry Everything S01 (WEB 1080p AVC EAC3) | Ore wa Subete wo Parry Suru (Batch)"