Vship Moving From Github

Here is the new release on codeberg: https://codeberg.org/Line-fr/Vship/releases

From Now on, the github repository will not receive any new commit or release. but files will remain there for a little more time.

I chose to finally go away from microslop github to avoid feeding their AI in their right with the terms of service of the platform.

VSHIP 4.0.2 : Bugfixes - HOTFIX 15/12/2025 15:58

Hi, here is a new minor version because it seems sable and bring some little bugfixes

changes:

• libvship exception system now allows to retrieve a more detailed error which is include in FFVship
• libvship now supports unaligned dimensions to the chroma subsampling (example: odd width with YUV422)
• libvship supports stride unaligned with the memory data type (example: 16 bits uint and odd stride)

bugfix:

• fix a bug in butter/ssimu2 and cvvdp that existed since the beginning for dimensions not multiple of 16
• no wild error messages in console when a hip internal error happens now, there is a dedicated libvship error

HOTFIX 15:58 : Vapoursynth Plugin had crop value uninitialized while using libvship which caused tons of issues.

I believe that this version is the most stable vship version. At least that is what I hope
So you can enjoy 🧈 without getting obese now!
Line

VSHIP 4.0.1 : BugFix, new api function and ffvship uses libvship API

Hi, a new minor release to stabilize a few things since 4.0.0

First, FFVship does not possess internally the whole libvship codebase duplicated. It now uses libvship.dll/so. To remain easy to use for windows users, the libvship.dll is still duplicated in the release by being inside the FFVship.zip
I created a testing binary which allowed me to find 3 bugs in test directory 🥳

Change:

• the precision of the values in the json from FFVship is greatly increased
• --verbose can display the colorspace detected by FFVship in the input videos

New API calls:

• Vship_ResetScoreCVVDP -> reset the internal score without resetting the temporal filter. This can be useful to implement scene scores for a clip instead of getting only a score for the whole clip
• Vship_LoadTemporalCVVDP -> allows to put frames inside the temporal filter without computing the score.

Bug Fixes:

• unsupported pixel format did not make FFVship crash
• fix incompatiblity of the local pixfmt.h and the one from the ffmpeg in the system by not having a local pixfmt.h
• Chroma 440 works now
• stride misaligned on sample type now works
• error when freeing handlers has been resolved

VSHIP 4.0.0 : Color conversion on GPU and CVVDP Metric!

The number of commit between last release and this one is 25% of the whole total commits to Vship. This release comes with some fixes bug will also potentially introduce new bugs on obscure videos that I have not tried. Don't hesitate to report them in Issues if you see one! I will probably be able to fix them without too much trouble.

• /!\ Breaking changes

  • Vship API changed a lot because vship handles color conversion internally now which means it is easier to integrate to a project but required breaking changes to the API which previously had ugly "-Uint16 float" functions for the 2 types. This is over and represented by a unified structure
  • vship.so/dll is now renamed as libvship.so/dll This is better for linking vship API and more coherent with what vship is today. This can create issues when installing the new version on Linux with conflicting dll. To uninstall the previous version before installing the new one you can use make uninstallOld
  • Some nice and weird colorspaces which used to work in FFVship (thanks to zimg color conversion) will not work anymore. But if you encounter such case, don't hesitate to create an issue. The code structure makes it easy to add a new supported colorspace. The list of supported colorspace can be seen in src/VshipColor.h

• New in Vship4

  • CVVDP Metric! I implemented https://github.com/gfxdisp/ColorVideoVDP in Vship. It is up to 8x faster than the original and fixes a lot of their issues (especially for the fact that they do not parse the metadata of the source). Where SSIMULACRA2 tries measures how bad artifacts are and butteraugli tries to measure the distance required to differentiate 2 sources, CVVDP tries to see the percentage of people who would select the lower quality source. CVVDP scores impressively high on independent benchmark research papers. CVVDP is also the first actually temporal Metric we get to see in Vship which means a single score per video is returned unlike others.
  • Butteraugli QNorm -> you can specify (--qnorm in FFVship) the norm used by butteraugli to allow arbitrarily strict severity on bad pixels for the resulting score.
  • No Zimg dependency -> Thanks to GPU Color conversion, we can remove this dependency. However, on GPU bottlenecked situation, you might see a slowdown in SSIMU2 computation speed but a lower CPU usage since the load of the color conversion is on GPU instead of CPU now.
  • Cropping support in FFVship. There are new options (check FFVship --help) that allows to crop the source or the encoded video easily

• BugFix

  • bad aligned memory for zimg on windows making random crashes
  • default transferFunction is sRGB when having an RGB Matrix
  • fix vapoursynth crash when cropping before passing to vship

This clearly is the biggest release I did so far and it was a lot of work but I am glad it is done now. I hope not too many bugs will have stepped in silently and I hope you will like this new CVVDP!
Happy christmas in advance ahah
Line

VSHIP 3.1.0 : FFVship fixes and extended support + Vship C API

Hi, it has been a while since the last release.

Overall news:

• the wiki has been moved to the doc folder for allowing PR on it and ...
• Vship C API now allows low level usage of Vship's internal metric system for other project to use efficiently. (more in the doc)
• Fixed wrong Transfer Function from being used in Vship and FFVship

FFVship news:

• You can cache the video indexes used by FFMS2 to speed up FFVship if comparing to the same files over and over using --cache-index
• bugfix: some segfault appeared in video having a width that is not multiple of 16.
• FFVship works on most images now, supporting a wider range of formats like full range YUV
• bugfix: the progress bar will now go to the end all the time.
• bugfix: errors will properly stop the computation and return error while writing in stderr instead of stdout.

Vship news:

• Breaking: vship.GpuInfo(gpu_id) got clockRate info removed because the information became unavailable in cuda 13
• bugfix: bad default metadata were being used on HDR content making the result completely delirious.

That was quite a long list for quite a long time since last release! I hope the API will prove useful and that Vship will continue to improve accessibility to better metrics.
Thank you, Line

VSHIP 3.0.2 PreRelease : Butter Speed + FFVship new options

Hi, Thank you for following the project

This version comes with a little butter speed boost and new options making FFVship possible to integrate within very specific scripts that would otherwise have required the vapoursynth plugin: --source-indices and --encode-indices. They are still experimental and I'd like to have your feedbacks on these features to eventually improve them.
Another option for scripters is --live-score-output to replace the stdout with scores for realtime score retrieval
You can now get FFVship version with --version
If your commandlines are too long because of indices, you can also use --config-file to extend CLI arguments using a file

Fixes: --list-gpu is fixed (doesnt require source and encoded file) and error code is properly set.

I am happy that this release is not only about FFVship arguments but also a little bit of 🧈❤️
That is all for today!

VSHIP 3.0.1 : FFVship improved a lot!

Hi and thank you for still being there to support vship!

@GreatValueCreamSoda Completely reworked the threading system of FFVship allowing a very impressive RAM reduction and speed increase on a lot of system. On top of that, I re-added -t argument to boost back systems that had been seeing slowdown from this rework. Default value is 1 so feel free to experiment with it.

FFVship now has a progress bar for users to know if the process is hanging or not.

With positional arguments, FFVship is more convenient to use than ever: FFVship path1 path2 is well enough to compute SSIMU2

To end all that for developper friends willing to compile FFVship on windows, I finally create a wiki page to do it flawlessly.

That is all for today, I wish you a good SSIMU2 night.
Line

VSHIP 3.0.0 : FFVship CLI Tool + SSIMULACRA2 VRAM Reduction

This release is the conclusion of a lot of hard work and I hope it will help making SSIMULACRA2 and Butteraugli more accessible.

• This update introduces the first version of FFVship ! Getting rid of bottlenecks rising from using vapoursynth, it is able to compute SSIMU2 much faster. Also, being Vapoursynth free, it is attractive for people unwilling to installing this massive dependency. Scores are little different but it is normal and will see potential improvments later. It also is for now only able to open videos, not images.

• /!\ Change that can impact your existing workflow: Default Intensity target in butteraugli is now 203. If you wish to go back to 80, you can set by yourself the option to 80.

• SSIMULACRA2 got a VRAM reduction update that also made it faster, you can expect about 1/3 VRAM usage compared to before. It also got debugged to work with huge images tested with up to 400 MégaPixels!

Upcoming changes for future releases are image support for FFVship and color conversion done entirely on GPU.

Thank you for your continuous support <3
I wish you a nice 🧈 cooking session.
Line

VSHIP 2.2.2 : Big Butteraugli speedup + some fixes

up to now, I mainly worked on speeding up SSIMU2 so there was a good margin of time to gain on butteraugli. Expect from 1.2x to 1.4x speedup

I also fixed vship when it received < -0.04 srgb values (I was not expecting negative inputs)
the downsample in butteraugli is now done after transition to linear RGB.

That is all for today!

VSHIP 2.2.1 : Important Speedup for non ROCm 6.3 users + numStream option

NVIDIA Users and < 6.3 ROCm users can expect up to 1.5x performance for ssimu2 with this update. Butter also improved but marginally.

It is now also possible to have a different number of stream used inside vship (which control VRAM consumption) than the number of vapoursynth threads (used for decoding). You can use the option "numStream" to set that.
(In general, stream managment was improved but it is pretty much invisible to the user)

Have a nice time computing your favorite metrics!