Blending modes (like Photoshop's Color/Luminosity) in Hybrid? / Histogram Preview

[mikazmaj]

Hi @Selur,
Is there an option in Hybrid to blend effects, similar to Photoshop’s layer blending modes (like Luminosity or Color)? Most major editors have this feature.
I would be particularly interested in a "Color" blend mode. For example, I'd like to run a model in the VS-MLRT plugin in a "Color" blend mode so that it doesn't affect the image's brightness/luminance.
Additionally, is there any way to display a histogram for each frame inside the VS Preview window?
Thanks!

[Selur]

Yes, you can set whether a filter (not resizer or deinterlacer or frame interpolation) affects only a specific plane by enabling "Filtering->Vapoursynth->Misc->Ui->Show 'Merge'-controls" and then use the 'Merge'-Controls next to the filter. (I think I mixed up 0 and 1 in the tool tip )

Additionally, is there any way to display a histogram for each frame inside the VS Preview window?

Filtering->Vapoursynth->Misc->Preview->Histogram

[mikazmaj]

Thanks for the advice. Quick question: Could you add a "Multi" option in the VS-MLRT plugin's filter section to enable multiple filters, similar to how it works in the Resize tab? Thanks!

[Selur]

For normal filters, use the filter queue instead of the filter order.
See: https://forum.selur.net/thread-1753.html

Cu Selur

[mikazmaj]

https://github.com/AmusementClub/vs-mlrt/issues/164

@Selur , could you take a look at this issue?
Right now, large ONNX models (over 2GB) only work correctly in vs-mlrt when they are first converted using TensorRT/RTX. If I want to load them for a quick preview with DirectML or CUDA onnxruntime, for example, it can't find the external data files (since the model consists of both .onnx and .onnx.data).
Also, could you implement flexible output for vs-mlrt? Multi-channel models are currently only allowed as inputs, while the output is restricted to 3 channels. It would be great to add a flexible output option to vs-mlrt.
Thanks!

[Selur]

Reading:
https://github.com/AmusementClub/vs-mlrt...4581956541
it sounds like: for non TRT/RTX backends adding a 'serialize' option, which when enabled, instead of:

clip = vsmlrt.inference([clip],network_path="PATH TO .onnx-Model", backend=Backend.XXX())

would use

import onnx onnx_model = onnx.load_model("PATH TO .onnx-Model"").SerializeToString() clip = vsmlrt.inference([clip],network_path=onnx_model, path_is_serialization=True, backend=Backend.XXX())

For TRT/RTX models, serialize would be ignored.

Is this correct?
Did you test this?
Does it actually work?

=> If you can confirm, that this is what should be done and that it works I can look at it tomorrow.

Also, could you implement flexible output for vs-mlrt? Multi-channel models are currently only allowed as inputs, while the output is restricted to 3 channels. It would be great to add a flexible output option to vs-mlrt.

Since I have no clue what "flexible output for vs-mlrt" means. Explain in details what should be done, what it is for and I can think adding support for it in Hybrid.
Otherwise: No, I can not implement something I have no clue about.


Cu Selur

Ps.: don't use old threads to ask unrelated questions.