When testing models I compare the upscaled to the original. being about to get the SSIM and PSNR values for one image compared to the other would be a useful tool and being able to use that in the node system would be super handy.

more of a stretch but if possible a VMAF node would be great too

We could do branching by using a special value representing a not-taken branch:

The main problem is that the current execution engine takes all possible paths. This is good and what we want it to do, but it makes branching (where we only want it to take some paths) difficult. So how about we use a special value, let's call it D, to represent a branch that is not taken?
To create a dead branch, a node simply has to output D in any of its outputs. Nodes would then handle D inputs as follows: if any of the required inputs is D, output D, and if an optional input is D, set the input to its default value. This is enough to implement branching (assuming no cycles).

This system would transparently implement branching. The D-input check would happen before actually running nodes, so our current nodes will just work in this system.

That being said, this only how to do branching. I don't have any good ideas as for how we could do the condition expressions.

For more information, see here.

The slider tracker being all the way on the left side should mean 100% A and 0% B.

I have found a bug where saving a workflow will just write an empty workflow to disk:

{"version":"0.5.2","content":{"nodes":[],"edges":[],"viewport":{"x":0,"y":0,"zoom":1}},"timestamp":"2022-04-27T17:36:16.425Z"}

I'm still trying to reproduce that bug and will make a separate issue for it.

In the meantime, please add a warning or make it impossible to save empty workflows. Saving empty workflows has no value anyway and would have made it easier to detect this bug.

Also, why is the JSON base64 encoded? Why not just save plain JSON and make the .cha files easy to debug?

Right now, there is no indication that the current chain has been saved successfully. Add a save notification (somewhere) so that the user knows that their chain has been saved (and where to).

As it is currently it appears you have to create your chain twice if you want to use it on a specific image and then decide to run a batch. Would be more intuitive and less work to have the iterator node removed and its functions integrated into the regular load image node for when you have a folder selected instead of a single image.
Or if it's less work to program just make the iterator node compatible with existing nodes instead of forcing it into its own empty window.

While I'm already posting I would like to add that implementing file drag&drop functionality to the load image and load model nodes would also be very useful.

I'm very new to upscaling, so apologies if this is a dumb question!

For cupscale, we have the option to use "RealESRGAN" or "ESRGAN" with NCNN. RealESRGAN is faster but only supports 4x models (not sure if it's a limitation of the implementation or RealESRGAN). Which one does chainner use?

If there is no cuda, do a check if dml is available and if it is then use that as the device

I chose to install with the portable install of chainner. I had hoped all dependencies would be installed under that portable folder but it doesn't seem to be the case (or I'm missing them).

Can someone confirm where those dependencies are? I know I can uninstall from the managed dependencies folder for some apps but python (integrated) cannot and I was curious where that was (I swear I alread had python installed).

Currently, every field in the Text Append node has to have something entered in order for the flowchart to run, which can be an issue when you only have two text inputs you want to append. Ideally, they should be able to be left blank.

Edit: Looks like this is actually implemented in utility_nodes.py, but it doesn't seem to be working as intended.

Is/will there be support for custom nodes?

I have a few things that I need chainner to support for it to be useful to me. I've already implemented these feature in a C# project of mine, but I can easily port them to any language.

Right now, the blur node is just broken.

1. It approximates Gaussian blur because the simple blur is iterated 16 times.
2. The amount input is not a radius (as is the case with Gaussian blur), but the kernel size. So if the amount is even, the kernel will move the whole image by one pixel (times 16).

Fixing either of these problems is a breaking change.

When I try to do an image upscale the upscaling fails and I see the following two things in the logfiles

Python subprocess exited with code null and signal SIGILL
kernel: [ 7750.737022] traps: python[176224] trap invalid opcode ip:7febea60213e sp:7feb82c920d0 error:0 in libtorch_cuda_cpp.so[7febe27a0000+61c96000]

So this seems to has something to do with my graphic card, which is a RTX3060, 12GB, because using CPU instead of GPU works.

I'm running this under Kubuntu 21.10.
Driver info: Driver Version: 470.103.01 CUDA Version: 11.4

If this is not directly related to chaiNNer, can you please add Information to the documentation, e.g. how to make pytorch to work with these Graphiccards?

I used the latest chaiNNer 0.4.3

Just like with File > New, show a little dialog that asks the user whether they really want to close without saving.

When PIL gets installed, old save files with Resize nodes break.

This is because the interpolation method input uses the literal interpolation method values of either PIL or OpenCV. The those values depend on what you have installed.

I'm currently working on a fix for this. Since the numerical values of PIL and OpenCV are incompatible with each other, I will choose the PIL values since most of our users likely have PIL installed.

AppImages run on all major Distributions, are portable, preventing QA issues from having different versions of Libraries on different systems.

You could drop .deb and .rpm support for it too, hopefully shortening your build times.

Some Relevant links:

• npmjs - AppImage - Electron Forge
• [appimage.org Electron Builder](https://appimage.org/
  https://www.electron.build/configuration/appimage.html)

Label ends of interpolation sliders for PyTorch and NCNN as 'Model A' and 'Model B'. Make sure functions match that ordering.

Examples:

The underlying image library we use does not support many DDS DXT10 formats.

This includes:

• ATI1 (legacy)
• ATI2 (legacy)
• BC1 (all variants)
• BC7 (all variants)

Right now, our custom alert have the problem that they can be undermined. E.g. the "Discard unsaved changes?" alert can be undermined by just opening a new file (see #219) or saving the current file. The same is also true for lots of other alerts.

The basic problem is that the user can still do other inputs by using the menu or shortcuts. This essentially bypasses the input lock imposed by the alert.

When you double-click or drag'n'drop a node from the node selector, the newly created node should be selected.

I don't think there is exactly an option to sharpen an image with OpenCV but the wiki shows a basic way to do Unsharp masking
https://en.wikipedia.org/wiki/Unsharp_masking#Digital_unsharp_masking

A note with a slider for sharpen amount would be great!

Remember window size on close so that the next time chaiNNer opens, it retains the same dimensions.

If I drag a node from the selection panel and drop it onto the canvas, it doesn't appear on the canvas.

I am running on a MacBook Air from 2020 (M1 / Apple aarch64, through Rosetta)

Video Recording

Right clicking a node should open a context menu similar to the ... button in the node footer, but at the point of right clicking.

This would be best accomplished via a context similar to the alert box context, where a single menu is used instead of multiple instances of it.

As the title says, they currently do nothing.

If I open chaiNNer on Windows, it will remain frozen until I either:

• Focus and unfocus the window
• Click anywhere else

Choose a better interpolation method than NN for the Resize nodes. Maybe cubic.

When saving a new workflow via Ctrl+S, the workflow won't be saved.

Steps to reproduce:

1. Make a new workflow
2. Add some nodes
3. Ctrl+S
4. Choose a file name and save
5. Open the saved file
6. All nodes will now be gone

This is because the saved file does not contain any information about the nodes. Example:

{"version":"0.5.2","content":{"nodes":[],"edges":[],"viewport":{"x":305.96171293776086,"y":-174.04899292324444,"zoom":0.5244057339865723}},"timestamp":"2022-04-27T17:50:34.429Z"}

This is something in electron. See docs here: https://www.electronjs.org/docs/latest/api/power-save-blocker#powersaveblockerstarttype

When closing chaiNNer, creating a new chain, or opening another chn file, pop up a prompt if there are unsaved changes to the current chain and ask whether to save.

As an AMD user, the automatic conversion from the pytorch model format to ncnn is really nice to have in cupscale. For now I can use that tool and use any converted models in chaiNNER but it would nice to support something similar natively in chaiNNER.

I know this was in the readme but I wanted to capture it as an issue so someone would be more likely to work on it :)

The node list is getting pretty long. Include a button to expand/collapse all node categories.

This is too good for us mortals

On Win10, the Nvidia 1080 Ti driver has a bug that causes all Electron applications to become blurry.

The bug has been present for 4 years now, and I lost all hope of it ever getting fixed. Luckily, there is a simple workaround. Disabling hardware acceleration for the Electron app fixes this issue. This workaround works for Slack, VS Code, and Discord, and it works for chaiNNer as well.

Please add an option in the settings for this, so I don't have to start chaiNNer from the command line with --disable-gpu anymore.

Right now, nodes only have a short summary of their function.

Add documentation and usage information for all nodes. The current idea is to have a little (i) icon next to the node name that opens a popup with the docs.

Add a small x to the search bar when it has some text to reset the text. Like this:

Trying to run chainner on Windows with Python 3.8.5 but I cannot get past this error.

11:15:12.966 > Attempting to spawn backend...
11:15:12.969 > Successfully spawned backend.
11:15:13.220 > Backend: No module named 'nodes'

11:15:13.221 > Backend: No module named 'nodes'

11:15:13.521 > Backend: No module named 'nodes'

11:15:13.521 > Backend: Traceback (most recent call last):
  File "F:\chainner\resources\backend\run.py", line 37, in <module>
    from nodes.node_factory import NodeFactory
ModuleNotFoundError: No module named 'nodes'

We often have issues where the devs need to know some system information and the current settings.

Add a little button that automatically collects all this info and copies it to the users clipboard, so the can send it to us.

This should include:

• All current settings
• OS
• CPU/GPU model + specs
• Environment variables

Installing NCNN gave an error output about restarting chainner. I did so but NCNN still wasn't installed. I tried again and again but still it wasn't installed.

Finally, I just let it sit and NCNN and the other dependencies were marked as installed. If chainner is installing, it'd be nice to have some indication and not get an error.

Running the Preview Image node would pop up a custom chakra window. This custom window would have the following features, as shown in the mockup below:

1. Window should be fully resizable.
2. Chain should pause while window is open to prevent window spam.
3. Preview Image should accept multiple image inputs (accept an iterable somehow in potential new iteration system?). All simultaneous inputs (so one set per iteration if chain is iterated) would be opened in different tabs of the Preview Window (as seen in the tab bar in the mockup).
4. The image displayed in the viewing window is swapped by clicking the appropriate tab or using the corresponding number key. Number keys would be assigned in order. If more than 10 images are permitted, could use key combos with the number keys to provide more hotkeys.
5. Scroll wheel or zoom buttons in toolbar (and maybe +/- keys) should zoom in and out of the images. Click and drag should pan the images.
6. Images should be scaled to the image with the largest dimensions (default nearest, but could allow other interpolations selectable from toolbar as well), as all tabbed images should share the same zoom scale and xy offset (that is, if you swap between different images, you should be seeing the same part of the image in the viewing window), as the primary purpose of this feature is to compare upscales of an image.
7. The Save button would pop up a save as window to save the currently selected image. Save All would save the images in all currently open tabs to a selected directory. Would need to figure out the best way to provide names for all of them, and possibly different behavior depending on whether file names are input to Preview Image. Could use a pattern like the GIMP Export all layers plugin (that plugin defaults to {image-name}-{layer-name}.png, just as an example), or save as the original input file name with the number appended, etc.
8. If possible: If preview is for a single image chain and not an iterator, have a 'Rerun current view' button that will rerun the chain on a cutout of the original image input based on the current section of the image visible in the viewing window, allowing chain to be rerun on partial images to compare specific sections only without having to process the full array.

Great for people that want to compile it themselves, contribute or package it for more platforms.

We recently had feedback where a user wanted to set their own defaults for the interpolation method of the resize nodes and the starting directory for image load/save and model load/save. This is definitely a useful feature, especially for power users.

Add settings for each node type that allow the user to override the defaults values/behavior of newly created nodes. Those settings should also be saved in localStorage like regular application settings. The setting should be displayed via a special menu that can be opened by clicking a gear icon on the node type in the node selector. The menu is also per node-type.

Note: These settings must only affect the default values and behaviors. The resulting chain must still be portable.

I ran into this issue with BoundedIntegerInput, but all other *IntegerInput classes also yield strings.

Why? The name says int, so why does it pass strings to the run method of nodes?

I went through all currently available nodes and all that use an *IntegerInput either:

1. Convert the string to int
2. Convert the string float
3. Let numpy convert it to float

So why does it pass strings?