Suggestions

[NEKolev]

It's great that you redesigned the interface and there are many new features from v1.4.

After I was able to launch version 1.5.1 on Windows, I have some suggestions and questions:

1. If you could add some videos or explanations with pictures of what works, because the interface has become quite complicated.
2. If an image is of low resolution, it stretches to the size of the screen, which I don't think is good.
3. It would be nice to add a translation option in the future. For example, via https://www.transifex.com/.

[pricklygorse]

Thanks for the feedback.

I plan to make some nice documentation when I've settled on the user interface and workflow, as I keep changing it. I agree its definitely a bit overloaded and clunky still. I hope the new workflow where the full generated and refined cutout is shown as a 'floating' preview on the output canvas which can be then added makes more sense than the previous workflow.

I can still in the meantime add some explanations and examples of the individual features though e.g. what the colour bleed correction does.

If the program gains a bit of popularity I can definitely look into translations, but for now that will be a future task.

[NEKolev]

A few more suggestions:

• if any object clipping or other action has been performed, there should be a warning when attempting to close whether the user is sure.

• the default format for saving "Quick Save" should be WebP or PNG. I suggest this from the perspective that JPEG does not support transparency.

• the clipped object should be saved in the middle of the image, and not analogously to the place from which you clipped it.

[pricklygorse]

1. Added, will be included in the next github build soon.

2. Quick save being JPG with white background is simply because it is what I need most frequently, not necessarily the best option. I'm not sure if there are enough save options to justify a dropdown of presets, but could add an additional "quick save png transparent" button, and/or set it to restore the last used settings when reopening the program.

3. This is by design. I made this program initially to fix and fine tune results from background removal models, so I could use a bit of one model, bit of another somewhere else, click a few areas with SAM etc. I'd then use the clipped image somewhere else (e.g. Inkscape, other image editors). Most of the open-weight models can't cut out my images without a couple of passes with other models/touching up areas. Now I've added effects to make the program closer to photoroom, maybe this workflow has become a little less apparent. There is a auto-crop export option. If it isnt working, its likely a model has added a stray pixel somewhere that the clean alpha function hasnt caught, which you can use the paintbrush to remove.

I've experimented with adding layers that can be moved but this confused the 'combine model output to make a perfect cutout' aspect. Photoroom uses a pop-out window to modify the model output, but they don't have a easy way to simply point and click on the original image to cut something out, or run a model on a specific area. So I'm still unsure about the best workflow

[NEKolev]

For me, saving in WebP (a better format than JPEG anyway) or PNG is the better option. Questions would arise why transparency, and the saved image has a white background.
In this case, I am experimenting with images of coins and cutting out a certain part of the coin. As you can see from the saved photo, the image is located exactly where it was in the space of the original image.

Interactive-BG-Remover_5xWuLxUbaI.mp4

It would be much more convenient to use one software, not several.

[pricklygorse]

Does the "Trim Transparent Pixels (Auto-Crop)" toggle on the export tab not work for this? It won't change the preview (other than darkening the bits that will be removed) but the exported image will be focused on just that part of the image.

If this behaviour was default, it would break workflows where removing backgrounds from, for example, a series of images of a rotating irregular shaped object on a turntable (one of my use cases). When stitching back together into a video the object would bounce all over the place as it won't be fixed in place in the frame.

[NEKolev]

Does the "Trim Transparent Pixels (Auto-Crop)" toggle on the export tab not work for this? It won't change the preview (other than darkening the bits that will be removed) but the exported image will be focused on just that part of the image.

If this behaviour was default, it would break workflows where removing backgrounds from, for example, a series of images of a rotating irregular shaped object on a turntable (one of my use cases). When stitching back together into a video the object would bounce all over the place as it won't be fixed in place in the frame.

Yes, with this option enabled, the cropped image was saved correctly.

[NEKolev]

A few more ideas:

1. It would be good to add the program version to the bar next to the name. For example, "Interactive Image... v1.6.0".
2. The program icon does not load in the Start bar in Windows 10/11.

3. I press the "m" key and the program displays a message that the module is missing. In my opinion, it would be more convenient to directly display the module download screen.

4. If the modules (for example, rembg) are updated in the repositories, is there a mechanism for updating them in the program?
5. A simple question. Do you have an idea to add, for example, image comparison and/or placing arrows, writing kexts, etc. on the images. It can also be as a separate program.

[pricklygorse]

1. Added

2. I still havent added an icon to the program, but I will at some point

3. I've added a prompt to open the settings menu to download models

4. I don't use the rembg library directly because how I cache the models in memory and other workflow differences. Some of the initial inference code was adapted from rembg however, and a few of the .onnx files are pulled from rembg repo. But all the models in the downloader are ones I've manually incorporated code for and have (mostly) tested. The main factor holding back incorporating more models is using onnx internally, not pytorch which nearly all models are released for. Converting formats doesnt always work, and I dont understand it well. I'm not sure adding a few more models is worth swapping to pytorch, which I think will also significantly increases the download size of the program releases.

5. I dont think at the moment I'll add too many more extra effects like text, arrows etc. There are so many programs that would do it far better than I can implement, like Inkscape, Illustrator etc.

[NEKolev]

1. Added

The version is visible when loading an image. Previously it was not visible. Maybe it was intended that way, but initially I thought it was not visible.
Here I forgot to suggest adding a check for a new version at startup and/or with a menu button.

Interactive-BG-Remover_18ymMw7mxs.mp4

4. I've added a prompt to open the settings menu to download models

When pressing the "I" button, the following error appears. I think the model download screen should have appeared here.

[ONNXRuntimeError] : 3 : NO_SUCHFILE : Load model from E:\backgroundremover-windows-cpu\Models\isnet-general-use.onnx failed:Load model E:\backgroundremover-windows-cpu\Models\isnet-general-use.onnx failed. File doesn't exist

6. I don't use the rembg library directly because how I cache the models in memory and other workflow differences. Some of the initial inference code was adapted from rembg however, and a few of the .onnx files are pulled from rembg repo. But all the models in the downloader are ones I've manually incorporated code for and have (mostly) tested. The main factor holding back incorporating more models is using onnx internally, not pytorch which nearly all models are released for. Converting formats doesnt always work, and I dont understand it well. I'm not sure adding a few more models is worth swapping to pytorch, which I think will also significantly increases the download size of the program releases.

Maybe I didn't express myself clearly. When a model is downloaded, is it updated over time by its creator or does it remain as it was created by its creator and if it is updated, will it be automatically updated by IIBR.

8. I dont think at the moment I'll add too many more extra effects like text, arrows etc. There are so many programs that would do it far better than I can implement, like Inkscape, Illustrator etc.

Too bad. I think it would have been a great addition :)

[pricklygorse]

Pretty much every model stays as is once released. Some have minor updates, like SAM 2 getting a 2.1 version, and some getting major updates (rmbg1.4 -> rmbg2). I'll add them as I find them in a compatible format. The current models however, paired with a point and click Segment Anything model and/or paintbrush touchup work perfectly for anything I've been able to throw at it so far.

There is a risk of there being so many models that it becomes overwhelming if I could add all the ones I know of, especially compared to paid solutions like photoroom that simply one-shot most things pretty well.

[NEKolev]

• if any object clipping or other action has been performed, there should be a warning when attempting to close whether the user is sure.

A warning is now displayed when closing, but when saving with Quick Save, Export Final Image and closing the program without making any other changes, there may be no need to ask - this way the user may be left with the impression that some change was made after saving, when in fact it was not.