bryced

Asked this and related questions to o3. I do not vouch for the answers at all but you may find it interesting. https://chatgpt.com/share/6876cdd1-dfbc-8011-a55f-6915a90275...

There is a smart mask feature. Add `--mask-prompt face --mask-mode keep`. I believe it outputs the masks as well

I believe you can just ignore that error. The cell ran.

Fair but I'd point out I also didn't make the algorithm that changes photos. I'm wrapping a bunch of algorithms that other people made in a way that makes them easy to use.

It's not just that GUI's are hard, it's that the "customer" base will inevitably be much less technical and I'd receive a lot more difficult to resolve bug reports. So no-gui is also a way of staying focused on more interesting parts of the project.

> would it be possible to use this tool to make automatic mask for editing in something like GIMP

probably but GIMP plugins are not something I've looked into

> REPL

already done. just type `aimg` and you're good to go

> GUI

GUIs add a lot of complexity. Can your file manager do thumbnails and quick previews?

I'm told Imaginairy secretly does run on Intel Mac. very slowly. I just don't want to be on the hook for support so the reqs are written that way.

Oh I see what you're saying. It's not actually downloading but there is a bug where it's not using the cache properly. will fix

If you're wanting to use Stable Diffusion 2.1 with imaginairy you just specify the model with `--model SD-2.1` when running the `aimg imagine` command.

This is itself it's own finetuned version of SD so now it won't work with alternative versions. img2img works by just running normal stable diffusion img2img on a noised starting image. As such it destroys information at all parts of the image equally. This new model uses attention mechanisms to decide which parts of the image are important to modify. It can leave parts of the image untouched while making drastic changes to other parts.

The console output tells you what happened.

Does dalle do prompt based photo edits now?

But yeah sometimes it doesn't follow directions well. I haven't noticed a pattern yet for why that is.

Here is a google colab you can try it in: https://colab.research.google.com/drive/1rOvQNs0Cmn_yU1bKWjC...

Here is a colab you can try it in. It crashed for me the first time but worked the second time. https://colab.research.google.com/drive/1rOvQNs0Cmn_yU1bKWjC...

It does work in non-pro colab apparently. Here you go: https://colab.research.google.com/drive/1rOvQNs0Cmn_yU1bKWjC...

If you're used to installing python packages it should be relatively easy. There are other projects with nice UIs but that's not what this library is for.

yes stable diffusion is not great about handling multiple ideas. New image models coming out soon though.

That is strange. I'm not sure what would cause that unless it was running in some ephemeral environment. What OS? Can you open a github issue with a screenshot?

I tried it out :-)

`aimg edit assets/girl_with_a_pearl_earring.jpg "make it pop" --prompt-strength 40 --gif`

https://user-images.githubusercontent.com/1217531/213912442-...

I'm running on a 2080 TI and an edit runs in 2 seconds. On my Apple M1 Max 32Gb edits take about 60 seconds.

dreambooth perhaps?

Yes if you're not careful it can ruin the face. You can play with the strength factor to see if something can be worked out. Bigger faces are safer.

I did the work to wrap it up and make it "easy" to install. The researchers who did the real work can be found here: https://www.timothybrooks.com/instruct-pix2pix

Shouldn't be downloads per prompt. Processing happens on your machine. It does download models as needed. A network call per prompt would be a bug.

I'm not aware of any pre-existing open-source model that selectively edits images (leaving some parts untouched) based on instructions. This new method is much better than the image2image that shipped with the original stable diffusion. I'm looking at the InvokeAI docs right now and don't see anything like this feature. We previously had smart-masks, but InstructPix2Pix mostly does away with the need for those as well.

If I am mistaken please provide links to these prior features.

Been doing a lot with prompts lately. What people are calling "prompt engineering" I'd call "knowing what to even ask for and also asking for it in a clear manner". That was a valuable skill before computers and will continue to be one as AI progresses.

I've been pretty disappointed to introduce ChatGPT to people in jobs where it would be a game changer and they just don't know what to do with it. They ask it for not-useful things or useful things in a non-productive way. "here is some ad copy I wrote, write it better". Whether you're instructing a human, chatgpt, or AI god... that's just too vague of instructions.

For the specific example you provide you could also use a prompt-based mask to prevent it from editing the person.

no. in imaginairy it's called `--prompt-strength`. In other libraries it's called CFG or "classifier-free guidance". For the image edits I vary the strength of the effect from between 3-25

Thanks Paul!

Works fine for me, you just need to adjust the strength of the edit.