Best Android app that can run generative AI locally?

gamma7

I've tried PocketPal, ChatterUI and Google Edge gallery, but all of the models I've downloaded and tried are seemingly text generators. I've also tried downloading Any-to-Any models from Hugging face and loading those into apps, but they still either answer that they can't generate images, or start writing weird links and . Am I doing something wrong, or do I need to find better alternatives? If the latter, I'd prefer not having to get them from Google Play, unless it's the only option.

Pocketstar

gamma7 I run SDAI (FOSS) locally.
https://github.com/ShiftHackZ/Stable-Diffusion-Android
It can generate up to 512x512 resolution images.
For local diffusion, it uses Microsoft ONNX and Google AI Mediapipe. (You can choose between those)
There a quite a few options to choose from, however the results and quality vary greatly. I only recommend it for low-quality works, but I'm happy with it myself.
I noticed it can throw an error when generating if installed on secondary user profiles.
Suitable models can be downloaded here:
https://github.com/ShiftHackZ/Local-Diffusion-Models-SDAI-ONXX
and here:
https://github.com/ShiftHackZ/Local-Diffusion-Models-SDAI-MediaPipe

The app ChatterUI can be used for image recognition by using a secondary image recognition model. (It does not generate images)

gamma7

Pocketstar SDAI FOSS can only generate images, and it hasn't been updated in over a year. I'm looking for something that can utilize Any-to-Any models, not just text-to-image and image-to-image. Unless there's no such thing on android that can do it locally?

Gouser

gamma7 I don't think mobile devices are powerful enough yet to do that locally. I don't know of any FOSS project that is able to do this.

Pocketstar Do you happen to have the signing hash of SDAI (FOSS) github version? I want to verify it with Appverifier after installing through obtainium but I can't find it on github or in the GrapheneOS comparing hashes post

Gouser

Pocketstar How did you load the Mediapipe models in local diffusion? The app and documentation only talks about ONNX. The ONNX packs have a proper directory structure but the Mediapipe models are a pack of files without any structure

Pocketstar

gamma7 Indeed, the last update seems over a year old, I honestly hadn't noticed that yet, time flies... However I believe it is not marked as archived or discontinued yet.

Pocketstar

Gouser Do you happen to have the signing hash of SDAI (FOSS) github version? I want to verify it with Appverifier after installing through obtainium but I can't find it on github or in the GrapheneOS comparing hashes post

Certainly! Here are both hashes, the FOSS and the full versions: (Version 0.6.8)

com.shifthackz.aisdv1.app.foss
89:74:EF:C8:53:E2:D5:E9:3E:B2:9D:D8:09:FC:57:CA:E7:41:1F:CF:AC:A0:3D:E3:F7:C1:52:F4:41:A9:00:64

com.shifthackz.aisdv1.app.full
89:74:EF:C8:53:E2:D5:E9:3E:B2:9D:D8:09:FC:57:CA:E7:41:1F:CF:AC:A0:3D:E3:F7:C1:52:F4:41:A9:00:64

gamma7

Gouser I don't think mobile devices are powerful enough yet to do that locally. I don't know of any FOSS project that is able to do this.

In that case, PC should be able to. Only kobold.cpp can do generation AFAIK but it's not available as Flatpak so it can't be sandboxed. Jan AI and GenAI are both broken and fail to install any model the last time I tried them.

Pocketstar

Pocketstar By the way, I have obtained these hashes by using AppVerifier, I haven't seen them anywhere else, so I'm not sure if it is a good idea to use as a comparison method.

Here is some other information obtained from my own installation: (foss and full version)

foss:
MD5sum: 8a392a3e69bf835a9249ea14ec16386e
SHA1sum: b2d4340a9a06d25c061cfc7dbe2edb74932298d4
SHA256sum: 9b910433145a208c5a3cb56538fdbf917ed661ff0ac5821109ebd8849629527a

C=UA,ST=Dnipropertrovska oblast,L=Dnipro,CN=Dmitriy Moroz

SHA256withRSA

CERTIFICATE fingerprints:
md5: 8dd1006a97e63a65e9de727a7bec887a
sha1: 74a6e5d6b5859ce4d766bd7377e3069c7a3ddd9f
sha256: 8974efc853e2d5e93eb29dd809fc57cae7411fcfaca03de3f7c152f441a90064

full:
MD5sum: fc47b52ef9b9a66e74fc5d8083156e50
SHA1sum: 564e53474786b76bc666e321d1af2baa134d4147
SHA256sum: 5837c334aaa6c4189fc3ca1d1ad3bb333d8da08c63a035aa76ff96fb3ef370f5

C=UA,ST=Dnipropertrovska oblast,L=Dnipro,CN=Dmitriy Moroz

SHA256withRSA

CERTIFICATE fingerprints:
md5: 8dd1006a97e63a65e9de727a7bec887a
sha1: 74a6e5d6b5859ce4d766bd7377e3069c7a3ddd9f
sha256: 8974efc853e2d5e93eb29dd809fc57cae7411fcfaca03de3f7c152f441a90064

Gouser

Pocketstar if the author of the app doesn't post the signatures the only way to verify it is by sharing it and let others check if it matches theirs. If nobody says they have different signatures when installing it then it should be fine. You still have to trust the developer but the amount of stars on the project (people following the code) and a matching signature is as good as safe as you can be.

com.shifthackz.aisdv1.app.foss
89:74:EF:C8:53:E2:D5:E9:3E:B2:9D:D8:09:FC:57:CA:E7:41:1F:CF:AC:A0:3D:E3:F7:C1:52:F4:41:A9:00:64

Matches with the signature you shared when I install the FOSS version. Thank you. If you want you can add it to the compare hashes post or I can add it

Pocketstar

Gouser If you want you can add it to the compare hashes post or I can add it

Thank you for the information!
Well, considering I am still quite unfamilliar with AppVerifier, you are welcome to add it. I am a total n00b with such matters.
However it is very important to verify the hashes, so I will do some study towards it. (I should've done that earlier)

Pocketstar

Gouser How did you load the Mediapipe models in local diffusion? The app and documentation only talks about ONNX. The ONNX packs have a proper directory structure but the Mediapipe models are a pack of files without any structure

The same way as the ONNX model; Make a directory in the Downloads folder, name it "SDAI", within that, make a folder called Local-Diffusion-Models-SDAI-MediaPipe and place its models in there, like stable-diffusion-v1-5.
The folder structure will be as follows;

Downloads/SDAI/Local-Diffusion-Models-SDAI-MediaPipe/stable-diffusion-v1-5/-all of the model files.- (like "cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc1.weight.bin" and all the other files)

If you want a second model, (like aniverse) you can make the following structure;

Downloads/SDAI/Local-Diffusion-Models-SDAI-MediaPipe/aniverse_v50/-model files.-

Same thing goes for ONNX;

Downloads/SDAI/Local-Diffusion-Models-SDAI-ONXX/-model files and folders.-

After that, set the model path for Mediapipe within the SDAI configuration as follows;
/storage/emulated/0/Download/SDAI/Local-Diffusion-Models-SDAI-MediaPipe/stable-diffusion-v1-5

For aniverse:
storage/emulated/0/Download/SDAI/Local-Diffusion-Models-SDAI-MediaPipe/aniverse_v50/

(Choose only one model at a time)

And for ONNX;
/storage/emulated/0/Download/SDAI/Local-Diffusion-Models-SDAI-ONXX

Gouser

Pocketstar I will give it a try, thank you.

Pocketstar You can find a topic on comparing hashes on this forum. I can often find a hash I'm looking for there. I will add this one to the list.

Pocketstar

Gouser You can find a topic on comparing hashes on this forum. I can often find a hash I'm looking for there. I will add this one to the list.

Awesome, thank you for your assistance! I shall give it a go.