NVDA 64-bit with on-device image description alpha build now available

220 views
Skip to first unread message

Quentin Christensen

unread,
Oct 3, 2025, 3:29:38 AM (2 days ago) Oct 3
to NVDA Screen Reader Discussion
Hi everyone,

We're bringing on-device image description to NVDA! More on that further down.

As you may know, up until now, NVDA has been a 32-bit program. We are working towards NVDA 2026.1 being 64-bit.  Almost all consumer PCs made in the last decade or so are 64-bit and Windows 11 only runs on 64-bit processors.

We've hit a milestone today - nvda_snapshot_alpha-52930,7778bb6c.exe now available, and is a 64-bit program! You can test it out from: https://download.nvaccess.org/snapshots/alpha/

Note, it IS still an alpha build, so there may be issues and bugs - but we wanted to let you know as we know a number of people are interested in trying it out. And we wanted to let you know a couple of caveats with it:

1) If you install a 64-bit build of NVDA, you cannot safely downgrade to 2025.3 or earlier.  You will need to uninstall NVDA and re-install the older version.
2) Windows 10 is now the minimum Windows version supported.
3) It won't work on ARM devices running Windows 10.
4) It won't work if you have a 32-bit version of Windows
5) This version is a "2026.1" alpha, so existing add-ons won't work.
6) When you upgrade, you may lose the secure desktop configuration (NVDA's configuration it uses on the logon screen) which may revert to the default config.

Now, to the exciting part: On-device image description!

Press NVDA+Windows+, to get an AI generated image description. This is generated locally on the device - no information is sent to the internet.

To use these image descriptions, NVDA will need to download approximately 230MB of data from the internet for the model the first time you enable it.

To enable AI generated image descriptions:
1. Press NVDA+control+g to open NVDA's general settings
2. Press CONTROL+TAB to get down to "AI Image Descriptions"
3. Press TAB to "Enable image captioner".
4. Press SPACEBAR to enable this option.
5. Press ENTER to close settings.
6. NVDA will prompt to download the data it needs. Press ENTER to allow this.

The data is downloaded (there is no progress bar, we are aware of that) and NVDA advises when it is done. After that, you can start using the image captioner.

Note that this ONLY gives English descriptions currently. Our plan is to bring internationalisation to this as soon as possible.

There are other exciting updates already in this alpha for 2026.1 you can check out:
- Report when multiple items can be selected in a list control
- The status bar can be reported in VS Code
- Performance improvements on ARM64
- Spelling errors can be reported with a sound or Braille rather than speech
- VirusTotal scan results are available in the details for add-ons in the add-on store
- In the add-on store you can see the latest changes for add-ons.

If you do test out this build, please do let us know how you find it. And don't forget to file any bugs you find on GitHub: https://github.com/nvaccess/nvda/issues

Kind regards

Quentin

--

Quentin Christensen
Training and Support Manager

NV Access

Subscribe to email updates (blog, new versions, etc): https://eepurl.com/iuVyjo

mattias

unread,
Oct 3, 2025, 6:01:34 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org

And wich key are ”windows?

I tryed with nvda + my left windows key but no

 

Sent from Mail for Windows

--
***
Please note: the NVDA project has a Citizen and Contributor Code of Conduct.
NV Access expects that all community members will read and abide by the rules set out in this document while participating in this group.
https://github.com/nvaccess/nvda/blob/master/CODE_OF_CONDUCT.md
 
You can contact the group owners and moderators via nvda-user...@nvaccess.org.
---
You received this message because you are subscribed to the Google Groups "NVDA Screen Reader Discussion" group.
To unsubscribe from this group and stop receiving emails from it, send an email to nvda-users+...@nvaccess.org.
To view this discussion visit https://groups.google.com/a/nvaccess.org/d/msgid/nvda-users/CAKsDpFgDt0%2BSQgyxZJR2uHVMBsWNkwd1GLN3ws5ZL2v_jKhNUw%40mail.gmail.com.

 

Quentin Christensen

unread,
Oct 3, 2025, 6:06:22 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org
Sorry Mattias,

The correct keystroke is NVDA+Windows+comma  (I copied the details from the release notes - which has the comma symbol - normally I try to write out symbols in words, to avoid it being lost just as it was here).

Also, to answer another question someone asked - yes you can setup this alpha as a portable copy of NVDA.  I got caught up with the caveats for installing it, that I forgot to mention the way many will do it - setting it up as a portable copy - and yes that should work as any other portable copy of NVDA.

Kind regards
Quentin

mattias

unread,
Oct 3, 2025, 6:11:04 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org

Oops my bad

My nvda dont reed special characters like comma

I have it turned off

Or special characters..

I have forgot the English Word for it

Fettah Pınar

unread,
Oct 3, 2025, 6:17:21 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org
Hello Quentin!
First of all, thank you so much for sharing this wonderful development with us. The innovations are quite exciting and impressive.
I would like to ask you a question that's been on my mind.
As I understand it, image recognition will work entirely within the device and will not require the internet.
My question is this: won't this significantly impact the quality of image recognition?  
I mean, isn't the accuracy of image recognition higher when done online?  
Which AI model are you using for image recognition, and what is the accuracy rate?

Quentin Christensen <que...@nvaccess.org>, 3 Eki 2025 Cum, 10:29 tarihinde şunu yazdı:
--

Quentin Christensen

unread,
Oct 3, 2025, 6:17:49 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org
Not at all, that's exactly why I usually write the word out.  Sorry for neglecting it!

Johann Tan

unread,
Oct 3, 2025, 6:45:08 AM (2 days ago) Oct 3
to NVDA Screen Reader Discussion, Quentin Christensen
Hi.
Is there a way for when running this alha to not have it read my current nvda config when launching the installer, because I'm afraid it will do something to my current NVDA config that would possibly corrupt it.

Gene Asner

unread,
Oct 3, 2025, 7:00:43 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org
You give a command to get descriptions but how do you tell the feature
what image or images you want descriptions of? Does it give
descriptions of all images on a page?

Gene

Mujtaba Merchant

unread,
Oct 3, 2025, 8:10:50 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org

Hi Quentin,

 

Thanks for sharing this exciting update about the new NVDA alpha build. I'm thrilled to hear about the on-device image description feature, which is going to be a game-changer for users.

 

I do have one question regarding the shortcut key – NVDA + Windows +comma to activate On Device Image Description. If someone is using the Desktop layout for their keyboard, using this shortcut might be a bit of a stretch (literally!). I tried it and my fingers felt like Spiderman trying to web-sling across the keyboard!

 

 

 

 

From: Quentin Christensen <que...@nvaccess.org>
Sent: Friday, October 3, 2025 12:59 PM
To: NVDA Screen Reader Discussion <nvda-...@nvaccess.org>
Subject: [NVDA] NVDA 64-bit with on-device image description alpha build now available

 

Hi everyone,

--

Gene Asner

unread,
Oct 3, 2025, 8:25:09 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org
It doesn't matter what keyboard configuration you are using. Set
capslock to be an NVdA key, you can change the setting in the keyboard
settings dialog and you can then use caps lock, Windows key with the
left and and type comma with the right. You can keep insert as another
NVDA key. Just set caps lock as an NVDA key and leave the insert
setting as it is.

Gene

On 10/3/2025 7:10 AM, Mujtaba Merchant wrote:
> Hi Quentin,
>
> Thanks for sharing this exciting update about the new NVDA alpha build.
> I'm thrilled to hear about the on-device image description feature,
> which is going to be a game-changer for users.
>
> I do have one question regarding the shortcut key – NVDA + Windows
> +comma to activate On Device Image Description. If someone is using the
> Desktop layout for their keyboard, using this shortcut might be a bit of
> a stretch (literally!). I tried it and my fingers felt like Spiderman
> trying to web-sling across the keyboard!
>
> *From:*Quentin Christensen <que...@nvaccess.org>
> *Sent:* Friday, October 3, 2025 12:59 PM
> *To:* NVDA Screen Reader Discussion <nvda-...@nvaccess.org>
> *Subject:* [NVDA] NVDA 64-bit with on-device image description alpha
> www.nvaccess.org <http://www.nvaccess.org>
>
> Training: https://www.nvaccess.org/shop/ <https://www.nvaccess.org/shop/>
>
> Certification: https://certification.nvaccess.org/
> <https://certification.nvaccess.org/>
>
> Subscribe to email updates (blog, new versions, etc):
> https://eepurl.com/iuVyjo <https://eepurl.com/iuVyjo>
>
> User group: _https://groups.google.com/a/nvaccess.org/g/nvda-users
> <https://groups.google.com/a/nvaccess.org/g/nvda-users>_
>
> Other groups: https://github.com/nvaccess/nvda/wiki/Connect
> <https://github.com/nvaccess/nvda/wiki/Connect>
>
> Mastodon: https://fosstodon.org/@NVAccess <https://fosstodon.org/@NVAccess>
>
> Facebook: http://www.facebook.com/NVAccess
> <http://www.facebook.com/NVAccess>
>
> --
> ***
> Please note: the NVDA project has a Citizen and Contributor Code of Conduct.
> NV Access expects that all community members will read and abide by the
> rules set out in this document while participating in this group.
> https://github.com/nvaccess/nvda/blob/master/CODE_OF_CONDUCT.md
> <https://github.com/nvaccess/nvda/blob/master/CODE_OF_CONDUCT.md>
>
> You can contact the group owners and moderators via
> nvda-user...@nvaccess.org <mailto:nvda-user...@nvaccess.org>.
> ---
> You received this message because you are subscribed to the Google
> Groups "NVDA Screen Reader Discussion" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to nvda-users+...@nvaccess.org
> <mailto:nvda-users+...@nvaccess.org>.
> <https://groups.google.com/a/nvaccess.org/d/msgid/nvda-users/CAKsDpFgDt0%2BSQgyxZJR2uHVMBsWNkwd1GLN3ws5ZL2v_jKhNUw%40mail.gmail.com?utm_medium=email&utm_source=footer>.
>
> --
> ***
> Please note: the NVDA project has a Citizen and Contributor Code of Conduct.
> NV Access expects that all community members will read and abide by the
> rules set out in this document while participating in this group.
> https://github.com/nvaccess/nvda/blob/master/CODE_OF_CONDUCT.md
> <https://github.com/nvaccess/nvda/blob/master/CODE_OF_CONDUCT.md>
>
> You can contact the group owners and moderators via
> nvda-user...@nvaccess.org.
> ---
> You received this message because you are subscribed to the Google
> Groups "NVDA Screen Reader Discussion" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to nvda-users+...@nvaccess.org
> <mailto:nvda-users+...@nvaccess.org>.
> To view this discussion visit
> https://groups.google.com/a/nvaccess.org/d/msgid/nvda-users/013e01dc345e%24bef28fd0%243cd7af70%24%40gmail.com
> <https://groups.google.com/a/nvaccess.org/d/msgid/nvda-users/013e01dc345e%24bef28fd0%243cd7af70%24%40gmail.com?utm_medium=email&utm_source=footer>.

Mujtaba Merchant

unread,
Oct 3, 2025, 8:28:58 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org
Thanks for the tip about setting the capslock as an NVDA key to make the shortcut easier to use. That way, I can use caps lock with the Windows key and comma without having to stretch my fingers too much. I'll definitely try that out. Your suggestion makes it much more manageable. Thanks again for sharing this helpful advice.
To unsubscribe from this group and stop receiving emails from it, send an email to nvda-users+...@nvaccess.org.
To view this discussion visit https://groups.google.com/a/nvaccess.org/d/msgid/nvda-users/7ebbc2f4-0aa0-b586-6805-1b66ba138ce6%40gmail.com.

muhammed ali çiçek

unread,
Oct 3, 2025, 9:09:32 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org
"Hello, I haven't tried it yet, but I think a small, open-source language model is being used. Regarding the quality of the descriptions, I assume it won't be able to provide the exact descriptions we want in the first versions. If you want a better description, try the Cloud Vision add-on. You're probably already using it, but I wanted to mention it.
Best regards."

Fettah Pınar <fettah...@gmail.com>, 3 Eki 2025 Cum, 13:22 tarihinde şunu yazdı:

Christo de Klerk

unread,
Oct 3, 2025, 10:01:53 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org

Hello


I have just installed the Alpha and with anxious anticipation gave the new image description a test run.


So far I am unimpressed in the extreme. This does not look anywhere near worth the effort. Here is one example, but I could give more: I asked it to describe an image I received via WhatsApp. Its description read: "A man looking at his cellphone". I showed Be My AI the same image and here is its description:


A screenshot of a Facebook feed on a mobile device. At the top of the screen, the time is 11:15, and the battery is at 57%. There are several navigation icons for Facebook at the top, indicating you are on the home feed.

The main visible post is from the group "Alberton What's Happening" by Nicole Moneron. The post is a warning reading: "Please be careful driving on the N12 towards voortrekker Bridge there's guys throwing rocks over the bridge and then a few guys on each side of the road ready to run." The post has received 4 angry reactions, 6 comments, and options to Like, Comment, and Share.

Below the post, Jonnno Moyle has commented: "Hate the human race."

There is a comment box with a profile picture of a child, ready for a new comment.

Below this, another post from "Dream Mansions" starts with the headline: "VIDEO SH0CK: The newly leaked 911 audio from the Charlie Kirk incident will leave you speechless. This is far more shocking than we were told..."

Above Nicole Moneron's post, part of a previous post is visible, showing images of outdoor scenes including a yard and a pool area, with a "+12" indicating more photos. This post has 22 likes, 1 comment, and 7 shares.


[End of description]


I realise that the image description feature is offline and based on a small language model, but it hardly seems usable. I will stick to the AI add-ons.


Kind regards


Christo

Sean Randall

unread,
Oct 3, 2025, 10:11:35 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org
I'm as excited as anyone about these llm's and the support they can give us. I regularly use them to interpret the visuals on my dauter's homework, describe family photos and memes and extract data from charts, visual pdf's etc. 
I still rely on standard OCR to read my food items and things like that though. Someone here said this integration was a gamechanger for us, but I disagree there. That level of description will add very little value to me on the regular. My little laptop's just not got the power to do much decent with generative yet when compared with the online models and that's fine. 

So although it's a cool feature, A bit like those people saying how amazing Talkback is now,  screen readers still do a heck of a lot without needing a language model on board to my mind. 
Thanks

Sean


On 3 Oct 2025, at 15:01, Christo de Klerk <christo...@gmail.com> wrote:



joseph....@gmail.com

unread,
Oct 3, 2025, 10:19:10 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org

Hi,

Regarding add-ons and 64-bit NVDA (a separate topic), any add-on with 32-bit only dependencies will not work on 64-bit NVDA (even after forcefully enabling them), including several speech synthesizers and global plugins.

Cheers,

Joseph

--

Richard Turner

unread,
Oct 3, 2025, 10:22:27 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org

When I went to the link provided for the alpha build, all I found were samples of snapshots.

I probably shouldn't be messing with alpha builds anyway, but I found nothing helpful at that link.

So, I will wait for a late beta or even the release…

 

 

 

Richard, USA

“Reality is the leading cause of stress for those who are in touch with it.

– Jane Wagner, from The Search for Signs of Intelligent Life in the Universe.

 

My web site: https://www.turner42.com

 

 

Sarah Alawami

unread,
Oct 3, 2025, 11:17:04 AM (2 days ago) Oct 3
to nvda-...@nvaccess.org

I use basiliskLLM for my image descriptions. I craft my prompt  and have it describe all of my images so I can do what I need to do later. so while this feature will be useful, there are other apps, better apps which will allow the use of more customizable  LLMs.

Aamir

unread,
Oct 3, 2025, 12:56:41 PM (2 days ago) Oct 3
to nvda-...@nvaccess.org, Quentin Christensen

Hello Quentin


I've tested the image descriptions, and I'm afraid I can't offer any positive feedback. Maybe it's just me, but so far, I haven't gotten a single result I'd call even remotely accurate.
It almost seems as though the AI is coming up with descriptions of random objects that have nothing to do with the actual image.
For example, here's what it told me about my desktop screen:
"a colorful advertisement for a k orean fashion show"
This is just one example; it has failed consistently. When I requested a description of a webpage, it told me it was a picture of a blue water bottle.

I'm hoping this will be improved before official release, as right now this just isn't up to the mark.
Thanks

muhammed ali çiçek

unread,
Oct 3, 2025, 12:59:30 PM (2 days ago) Oct 3
to nvda-...@nvaccess.org
"Hello, I'm very happy that this feature is coming. Which language model does it use, and when do you plan to release the stable version?
Best regards."

Quentin Christensen <que...@nvaccess.org>, 3 Eki 2025 Cum, 10:29 tarihinde şunu yazdı:
Hi everyone,

--

joseph....@gmail.com

unread,
Oct 3, 2025, 1:00:16 PM (2 days ago) Oct 3
to nvda-...@nvaccess.org

Hi,

This might be due to the model in use (looks like the model is Xenova GPT2 model as NVDA tries to download this model from huggingface.com (the version control repo used by OpenAI) he first time AI image captioner is enabled.

Cheers,

Joseph

Zvonimir Stanecic

unread,
Oct 3, 2025, 5:03:29 PM (2 days ago) Oct 3
to nvda-...@nvaccess.org, Gene Asner
Hi gene,

you open the image in some image viewer or in any place where you need a
description.

and then use the shortcut

W dniu 3.10.2025 o 13:00, Gene Asner pisze:
Reply all
Reply to author
Forward
0 new messages