Skip to content →

The Idea Place Posts

Image Description Toolkit 3.6.0 Available

A 3.6.0 release for my Image Description Toolkit is now available. The release features a new –redescribe option as well as support for two Microsoft Florence-2 models via Hugging Face.

Redescribe

The redescribe option can be used with the command line version of the toolkit to redescribe the same set of images without the need to extract video frames or convert image files to JPG from HEIC. This can save time and allow you to quickly test different prompts on the same set of images. To use the redescribe option, simply add –redescribe as the first option in an “idt workflow” command followed by the standard –provider, –model and –prompt-style options. Use the workflow directory from a previous image description workflow as the file directory for images to be described. Product documentation has full details.

Hugging Face and Florence Models

Hugging Face is described as “The platform where the machine learning community collaborates on models, datasets, and applications.” It contains a wide range of AI models, datasets and apps across the full spectrum of AI tasks.

The Florence-2 AI model is described in part as, “an advanced vision foundation model that uses a prompt-based approach to handle a wide range of vision and vision-language tasks.” It runs locally and requires no pre-installation. Note that the model will download on first use so expect a longer image description time. Also be aware that for a variety of reasons, the Florence-2 model is only supported in the command line version of the Image Description Toolkit.

You can get the latest update for the Image Description Toolkit from my projects page at www.theideaplace.net/projects or the GitHub release page.

2 Comments

Give Accessibility a Seat at the Notepad Table Please

Unreal, I read today that Microsoft has started rolling out table support for Notepad. But guess what, if you are using a screen reader, the fact that you are in a table when editing is not communicated at all. This despite the fact that Office has clearly communicated tables when editing for many years with screen readers, including Narrator. So, clearly technical solutions exist and have been used for a long time. Why is something this basic once again left out for accessibility?

If you are so inclined, you can up vote feedback I submitted on this issue at this feedback link.

This is another example of why I say we need more accountability around accessibility. If Microsoft is going to enhance Notepad, accessibility needs an equal seat at the table, not something missed yet again.

2 Comments

Windows Copilot Serves At Best Half an Answer to Screen Reading Users

If you follow me on social media, one of the themes you’ll find I talk about fairly often related to accessibility is how text that isn’t shown visually needs to have better verification. This can be anything from alt text on web pages, to the accessible names used in different applications for controls such as menus and more. For several months now, the Windows Copilot app has had a serious problem that fits in this category. I know I reported it to Microsoft at the start of March of this year and the problem remains as broken today as it was then.

In this case, the challenge is that the text communicated for accessibility purposes is clipped such that when reviewing answers with Copilot, The answers you receive are incomplete.

If I ask Copilot for a list of the fifty United States, I’m led to believe that there are only 27 states. In the way the list is rendered to a screen reader, the List ends at the State of Nebraska. I guess Wisconsin, where I am writing this from, has disappeared?

If I ask for a list of the U.S. presidents, the list when read with a screen reader only shows 10 presidents and in fact cuts off the last name such that you end up getting only a partial name. At least my history never said we had a president named “10. John T.”

You can try various modes with the different screen readers, such as Narrator’s Scan Mode, the JAWS Touch Cursor and more and the problem is not improved. In most cases, such as Narrator’s Scan Mode, you end up reading some text twice but still never get the full answer.

It should not be necessary to copy a response and paste it into another app just to read the answer from an AI system or any app for that matter.

The frustrating thing is that none of this was an issue at the start of 2025. Microsoft appeared to be coalescing around a fairly standard user experience for the multitude of Copilot experiences available to customers. The foundation for the Windows app was a well-structured HTML experience with headings and more that made reading answers quite straight forward.

Sometime after the start of the year, the Windows Copilot app appears to have changed some of the foundational technology used that has caused this sorry state of incomplete answers when read with a screen reader. One can only wonder how long this busted experience will be allowed to exist. It is already several months too long in my opinion.

To be honest, it is incredibly frustrating to write blog posts like this. I doubt they do much good, but I am not sure what other avenues to pursue. As I said, I reported this situation to Microsoft in early March of this year. Through multiple email exchanges, all I know is that Microsoft claims they are taking the issue seriously.

It is repeated situations like this, where accessibility issues linger on and on where I say we need a shift to true accountable accessibility. I outlined a proposal for such earlier this year as well.

This is just one of several significant accessibility challenges with the Windows Copilot app. I’m a firm believer that all experiences should be usable and accessible, but this is especially true for AI-based situations. We need to be able to access the responses put out from AI effectively and obviously comprehensively.

In fact, if I ask the Windows Copilot app to give me a definition of accessible software, the experience is to me completely unacceptable. Things start out well enough with automatic reading of a message that Copilot is preparing the answer. They then progress to a reading of the beginning of the answer, only to have that reading interrupted by a restart of the reading of the answer that starts over from the beginning, only to read more of the full answer automatically.

Ignoring the multiple tabs or shift+tabs it requires to get from the edit box for communicating with Copilot to the chat history, the answer in the chat history is again clipped from what you heard with the automatic reading. Here is my chat history as read by Narrator when moving to the Copilot chat history for the accessibility prompt and response.

give me the definition of accessible software please.
**Accessible software is software designed to be usable by people with a wide range of disabilities, ensuring equal access to digital tools and content.**
### 🧩 What Accessible Software Means
Accessible software incorporates features and design principles that remove barriers for users with:
– **Visual impairments** (e.g., blindness, low vision, color blindness)
– **Auditory impairments** (e.g., deafness, hard of hearing)
– **Motor impairments** (e.g., limited dexterity, paralysis)
– **Cognitive or neurol

If I copy the answer, you can read just how much is skipped and also the fact that any of the structure, such as lists, is not communicated when reading in the Copilot app. Here is the full copied answer to my question.

Accessible software is software designed to be usable by people with a wide range of disabilities, ensuring equal access to digital tools and content.

🧩 What Accessible Software Means

Accessible software incorporates features and design principles that remove barriers for users with:

  • Visual impairments (e.g., blindness, low vision, color blindness)
  • Auditory impairments (e.g., deafness, hard of hearing)
  • Motor impairments (e.g., limited dexterity, paralysis)
  • Cognitive or neurological conditions (e.g., dyslexia, ADHD, autism)

The goal is to ensure that everyone, regardless of ability, can interact with the software effectively and independently.

🔍 Key Features of Accessible Software

  • Keyboard navigation: All functions can be accessed without a mouse.
  • Screen reader compatibility: Proper use of semantic HTML, ARIA roles, and alt text for images.
  • Color contrast and scalable text: Ensures readability for users with low vision.
  • Captions and transcripts: For multimedia content, aiding users with hearing impairments.
  • Logical structure and focus management: Helps users navigate and understand content flow.
  • Customizable interfaces: Allows users to adjust layout, font size, and input methods.

📐 Standards and Compliance

Accessible software typically adheres to guidelines such as:

  • WCAG (Web Content Accessibility Guidelines) – the global standard for digital accessibility.
  • Section 508 (U.S.) – mandates accessible technology for federal agencies.
  • EN 301 549 (EU) – accessibility requirements for ICT products and services.

🧠 Why It Matters

  • Inclusivity: Promotes equal participation in digital spaces.
  • Legal compliance: Avoids lawsuits and meets regulatory requirements.
  • Better UX for all: Accessibility often improves usability for everyone.
  • Ethical responsibility: Reflects a commitment to social equity and universal design.

If you’d like examples of accessible software in action or want to benchmark a specific app, I can help you analyze it.

2 Comments

Image Description Toolkit 3.5 Beta Featuring Geolocation Data and Web Image Downloads

With more AI-driven development I have another sizable update for my Image Description Toolkit or IDT. There is a full What’s New document available.

Highlights for this beta release include use of geolocation data when present in images, the ability to download images from a specified web address and have them run through the image description system and numerous other enhancements.

You can also keep current with all my projects from my Projects page.

One Comment

Feedback on IDT Demo Gallery

I’m looking to crowd source some feedback. I’ve mentioned here a few times a collection of tools I’ve created called the Image Description Toolkit. The short version of this is that it is a way to get image descriptions that you can save and customize the level of detail you get. This can be a bit of an abstract concept in a world where many still do not understand alt text.

So, I’ve put together a demo page at www.kellford.com/idtdemo. It has the traditional image gallery but then a Description Explorer. The Description Explorer allows you to see how different AI prompts result in various image descriptions and how different AI providers do at image descriptions. There are a total of four prompts (narrative, colorful, technical and detailed) using 10 different AI provider/model combinations.

For example, choose Description Explorer and then the option for all prompts from a provider. Note how the descriptions built on each other in a way from Narrative to Colorful to Technical.

The point of this demo is to showcase the sort of data my toolkit can make available. Whether you are an individual like me who wants more access to my pictures with different descriptions, or you want longer descriptions for other purposes, this is an example of what my toolkit makes possible.

This is not the one-off random describe this picture type of system. There are hundreds of those. This is the I want permanent descriptions at scale type of system.

Feedback I’m love to have. First off, does the web page look reasonable and free from glaring problems? Do the concepts of what info you can have from my toolkit make sense from this demo? If not, what would help?

One very interesting challenge. AI vision models are in my experience not great at generating alt text. I tried a range of prompts to get them to do so. In the end, the alt text (not my longer descriptions) was created by taking the Narrative prompts created by AI and running those through AI again asking for alt text to be created. You can see an example of this in action by using the Image Browser and choosing to show the alt text visibly. Note, choosing this mode with a screen reader will result in alt text reading twice–once as alt text on the images and once as the visible version of the alt text. I debated what to do about this situation and, so far, opted to turn off the visible display of alt text on page load. I do want people to see the alt text on demand because it is part of the overall system.

The toolkit allows for this sort of data gathering and gallery creation to be done all automatically. Just point the tools at a collection of images and an AI provider and you can choose how the info is shown.

Again, visit http://www.kellford.com/idtdemo for the gallery. Visit https://github.com/kellylford/Image-Description-Toolkit/releases/tag/v3.0.1 for the toolkit itself and https://theideaplace.net/image-description-toolkit-3-0-available/ for my latest blog post on the toolkit.

Leave a Comment

Image Description Toolkit 3.0 Available

I have a vastly updated version of my Image Description Toolkit (IDT) available. The 3.0 release builds on the command line workflow tools I released earlier with a guided workflow creation system, analysis tools for combining descriptions, a range of performance statistics on your descriptions and tools to review content from various AI models and prompts.

The IDT also includes several other tools for working with image descriptions. A results viewer allows you to browse results, monitor workflows as they happen, copy descriptions and images.

If you want to focus on detailed descriptions for individual images, the Image Describer tool, allows you to load a directory of images, and easily request descriptions for individual images and do so using multiple providers and prompts and save all your work for future use.

A Prompt Editor allows you to create and edit the prompts used to guide AI descriptions for all the tools in the IDT and configure defaults.

In addition to Ollama, which powered earlier versions of the IDT, the IDT now includes support for multiple AI models from both OpenAI and Anthropic if you provide your own API key.

A comprehensive User Guide is available for the various tools in the IDT.

Whether you want detailed descriptions for one or one thousand images, the IDT has a tool for you. Grab IDT3.zip, extract the files to your computer and run install_idt. Be sure to install Ollama and or get API keys for OpenAI or Anthropic and configure your AI models. Then open a command prompt and run:

idt guideme to generate your first descriptions.

15 Comments

RSS Quick, A Basic RSS Reader for Windows 1.0 Available

I’ve been dabbling further with AI-driven development and have another app for exploration. This time it is a basic RSS Reader for Windows I’m calling RSS Quick. Get the full details and program from Release RSS Quick 1.0 – Initial Release · kellylford/rssquick or my alternative download location.

The name RSS Quick reflects the intent of the app to be a simple and fast way to quickly browse RSS feeds. The program does not handle feed management at this time but will allow you to import an existing OPML file and quickly browse the feeds. RSS Quick includes a default OPML file named RSS.OPML that will be put in the default folder where you unzip the software package. It has dozens of feeds from a wide range of sources already included. Even if you don’t use the app, the OPML file may be handy as another source of feeds.

To use RSS Quick, you need to download the .ZIP file for your computer. In most cases this will be the X64 version. If you are using ARM hardware with Windows, you’ll want the ARM version. Extract all the files to any location on your computer and run RSSQuick.EXE. The program will load the default OPML file that has been included. If you do not have the necessary version of .NET, you will be prompted to download that first.

You can learn more about RSS Quick from the readme for the program. Give it a try and consider getting involved or submitting issues on the GitHub page.

RSS Quick is one of three projects I have available. I also have my Sports Scores app and an Image Description Toolkit. available. All projects can also be downloaded from my separate release archive.

Leave a Comment

Making Sports Information More Accessible: A Preview of the Sports Scores App”

I’ve gone back to the AI workbench, much like I did with my Image Description Toolkit, to create another app. This time it is to address the goal to make sports scores and play-by-play information easier to read and avoid some of the long-standing accessibility challenges on leading sports web sites.

I’m calling this app a preview release because not everything is working yet but I have enough functionality available that I wanted to share it now.

I have an audio demo of the app along with a transcript of the demo. The release page with notes and links to download the Windows program are at Release Sports Scores 0.5 Preview Release · kellylford/Scores.

The program uses lists, trees and tables to show things like game scores, play-by-play action, box scores and standings. The info itself comes from APIs made available by leading sports providers. I’m just showing it in ways that work better in my opinion with screen readers and other assistive technology.

Perhaps the most interesting feature I have so far is the ability to hear baseball pitch location and speed by converting pitch data to audio sounds. The demo has a sample of this and within the app you can press Alt+p to hear the location of any pitch when you are on that pitch in the play-by-play details. Press Alt+s to hear the entire pitch sequence. Bring up the context menu on a pitch for these options plus a Pitch Explorer feature to get a sense of how the sounds for different locations and pitches will be presented.

In addition to the release page, the GitHub project has a readme with full details on the app. An issue list where you can report new issues or discover what’s already known is available.

Give the app a try and if you are interested in contributing to the project, please do so.

Leave a Comment

Image Description Toolkit V2 Available

I’ve made another series of updates to what I’m calling the Image Description Toolkit since my last announcement. As a recap, the goal of this toolkit is to take collections of images and videos and create descriptions you can save and do this all with local AI models. Dozens of tools provide descriptions, but it is still difficult to save those descriptions for future review. With the Image Description Toolkit, you get nicely formatted HTML pages to read through all your image descriptions.

The newest enhancements include a comprehensive testing system to experiment with model prompts, a workflow script that allows for all tasks to be run with one command versus individually running each script and numerous small adjustments throughout the system. The code here is still all AI-generated with my ideas powering what’s created.

I’m sure I’m not objective but for me this has turned into something that started as a curiosity, moved into a better understanding of how AI code generation could work and is now something I’m using regularly. Over the weekend I attended several musical events and was able to generate more than 400 image descriptions from photos and videos I took.

The project lives on GitHub and has a readme that covers the basics of getting started. A guide for using the prompt testing script is also available. This is particularly heklpful for trying out different models.

I’m always curious how AI writing works as well so asked GitHub Copilot to generate a second blog post about project developments. And of course, it is software, so there is also an issue list.

I won’t say for certain what’s next but my current plan is to work on a graphical version of the project to understand more about that environment with Python, create a prompt editor so changing the default prompts is easier and get this all working with Python packaging so install is easier.

Contributions, suggestions or pointers to tools that already do all of this are always welcome.

Leave a Comment

Updates to Image Description Toolkit

Several months ago I announced a highly experimental set of Python scripts I called The Image Description Toolkit. Consider it a fancy name for solving my goal of wanting a way to get thousands of pictures taken from my iPhone and also for the past several decades from whatever phone I was using described and having a permanent description of the photos. I’ve made some key updates, although I’d still say this is categorized as highly experimental.

Most notably, I’ve made it possible to build custom AI prompts, choose the model you use and adjust the parameters used with the model and have all of this done through a configuration file.

I’ve also updated the script that will convert files in the .HEIC format to .JPG and streamlined the output to HTML with a script that can be run. To be very clear, when I say I’ve done these things. All the code in this project was generated with AI through my prompting and refinement.

A readme for the project explaining how all this works is available. I also had AI generate a blog post about the project. You can find the full project on GitHub.

With all of those qualifications, I have found these tools of value. I’ve now generated more than 10,000 image descriptions running on my local computer. The Moondream model used through Ollama has been excellent. It is incredibly fast when used for batch processing, has some of the lowest memory requirements I’ve found and still gives rich details and is highly responsive to different prompts.

I plan to continue experimenting here over time. I want to make setup easier and know about Python packaging but have found it doesn’t always work so this all still requires manual install of Ollama, Python and the individual scripts. The readme file should walk you through this though.

If you have feedback, know of other ways to accomplish these same tasks or suggestions on what else I should include here, feel free to let me know. I’ve leanred a great deal about image processing from AI, using Python and AI code generation from these experiments. And of course, I now have permanent descriptions of more than 10,000 pictures.

Leave a Comment