@Karen E. Lund Would you lose points for very long alt-texts/image descriptions?

I tend to describe my original images in extremely high details. Recently (for any definition of "recently" because I haven't posted a single image in over a year due to the huge effort of describing them), my alt-texts tend to reach 1,500 characters or one or a few below that. At least ca. 900 characters are actual image descriptions, sometimes up to ca. 1,400.

And that's what I consider a "short" description. Because the rest of the alt-text explains where to find the "long" description. It's in the post itself. It includes verbatim transcripts of every last bit of text anywhere within the borders of the image, readable or not. And it includes all explanations which I deem necessary for everyone to understand my images.

This long description exceeds any known arbitrarily defined character limit anywhere in the Fediverse by magnitudes. I can post such long image descriptions because the only character limit I have here on Hubzilla is the maximum size of the database field for the post text.

Yes, you've read that right. I describe each one of my original images twice.

And I must write my image descriptions that long. I don't post real-life photos, nor do I post social media screen shots. I post renderings from extremely obscure 3-D virtual worlds. Maybe one in 200,000 Fediverse users has even only heard of the technology that drives them.

Thus, I cannot assume anything in my images to be familiar to anyone out there. I can't assume that anyone out there knows what anything in my images looks like, also because my images tend to contain things which simply do not exist in real life in any shape or form.

At the same time, my impression is that especially Mastodon users expect all information which they don't have to be served on a silver platter immediately with the image description. If you mention something in your image, and somebody doesn't know what it looks like, you're obliged to describe it right away. Expecting anyone to ask you anything about your image after the fact feels like being considered ableist. I mean, you could just as well expect people to ask you to describe the whole image in the first place, right?

Same goes for explanations. Given the choice between looking stuff up themselves, being given links where they can look stuff up and being explained everything right there, right then, Mastodon users appear to greatly prefer the latter and only consider the latter really accessible.

And I have to explain a lot. When I tell you where I've taken an image, this alone takes me more characters than some of you use for a whole day's worth of alt-texts. The whole topic is so obscure that I have to explain explanations of explanations.

My personal record (warning: technically outdated image descriptions): 1,500 characters of alt-text, 1,400+ of which are image description; 60,000+ characters of long description for one image. That's about 10,000 words. It took me two full days, morning to evening, to research for and write the image descriptions. It takes a screen reader about three hours to read the long description out loud. But someone somewhere out there might be interested in all this information and displeased if they had to ask me about it to get it.

As for bilinguality, I should add to my WIP wiki on image descriptions and alt-texts that an alt-text must never include more than one image because screen readers cannot switch between languages mid-alt-text.

What I do, and I'm not even sure if that's such a good idea, is transcribe text in images that is not in English verbatim, literally letter by letter, and then translate it into English as closely as possible. I'm torn between a verbatim transcript which a screen reader cannot read out correctly and only giving a translation which would not be a verbatim transcript.

In fact, I've once had a situation in which I had to transcribe a sign in English, (broken) German and French. So I gave

• a 100% verbatim transcript of the English text
• a 100% verbatim transcript of the German text, all mistakes included
• an English translation of the German text that's as close to the original as possible
• a 100% verbatim transcript of the French text
• an English translation of the German text that's as close to the original as possible

Nowadays, I'd simply avoid posting images with non-English text anywhere in it like the plague.

CC: @Kim Possible :kimoji_fire: @Jayne

#Long #LongPost #CWLong #CWLongPost #AltText #AltTextMeta #CWAltTextMeta #ImageDescription #ImageDescriptions #ImageDescriptionMeta #CWImageDescriptionMeta

@Christian Kent   𝘊𝘒 :＼﹥ I can't really go with that proposal.

First of all, if you turn the automatic generation of alt-text on by default, then fully-abled people with tons of spoons who could write their own alt-texts with no problems will use that automatic alt-text-generating AI, too. Especially newbies who'll keep everything at default settings for quite a while. This would make alt-text quality and especially accuracy plummet.

Trust me. I've pitted LLaVA against my own hand-written image descriptions twice. In both cases, the images were renderings from very obscure 3-D virtual worlds with a resolution of only 800x533. The results were utterly devastating by my standards. Granted, the AI only had the images whereas I described the images while looking around in-world, being able to move the camera and zoom in on things and such.

So instead of turning AI-generated alt-text on by default, people should be educated about alt-text and told to write them themselves. Only if they can't, they should be told about the alt-text generator switch.

Besides, having this established on Mastodon is one thing. Trying to force the whole Fediverse to implement an image-describing AI is something else. And: It will never come to pass.

The Fediverse is far from being only Mastodon. The Fediverse is not even a) Mastodon, b) Mastodon forks and c) stuff (actually or seemingly) glued onto Mastodon as add-ons after the fact. There's stuff in the Fediverse that's vastly different from Mastodon while at the same time being in direct competition against Mastodon.

For example, you have rather minimalist stuff like snac2 or GoToSocial. I'm not sure if their devs are too keen on including an AI-based automatic alt-text generator even if the AI is external.

On the other end of the spectrum, you have Friendica, Hubzilla, (streams) and Forte. These four are vastly more powerful than Hubzilla by various degrees. And they're so modular that, in theory, it'd be possible to build an add-on that uses an AI to automatically describe images.

Now, the first issue is that neither of these four strives to be more like Mastodon. In fact, they've got their own cultures that date back to times when Mastodon wasn't even an idea. After all, Friendica is five and a half years older than Mastodon, and the others are part of the same big family tree of forks and forks of forks. And, unlike Mastodon, they're Facebook alternatives with a side of fully-featured long-form blogging and not minimalist Twitter alternatives (and Hubzilla is much much more than that even).

Their cultures are vastly different from Mastodon's and will always be. They don't aim for brevity, what with "character limits" with eight digits. They don't ask for CWs in a dedicated CW field. What's the CW field on Mastodon has been the abstract field on Friendica for almost seven years longer and on StatusNet for some nine years longer, and like Friendica, the other three use it for summaries. All four have always been able to automatically generate CWs individually for each reader, depending on a) the presence of keywords in a message and b) the keywords in that reader's "NSFW" keyword list. Something that Mastodon didn't have until October, 2022 when Mastodon 4 came out, and that is not part of Mastodon's culture because it came that late.

As for alt-text, users on all four are completely unaware of its existence at best. That's because they're completely out of touch with Mastodon and its culture. In fact, some Hubzilla users chose to leave ActivityPub switched off, and some (streams) users decided to turn it off themselves, in both cases with the explicit intent of locking Mastodon out. Of those few who have heard about alt-text in the Fediverse, most see it as just another stupid Mastodon fad that Mastodon is trying hard to force upon the rest of the Fediverse. Like that 500-character limit.

Also, especially the Hubzilla devs and the (streams) and Forte dev are not too keen on implementing Mastodon features to say the least. For now, they only do what's absolutely necessary to federate with Mastodon, but not much beyond that. These three will not make an image-describing AI add-on a priority task. Or a task at all. And don't ask for building that image-describing feature into the core. This will happen even less. Also, don't expect either of them to hard-code a connection to a power-hungry corporation like Alphabet (Google) or OpenAI into an add-on, much less into the core. Despite the fact that Hubzilla still has its old (optional, off-by-default for both hubs and channels) Twitter connector.

By the way: (streams) and Forte are the only server applications in the Fediverse with a server-wide user agent filter that's not only capable of blocking Mastodon in its entirety, but that was actually announced and advertised as having that capability while having actually been designed to keep Threads out. This should give you to think.

There's another obstacle: Where would the alt-text be stored? After all, the handling of images is, again, vastly different on these four than on Mastodon. On Mastodon, images are files attached to posts with their own alt-text data fields. And upon uploading the images, they end up in some data nirvana.

In stark contrast, Friendica, Hubzilla, (streams) and Forte all have a cloud file storage built into each account (Friendica)/channel (Hubzilla, (streams), Forte), complete with its own simple file manager and, except for Friendica, even WebDAV connectivity. It's there where image files are uploaded. And then they aren't attached to posts as files. Rather, they're embedded into posts just like into webpages or blog posts. You can have text, then an image, then more text, than another image, then more text. These four can generate something that Mastodon still staunchly refuses to even display.

Now, I know that (streams) and Forte do have a field for an image description. It can be filled out during upload or in the image editor after the upload. I'm not sure if Friendica has at least got a data field for alt-text. But I know for a fact from daily-driving it that Hubzilla does not have any data field for alt-text, not in the UI and most likely not in the database either.

In the posts, the alt-text is not added to the images by referencing that data field. The posts don't work on some murky hocus-pocus in the background. By default, they're composed in a markdown language in plain sight. Friendica uses expanded BBcode or optionally Markdown. Hubzilla uses even more expanded BBcode. (streams) and Forte can use the same expanded BBcode as Hubzilla and Markdown and HTML within the same post. And the alt-text is always part of the image-embedding code.

And this is also the only place on Hubzilla where alt-text can come into play. This means that if Hubzilla wanted to include automatic alt-text generation, the alt-text would have to a) be generated whenever an image is embedded into a post while composing the post (= you embed the same image file a hundred times, and a hundred times will the AI generate a new alt-text for it from scratch) and then b) be woven into the image-embedding BBcode.

Speaking of image storage, WriteFreely, a specialised long-form blogging server application and basically a Medium alternative, doesn't even have that. You can embed images into posts on WriteFreely, but only as long as you can store them someplace that allows for hotlinking because you will have to hotlink your images from there. But if WriteFreely doesn't handle images itself at all, if images in WriteFreely posts don't actually go through WriteFreely until after the post is sent (namely when someone views the post for the first time), it can't use an AI to automatically describe them either.

Unlike WriteFreely, Plume does have its own dedicated image storage. However, while Plume is not dead, its development has come to a complete halt for the foreseeable future because its devs don't have any time to work on it. And even if they got back to it, they'd probably have more important things to work on after several years of neglect.

And then there is software that is actually dead, that was officially pronounced dead by its own last maintainers, but that still has running servers. Calckey. Firefish. /kbin. But if it's dead, it won't even get security patches anymore, much less new features such as automatic AI-driven alt-text generation.

CC: @James Edwards @Danny Boling

#Long #LongPost #CWLong #CWLongPost #FediMeta #FediverseMeta #CWFediMeta #CWFediverseMeta #AltText #AltTextMeta #CWAltTextMeta #ImageDescription #ImageDescriptions #ImageDescriptionMeta #CWImageDescriptionMeta #AIVsHuman #HumanVsAI #CW #CWs #CWMeta #ContentWarning #ContentWarnings #ContentWarningMeta #Fediverse #MastodonCulture #FediverseCulture #snac #snac2 #GoToSocial #Friendica #Hubzilla #Streams #(streams) #Forte #Plume #WriteFreely #Calckey #Firefish #/kbin

@Baranduin

Oh does this feel like my inner monologue when I post a photo. It is a bummer that it, at times, prevents me from posting more photos, but I hope this make me a little more quality over quantity.

I actually keep entire categories of things out of my images because I can't describe them up to my own standards. This includes realistic buildings. I would first have to look up loads of architectural terms to describe all details of a building, and then I would have to explain each and every one of these architectural buildings in a way that absolute laypeople understand my image description without ever having to ask me anything or look anything up themselves.

The last time I posted an image with a building was this post. I actually went around and looked for a nice motive for a new image post for quite a while. There was one harbour scene which I thought looked spectacular enough to show, but which was impossible to describe. So I fell back to this motive. I thought it's not too bland, not too simple and at the same time not too complex. Besides, the one building in the image is totally unrealistic and without all the tiny details that would make up a realistic building.

And then I ended up taking some 30 hours over two days to describe the image in over 60,000 characters. The building alone took up some 40,000 or so. This is still the longest image description in the whole Fediverse, I think. Here is the image describing log thread.

My last image post before that was this one with still over 25,000 characters of description for one image, and I consider it outdated slop.

It was the last time that I described an image in my image with more details than visible in the original of that image itself. And that's where I got sloppy. I completely forgot to transcribe what's written on the license plate above the office door of the motel in that image in my image. And I couldn't be bothered to give detailed descriptions of the two 1957 Chevy Bel Airs parked in front of the motel because I really wanted to get that description done. In the actual image, all of this is sub-pixel-sized. You wouldn't know it's even there if I didn't mention it. I did describe the motel, but it's a fairly simple building, and I decided against describing what's visible through the windows with open blinds from the camera angle in the image in my image.

In the next image, the one with 60,000+ characters of description, I stopped describing images in the image beyond what I can see in the place where the image itself was taken. That was because one image is a destination preview image on a teleporter. The destination is a kind of teleport hub. The preview actually (if only barely so) shows over 300 single-destination teleporters, a few dozen of them with their own preview images.

So I teleported to that hub to describe it in detail. And I looked at the teleporters and their preview images. Turned out, not only do these preview images pretty much all have text in them and not necessarily few bits of text, but some of them actually have images within themselves again.

I would have had to describe that image in my image, dozens of images in that image in my image and a number of images in these images in that image in my image. For each of the latter, I would have had to teleport three times from the place that I originally wanted to describe. I would also have had a whole lot more text to transcribe. All on a sub-pixel scale several times over.

Not only would that have been a humongous task, but more importantly, it would have inflated my image description and my whole post to more than 100,000 characters. Mastodon would probably have rejected my post for being too long. And this would have rendered the whole effort futile. In the few places in the Fediverse that would still have accepted my post, nobody cares for image descriptions.

AI certainly can't get inside my brain well enough to write accurate descriptions. Even if it could would I? hmmm.

I've only used AI to describe images twice. And in both cases, that was to show just how bad AI is at describing images about an extremely obscure and quickly changing niche topic at the level of accuracy and detail which I deem necessary for that topic.

I guess one problem that you're facing is that next to nobody in the Fediverse can even grasp what you're thinking about, what you're taking into consideration for your image descriptions. That's why you got next to no feedback upon your first comment in this thread.

I have one advantage here: What you're pondering, I have actually done. If I feel like people won't understand what I'm thinking about, I point them at one or several of my actual image posts, and/or I post a quote from one of my actual image descriptions. Still, almost nobody actually goes and reads through any of my image descriptions, but I guess they get the gist, especially when I post snippets from my actual descriptions.

CC: @Icarosity

#Long #LongPost #CWLong #CWLongPost #AltText #AltTextMeta #CWAltTextMeta #ImageDescription #ImageDescriptions #ImageDescriptionMeta #CWImageDescriptionMeta

@Freeeeeeeeeeeeeeeeeeeee :ablobfoxbongohyper:‮‮‮‮‮‮

Here are some more:

• Always keep in mind that the primary target audience of your alt-text are blind or visually-impaired people. Do not write your alt-text for sighted people first and foremost.
• Do not start with "picture of", "image of" or "photo of". "Picture of" and "image of" are redundant because screen readers announce images anyway, and "photo of" is redundant because digital photographs are the default for online images. Do mention all other media, though.
• Do not add line breaks. Line breaks are not a standard part of alt-text, and they are useless for blind or visually-impaired people because screen readers won't read them out aloud. Not all frontends support them, and they may cause nasty side-effects.
• Do not add quotation marks from your keyboard. Quotation marks are not a standard part of alt-text, and they are useless for blind or visually-impaired people because screen readers won't read them out aloud. Not all frontends support them, and they may cause nasty side-effects or even severe breakage.
• Do not add emoji or fancy Unicode characters. They will disturb the reading flow for screen reader users because screen readers will call all of them by name.
• Do not add hashtags. They won't be parsed as hashtags, and links don't work in alt-text anyway.
• Describe dimensions and distances relatively to what other dimensions and/or distances are in the image.
• Do describe colours. Not everyone who needs a screen reader was born completely blind. Some people used to be sighted, and they have lost their eyesight later in their lives, but they still remember colours. But always describe colours using brightness, saturation and one or two familiar basic colours: red, blue, yellow, green, orange, purple, violet, pink, brown, gold, silver, black, grey, white. Do not simply use the name of a specific hue or shade. Not everyone who isn't sighted knows what it looks like. If you absolutely have to mention the name, do describe the colour afterwards.
• Think about what someone who comes across your post knows and what they don't know. Don't assume that everyone is familiar with the same things as you.
• Transcribe all text within the borders of the image 100% verbatim. (Sorry, I don't have any steadfast rules for text in foreign languages, for misspellings or other mistakes or for text that's unreadable in the image, but that you can read elsewhere and therefore transcribe.)
• Do not use all-caps for words or entire sentences. Some screen readers will spell them out, letter by letter, even if that isn't your intention.
• Do not use technical terms and/or jargon, no matter what you think people are or should be familiar with.
• Do not mention a person's race. If at all, mention their skin tone (light-skinned, medium-light-skinned, medium-skinned, medium-dark-skinned, dark-skinned).
• Do not mention a person's gender if you have to assume it. Only do it when a person actually performs their gender in an image, or that person has personally declared their gender to you, or you have a source of that person publicly declaring their gender, or you have a similarly rock-solid, undeniable source for that person's gender.
• Do not add any information that is not available in the post text and/or the image itself. This includes explanations, image credits and license information. Not everyone can access alt-text. Some people have physical disabilities that prevent them from accessing alt-text. Any information that's only available in alt-text is inaccessible and lost to them.

#Long #LongPost #CWLong #CWLongPost #AltText #AltTextMeta #CWAltTextMeta #ImageDescription #ImageDescriptions #ImageDescriptionMeta #CWImageDescriptionMeta