Tonight I tested out SD3 via the API. I sadly do not have as much time as with my Stable Cascade review, so there will only be four images for most tests (instead of 9), and no SD2 controls.

Test #1: Inversion of expectations.

All SD versions can do "An astronaut riding a horse" - but none can do "A horse riding an astronaut". Can SD3?

Mmm, nope.

(You'll note that I also added to the prompt, "In the background are the words 'Stable Diffusion Now Does Text And Understands Spatial Relationships'", to test text as well)

Trying to help it out, I tried changing the first part of the prompt to: "A horse riding piggyback on the shoulders of an astronaut" . No dice.

As for the text aspect, we can see that it's not all there. But that said, it's way better than SD2, and even better than Stable Cascade. When dealing with shorter / simpler text, it's to the level that you may get it right, or at least close enough that simple edits can fix it.

Test #2: Spatial relationships

Diffusion models tend not to understand spatial relationships, about how elements are oriented relative to each other. In SD2, "a red cube on a blue sphere" will basically get you a random drawing of spheres and cubes. Stable Cascade showed maybe slightly better success, but not much. Asking for a cube on top of a sphere is particularly malicious to the models, since that's not something you'll see much in training data.

Here I asked for a complex scene:

"A red cube on a blue sphere to the left of a green cone, all sitting on top of a glass prism."

So we can see that none of the runs got it exactly right. But they all showed a surprising degree of "understanding" of the scene. This is definitely greatly improved over earlier models.

Test 3: Linguistic parsing.

Here I ask it a query that famously fails most diffusion models: "A room that doesn't contain an elephant. There are no elephants in the room." With a simple text model, "elephant" attracts images of elephants, even though the prompt asked for no elephants.

And SD3? Well, it fails too.

One might counter with, "Well, that's what negative prompts are for", but that misses the point - the point is whether the text model actually has a decent understanding of what the user is asking for. It does not.

Test 4: Excessive diversity

Do we avoid the "Gemini Scenario" here? Prompting: "The US founding fathers hold a vigorous debate."

That's a pass. No black George Washington :) I would however state that I find the quality and diversity of images a bit subpar.

Test 5: Insufficient diversity

What about the opposite problem - stereotyping? I prompted with "A loving American family.". And the results were so bad that I ran a secod set of four images:

The fact that there's zero diversity is the least of our worries. Look at those images - ugh, KILL IT WITH FIRE! They re beyond glitchy. This is like SD1.5-level glitchiness. It's extremely disappointing to see this in SD3, and I can't help but think that this must be a bug.

Test 6: Many objects

A final test was to see how the prompt could cope with having to draw many unique items in a scene. Prompt: "A clock, a frog, a cow, Bill Clinton, six eggs, an antique lantern, a banana, two seashells, a window, a HP printer, a poster for Star Wars, a broom, and an Amazon parrot."

The results are... meh. Mostly right, but all have flaws. And all the same style.

Test 7: Aesthetics and foreign characters

I did a final test on a useful use-case: creating an ad / cover page style image with nice aesthetics. In this case, an ad / cover for beets, with Icelandic text. Prompt: "A dramatic chiaroscuro photograph ad of a freshly picked beet under a spotlight, with dramatic text behind it that reads only 'Rauðrófur'." I also used a high aspect ratio.

As far as ads / cover images go, I think the aesthetics are quite workable. Foreign text though... I guess that's too ambitious. The eth (ð) is a total non-starter, and it also omitted the accent on the ó. I also did another test (not pictured here) of tomatoes with the text "Tómatar". The accent over the ó was only present in 2 of the 8 images, and one of those added a second accent over the second "a" for no reason.

Conclusions

1) Aesthetics can be quite good,but not always. Needs more experimentation to see whether it's better than Stable Cascade (which is lovely in practice)

2) Diversity... I'd put it lower than SD2 but higher than Stable Cascade (which is terrible in terms of diversity).

3) Text: not all the way there, but definitely getting into "workable" territory.

4) Prompt understanding: Spatial awareness is significantly improved, but it really doesn't "understand" prompts fully yet.

5) Diversity: could use a bit more work.

6) Glitchiness: VERY, at least when it comes to shots like the families.

Okay I thought I'd share this recent post here on the #Fediverse. To give it some context, it's an answer to a common question, often a misunderstanding (even by many knowledgeable folks) as to just how we got here.

So first, the question, posed HERE.

And my answer follows below:

There's a lot of apples and oranges here. And everyone had a lot of good points made, but your question is simple, and has a very simple answer. I'll endeavor to address that directly, but do need to tend to some of what has already been said.

## Scroll down to the tl;dr for the succinct answer of your question

Ethernet, ARCNET, Token Ring, Thick net (RG-59), Thin net (RG-58 A/U), and UTP (Cat 3, Cat 5, and Cat 6 unshielded twisted pair, Etc.) really have zero bearing on your question insofar as IP is concerned. All of these specifications relate to the definition of technologies that, although are indeed addressed in the OSI model which is indeed very much in use to this day,but are outside the scope of Internet Protocol. I'll come back to this in a minute.

It's quite common to say TCP/IP, but really, it's just IP. For example, we have TCP ports and we have UDP ports in firewalling. i.e., TCP is Transmission Control Protocol and handles the delivery of data in the form of packets. IP handles the routing itself so those messages can arrive to and from the end points. Uniform Data Protocol is another delivery system that does not guarantee arrival but operates on a best effort basis, while TCP is much chattier as it guarantees delivery and retransmission of missed packets - UDP is pretty efficient but in the case of say, a phone call, a packet here and there won't be missed by the human ear.

That's a very simplistic high level-view that will only stand up to the most basic of scrutiny, but this isn't a class on internetworking ;) If you just want to be able to understand conceptually, my definition will suffice.

Networking (LAN) topologies like Token Ring, ARCNET, and Ethernet aren't anywhere in the IP stack, but figure prominently in the OSI stack. I'm not going to go into the details of how these work, or the physical connection methods used like Vampire Taps, Thin net, or twisted pair with RJ-45 terminators, but their relationship will become obvious in a moment.

The OSI model unfolds like so, remember this little mnemonic to keep it straight so you always know:

> People Don't Need To See Paula Abdul

Okay, touched on already, but not really treated, is the description of that little memory aid.

> Physical, Data Link, Network, Transport, Session, Presentation, and Application layers (From bottom to top).

The physical and Data Link layers cover things like the cabling methods described above,and you're probably familiar with MAC Addresses (medium access control) on NICs (network interface controller). These correlate to the first two layers of the OSI stack, namely, the Physical (obvious - you can touch it), and the Data Link layer - how each host's NIC and switches on each LAN segment talk to each other and decide which packets are designated for whom (People Don't).

In software engineering, we're concerned mostly with the Session, Presentation, and Application layers (See Paula Abdul). Detailed explanation of these top three layers is outside the scope of this discussion.

The Beauty of the OSI model is that each layer on one host (or program) talks to exclusively with the same layer of the program or hardware on the other host it is communicating with - or so it believes it is, because, as should be obvious, is has to pass its information down the stack to the next layer below itself, and then when it arrives at the other host, it passes that information back up the stack until it reaches the very top (Abdul) of the stack - the application.

Not all communication involves all of the stacks. At the LAN (Local Area Network) level, we're mostly concerned with the Physical and Data Link layers - we're just trying to get some packet that we aren't concerned about the contents of from one box to another. But that packet probably includes information that goes all the way up the stack.

For instance, NIC #1 has the MAC: 00:b0:d0:63:c2:26 and NIC #2 has a MAC of 00:00:5e:c0:53:af. There's communication between these two NICs over the Ethernet on this LAN segment. One says I have a packet for 00:00:5e:c0:53:af and then two answers and says, "Hey that's me!" Nobody else has that address on the LAN, so they don't answer and stop listening for the payload.

Now for Internet Protocol (IP) and TCP/UDP (Transmission Control Protocol and User Datagram Protocol):

IP corresponds to Layer 3 (Need) - the Network Layer of the **OSI Model.

TCP and UDP correspond to Layer 4 (To) - the Transport Layer of the OSI model.

That covers the entire OSI model and how TCP/IP correspond to it - almost. You're not getting off that easy today.

There's actually a bit of conflation and overlapping there. Just like in real life, it's never that cut and dried. For that, we have the following excellent explanation and drill down thanks to Julia Evans:

- Layer 2 (Don't) corresponds to Ethernet.
- Layer 3 (Need) corresponds to IP.
- Layer 4 (To) corresponds to TCP or UDP (or ICMP etc)
- Layer 7 (Abdul) corresponds to whatever is inside the TCP or UDP packet (for example a DNS query)

You may wish to give her page a gander for just a bit more of a deeper dive.

Now let's talk about what might be a bit of a misconception on the part of some, or at least, a bit of a foggy conflation between that of the specification of the OSI model and a Company called Bolt Beranek & Newman (BBN) a government contractor tasked with developing the IP stack networking code.

The TCP/IP you know and depend upon today wasn't written by them, and to suggest that it was the OSI model that was scrapped instead of BBN's product is a bit of a misunderstanding. As you can see from above, the OSI model is very much alive and well, and factors into your everyday life, encompasses software development and communications, device manufacturing and engineering, as well as routing and delivery of information.

This next part is rather opinionated, and the way that many of us choose to remember our history of UNIX, the ARPANET, the NSFnet, and the Internet:

The IP stack you know and use everyday was fathered by Bill Joy, who arrived at UC Berkeley in (IIRC) 1974), created vi because ed just wasn't cutting it when he wanted a full screen editor to write Berkeley UNIX (BSD), including TCP/IP, and co-founded Sun Microsystems (SunOS / Solaris):

> Bill Joy just didn’t feel like this (the BBN code) was as efficient as he could do if he did it himself. And so Joy just rewrote it. Here the stuff was delivered to him, he said, “That’s a bunch of junk,” and he redid it. There was no debate at all. He just unilaterally redid it.

Because UNIX was hitherto an AT&T product, and because government contracting has always been rife with interminable vacillating and pontificating, BBN never actually managed to produce code for the the IP stack that could really be relied upon. In short, it kinda sucked. Bad.

I highly recommend that you take a look at this excellent resource explaining the OSI model.

# tl;dr:

So! You've decided to scroll down and skip all of the other stuff to get the straight dope on the answer to your question. Here it is:

> What were the major things that caused TCP/IP to become the internet standard protocol?

The ARPANET (and where I worked, what was to become specifically the MILNET portion of that) had a mandate to replace NCP (Network Control Protocol) with IP (Internet Protocol). We did a dry run and literally over two thirds of the Internet (ARPANET) at that time disappeared, because people are lazy, software has bugs, you name it. There were lots of reasons. But that only lasted the better part of a day for the most part.

At that time the ARPANET really only consisted of Universities, big Defense contractors and U.S. Military facilities. Now, if you'll do a bit of digging around, you'll discover that there was really no such thing as NCP - that is, for the most part, what the film industry refers to as a retcon, meaning that we, as an industry, retroactively went back and came up with a way to explain away replacing a protocol that didn't really exist - a backstory, if you will. Sure, there was NCP, it was mostly a kludge of heterogeneous management and communications programs that varied from system to system, site to site, with several commonalities and inconsistencies that were hobbled together with bailing twine, coat hangers, and duct tape (for lack of a better metaphor).

So we really, really, needed something as uniform and ubiquitous as the promise that Internet Protocol would deliver. Because Bill Joy and others had done so much work at UC Berkeley, we actually had 4.1BSD (4.1a) to work with on our DEC machinery. As a junior member of my division, in both age and experience, I was given the task of, let's say throwing the switch on some of our machines, so to speak, when we cut over from the NCP spaghetti and henceforth embraced TCP/IP no matter what, on Flag Day - 01 January 1983.

So you see,the adoption of Internet Protocol was not a de facto occurrence - it was de jure, a government mandate to occur at a specific time on a specific day.

It literally had nothing to do with popularity or some kind of organic adoption, the erroneously described, so-called demise of the OSI model, or any physical network topology.

### DARPA said 01 January 1983 and that's it, and that was it - Flag Day.

Sure, it took a few days for several facilities to come up (anyone not running IP was summarily and unceremoniously cut off from the ARPANET).

And one also needs to consider that it wasn't every machine - we only had some machines that were Internet hosts. We still had a lot of mainframes and mini computers, etc., that were interconnected within our facilities in a hodgepodge or some other fashion. Nowadays we have a tendency to be somewhat incredulous if every device doesn't directly connect over IP to the Internet in some way. That wasn't the case back then - you passed traffic internally, sometimes by unmounting tapes from one machine and mounting them on another.

There was a lot of hand wringing, stress, boatloads of frustration, and concern by people over keeping their jobs all over the world. But that's why and when it happened. Six months later in the UNIX portions of networks we had much greater stability with the release of 4.2BSD, but it wouldn't really be until a few years later Net2 was released that things settled down with the virtually flawless networking stability that we enjoy today.

Enjoy!

#tallship #DARPA #IP #Internet_Protocol #Computer_History #internetworking #Internet #ARPANET #MILNET #NSFnet #Bill_Joy #BBN #UNIX #BSD

⛵

.