It's not about how good a model is in the general [music] sense. It's about how good is this model for my use case. For a lot of design and marketing
use cases, we need editable design, not a single flat image. >> It's super impressive, honestly, reaching the level of things like Nano Banana or GPT
image with an open-source model. Why did you think that was important? >> We really want our models
to have taste. Every artist, they can really customize this model to the new answers of their style, the texture of their canvas, and really get 2K
output, and hopefully make that part of their workflow. >> One thing we were always wondering is uh this release open-source model is so small. It's
9.3 billion parameters. Like previously a SOTA is probably like 80 >> [laughter] >> billion parameters.
It's like 9x a difference. How did you do it? >> We focused on >> So, today we're excited to have Muhammad, CEO and founder of Ideogram, a
Toronto-based generative AI company that just released their first open weights image model. Congrats on the huge release. >> Congratulations. >>
Thanks for having me. >> We're really excited to talk through something that everyone has been buzzing about, which
is the fact that the model is open weights. Um the previous Ideogram models have been closed source, so we'd love to hear how you made the call to
make it open this time. >> What has happened is there has been a lot of progress in the industry. And uh we used to do everything. Basically, we had
our own first-party app, as well as our own first-party API. And model development itself is a lot of
work. And we decided to focus a little more on the model side. We think that's where a lot of potential exists. We still want to continue to own the
uh interaction with the users. We think there's a lot of important um feedback you can get from the users directly, but then we want to focus more on
building the model. And by releasing the weights, you're actually extending ourselves and working
with inference providers, working more directly with large enterprise. They have every ability to customize the models or host it on prem or optimize
it for device. And we would love to work with, you know, the best chip makers to really uh optimize the model, the best uh inference providers. So,
this is basically us saying, "Hey, we're very serious about building the foundation model and we would
like to work with you, wherever you are, whether you're app developer or a chip maker or an inference provider." >> I think you already kind of
touched on this. Um the new open-source model is very exciting in that it unlocked a lot of new use cases. It's like very photorealistic. I think it
can generate up to 2K with a, you know, smaller model, too. Um you know, like obviously there's very
precise layout control as well. Do you want to talk about some of the net new use cases that's unlocked by this model? >> So, this is actually a
foundation based on which we will we're going to release some more exciting features next. So, this is just the first release, you know, we're just
testing the waters, figuring out how to work with Hugging Face and the open-source community, ComfyUI, etc.
Um what I'm personally most excited about is something we haven't released yet, which is editable text and layout control. And I really believe for a
lot of design and marketing use cases, we need editable design, not a single flat image. And uh we haven't released that yet. We kind of showed the
teaser in our video, but um I'm personally most excited about that. On the technical side, what we've
done is we went really detailed on the prompting. And if you look at our prompts, it's like thousands of words each element in the image, where it is
in the image. Like we have layout control bounding box and a number of elements. And that's one of the key innovations here that unlocks a lot of
again design use cases because you clearly want font control, you want layout control. And this model is
very versatile. allows you to really fix certain elements, fix positioning, and control the image generation in every detail possible. >> Amazing. And
one of the other things I immediately noticed about the model was how you can render super long texts, like paragraphs of texts completely accurately,
which you either give the model in the prompt or you ask the model to kind of come up with
something and it does it really well. Um and it's super impressive, honestly, reaching the level of things like Nano Banana or GPT image with an
open-source model. Was that something you guys really focused on and sort of why did you think that was important? >> I don't know if you remember,
but the very first model we released 3 years ago. And at the time uh image generation was synonymous with
garbled text. And there were memes about DALL-E 2 generating travel posters with incorrect city names. >> Yeah. >> Which is fun to look at. So, I
remember at the time we were just a few people building these models. And the question was, how can we differentiate? What's unique about our model?
And we said, "Okay, text generation, accurate text is something we have." And then we released it and we
were really surprised. It just so many people were so excited about text generation. And then we realized, "Oh, actually, that's the whole graphic
design and you know, storytelling industry. Like text is very important part of image generation. And that became a very important part of our brand.
So, if you search Ideogram, people talk about the quality of typography, the quality of text. Um we are
known for, you know, really stylized typography for logo, t-shirt design, uh graphic design in general. And so, we continue to push forward. I think
our previous model wasn't really beating uh the state of the art in text generation, but we continue to focus on that and we had a bunch of research
breakthroughs, and with this model, despite the fact that it's very tiny, the text generation is very,
very accurate. >> One of the things that stood out to us, which is what the community has been chatting about, is how there's new uh ways of
processing data as you are training the model, uh which is like you kind of let the model learn what is a bounding box and how to do the layering and
color palettes. Do you want to talk more about some of the innovations you had during the training process on
what made this model so good with these differentiating features? >> Yeah, it's it's kind of difficult to exactly describe what what resulted in such
an amazing model. I think a lot of it is uh focus and evaluation. Evaluating image models is actually a very difficult thing to do. There are lots of
benchmarks out there, but people look at them and they're like, "Okay, this doesn't correlate with
pixel fidelity that I care about, the realism." Uh you don't really want novice users to judge the quality of these models because they may be looking
at small monitors um that aren't really adjusted for color accuracy, and um they all just care so much about, you know, quality, photorealism, and
again, text accuracy. So, throughout training, we always measure text accuracy and we apply very
detailed changes to the model and data and see how that results in performance. So, I would say a lot of it is really listing all the possible changes
and very carefully tuning each element of the model and see what happens. Obviously, we try to gather as much data as possible. One of the standard
recipes in the industry is that we take images and we turn them to text using visual language models.
The very first models we were training 3-4 years ago uh would be based on the alt text that you can find on internet. That is, each image on the
internet may have an alt text field associated with it, but which describes what's in the image. But, the problem is the alt text is often very short
or inaccurate. And what we do now is we train models to go from image to text. And in this case, image to
text with detailed bounding box information, detailed element information. If you care about text and we really want to make sure all the text in the
images correctly described. And then we go from text to image backwards. So, it's kind of interesting. We gather all the images from the internet.
Some of them may have alt text, some of them may not have alt text, and then we use AI to go from image
to text, and then we train another AI model to go from text to image. So, that's one of the key recipes that results in very good models. >> I saw a
lot of JSON prompting in your technical blog, uh which is very unique. And as a strategy model, it seems like it was translating the text, the prompt,
to a JSON representation with implicit structure. >> Right. >> Do you think JSON is a representation
for image models going forward, or do you think there's another representation there? >> It's a very good question. Uh I don't know if you've seen uh
the open source community is a little upset because of uh the safety image that shows up. >> [laughter] >> They always are, but about that to be >>
was really lashing out at our engineers and one of our people said, "Oh, we might fix this." And they
were like, "You might?" >> [laughter] >> But the fact is um the community to also read the documentation and bear with us. This model is only trained
with JSON prompting. And you have to provide JSON with that particular structure for you to get good quality output. So, I don't know if it's a
feature or bug. We did have some safety built into the model, but that is also uh detecting incorrect
prompts. So, if you just give it a one-word prompt, then you get a this image is blocked by safety uh image back, but that's because your prompt is
not a well-specified JSON. Um now, we don't want people to write in JSON. We don't think that's a natural way of interacting with these models, but I
do strongly believe that we need to use all the AI innovation to build the best image generation and
editing models. And there had been a lot of progress in language models in the in the text space. So, the question is if you want to go from some
vague idea to an image, what's the exact process? How much of the thinking happens in the language space, and how much of the thinking happens in the
actual kind of pixel generation space? I know you're an artist, so you should probably tell us um when
for example, we always had this prompt meaning of life. We we test our models based on this prompt. So, if you have meaning of life as your prompt,
then do you want an image gener- and a diffusion model to decide what the meaning of life is, or do you want a language model to kind of think and go
back and forth and come up with uh a description of a scene that's explaining uh the meaning of life.
And that's kind of the context of JSON prompting. It's the intermediate representation that we think language models can describe images in that
format, and then image generation can happen. In general, we see a lot of editing happening in the field, and that's the new frontier. So, I don't
think we should expect the interaction to be only through text or JSON, but it's a combination of JSON and
image, if I were to make a guess. >> Awesome. So, it sounds like a lot of it is basically taking often a relatively simple prompt that someone puts
into the model, and then translating it on the back end with the magic prompt and into JSON so that the model can make something probably more
detailed and interesting than the person with the short prompt may have even imagined. >> Exactly. And then I
think everybody else does it, too. OpenAI does it, Google does it, but then they don't give you the actual uh input to the model. >> Yeah. >> But
again, for professional use cases, you don't want to just roll the dice and then get some other completely different image interpretation of your
prompt. We show you the actual input to the model as well in the JSON format, and we think that will foster
more innovation and creativity. >> So, and for people who want control or like consistency, I think that'd be key. >> Yeah, and to Justing's point,
like exactly like what is the implication for the professional um use cases? Like what like with the JSON prompting, what can they do more easily now
with this capability compared to before? >> Maybe zooming out a bit, we the world is changing, right?
I'm sure your work has changed a lot. My work has changed a lot. I'm actually writing TRs now, which is very exciting. >> [laughter] >> We're in
collaboration with AI. AI does the writing. And then for creative professionals as well, the word is changing. They are very excited to Okay, a large
number of them are very excited. I think more and more of creatives are excited about adopting AI and
they see the potential. They think ideation is still the most important part of the creative process and humans are very good at putting context into
these models and their understanding of situation, creativity will result will result in the best ideas. I'm very excited about a future where, you
know, every kid will have these models at their fingertips and then they can be much more creative and
we're going to experience a much more beautiful world. Now, as it comes to JSON prompting, again, um because the JSON prompt describes every detail in
the scene, you can take it and change one element in the scene, and that results in very consistent output. So, you could be describing like a tiny
detail in the corner of the image and then leave everything else the same. And we think this also has
a big implication on editing. We haven't released We haven't released our editing models yet, but they will also utilize the same JSON prompting
approach. And uh it's just more control and uh with layout as well, you can uh imagine uh for every brand you have brand guidelines in terms of okay,
the size of text, the font of text, and we think uh this kind of foundation allows us to really get into
a lot of the uh enterprises cases. >> Amazing. We've been talking about some of the things you focused on with this model. Obviously, there's always
trade-offs in model training in terms of what you want the models to be really amazing at and what you focus on. Would love to hear what you focused
on for this release and also, do you consider more capability like we want to be the top for this
specific prompt adherence or something like that? Or is it more thinking about the end user and holistically, where the what are the different vectors
where they want the model to be performing? >> Right. So, we care about a couple things. One is graphic design in general. Again, text rendering is
part of that. We think um basic graphic design is everywhere. Like, you go you go in a city, you open
your eyes, you see billboards, you see storefronts, they all have text. And actually, it's much more important than um I guess photography is part of
graphic design. But like, it's graphic design is actually the frontier for a lot of business use cases, for storytelling. So, we definitely focused a
lot on graphic design uh since the release of our first model, which was good at text rendering. And
in addition, we think taste is extremely important. And and we really want our models to have taste. And it's very hard to explain it. >> Yeah. >>
What what is what exactly taste is. Uh one element of taste is kind of being um going outside of the normal a little bit and uh not conforming to the
average uh opinion, which is a little against being on top of the top of the leaderboard. >> [laughter]
Yeah. >> Which is kind of interesting. >> Your own leaderboard. >> Yeah, we just care about the Ultimately, we work with all the arenas. We hope
all of the arenas will improve in detecting the nuances of uh images and image quality, but we care about our own internal evaluation. Uh and
unfortunately, we see that AI is not very good at doing the actual taste evaluation yet. So, we work with
designers and we have side-by-side comparisons between different versions of the model as well as other models to really push on the taste. So, we
care we really care about taste. I think there's a still so much more to do, obviously. >> Yeah, I was going to my follow-up was going to be do you
have like one vibe guy or vibe woman internally who's like the taste arbiter cuz it can be hard to
measure taste, but it sounds like you have a a group of designers, which is probably better. [laughter] >> I think yeah, if we need to find that taste
maker. >> Yeah. >> Yeah. >> Um and then I guess one thing we were always wondering is this release open source model is so small. It's 9.3 billion
parameters. Um you know, like previously a soda is probably like 80 billion parameters. So, it's like
9x difference and then you can run it on a single GPU instead of having a lot of computer footprint. Which really opened up opportunity for people to
use it. So, the question is how did you do it? >> We we focused on the details of the model and we know we can't win on scaling. I I don't think I
used to work for Google. I don't think even if you raise 10x the amount we raised so far, we can beat
Google in terms of the number of chips that they can dedicate to each model training. So, instead we focused on innovation. We think there's a still
so much more to do to innovate. We are also focusing on differentiation. I don't think a lot of labs are focusing on design in graphic design in
particular, editable text that I'm talking about. And then we also decided to go open weight to really
partner with a lot of other platforms to um be at least another option uh for people who care about design. Uh and so yeah, so we focus on the small
model primarily because uh, we think there's a still so much to do. We think now is actually a good time for us to scale. Given the quality of the
model at 9.3 billion parameters, you should imagine what if this model is 100x bigger and uh, mixture of
experts architectures that don't make the model necessarily store, but they make the model a lot more powerful. So, I think that's one new frontier
for us to kind of scale this model 10x, 100x. >> I imagine because it's a smaller model as you mentioned, like it's harder to win on scaling and
counting number of chips, but it is possible to win on a specific domain or optimizing for a different, you
know, thing in a different domain. So, what was the trade-off for the research team when training this model to decide what to focus on? >> Right, so
one thing that you kind of alluded to is the fact that this can run on consumer GP you know. And we think there's a new frontier that, you know, you
do a lot of editing on your phone, a lot of image generation on your phone, and it's not only about
pushing quality at, you know, 100 billion parameter, 1 trillion parameter range. Uh, we think it's really important to have small models that can run
on device. Obviously, a lot of companies care about privacy, and we are really excited to partner with the industry to push the kind of small model
size quality further. Um, now in terms of the research team, um, it's it's an interesting question
whether you can focus on a very small, narrow field in image generation. I sort of believe that you need a general understanding of the world in order
to even be good at logo generation or be good at illustration uh, style. But, once you have a general base, then you can customize the model for
certain use cases, and it can be the best at that particular use case. So, we're really excited about
customization. We think that's a new frontier. And again, that's an important reason for releasing the model with open weights. That is, every artist
who has at least like 50 pieces of art or hopefully like a little more, they can really customize this model to the new answers of their style, the
texture of their canvas, and really get 2K output, and hopefully make that part of their workflow. And
you know, be augmented with AI and be a lot more productive and creative. And we actually have worked with some artists in resident who said to us,
"Okay, this at least made me 3x faster in making this comic book." Uh so, that's one frontier. Another frontier is enterprise again, cuz it's not
about how good a model is in the general sense. It's It's about how good is this model for my use case,
right? Uh at the end of the day, I may not care about many of the general purpose use cases. I may not care about character consistency, for example,
as an enterprise. But, even though this model is a small, we think it can be the best model for particular use cases, whether that's a certain artist
or whether that's an enterprise. >> Were you getting a lot of demand from enterprises when it was a
closed model that wanted to fine-tune it on their own data? Were there use cases where you're like, "We really need to open source it now because
there's so many cool things that people want to do"? >> Yeah, yeah. So, some of the enterprises we work with are very sensitive. They don't want to
talk about using AI in their visual uh of things, but what we've seen uh over and over again is companies
come to us and say, "We tried these generic models, and they don't meet our design bar. They don't follow our style. They don't follow our brand
guideline." And once we train custom models for them, they're like, "Wow, this understand my brand DNA now. We can use this for design ideation, or we
can use this for um marketing." And we thought with open rates release, we can give a glimpse of
customization to, you know, developers within enterprise, and kind of scale this side of our business further. Uh and we're really excited about that.
For enterprises, as you alluded to, I mean, customizing customization is really top of mind, whether that's a brand care or it's something that
your stylist typically just done, but hard to encode that style into just a doc. So, I imagine
customization on top of the open source model is just the best way to go. So, the question becomes like, what is the ramp-up for the customers or the
artists to start, you know, post-training or fine-tuning on top of the ideal brand model? What do they have to do? >> So, one thing I should say is we
will work with the open source community to make it as customizable as possible, but because we are
the model developer, we have some secret sauces that can make it even better. So, what will happen is there will be different ways of customizing. One
is in the open source based on the quantized model that's already released. The other is we already have a product that allows you to customize by
just uploading certain number of images to our custom model training app. And we haven't released the
full old version of that yet, but we are hoping to release that as well. And then the kind of third category is when enterprises work with us and we
really describe these detailed prompts for every image. We work with their design team to understand what words they want to use cuz each team has
different set of keywords. Each company may have certain mascots, who have certain names. And so um our
annotation team gets involved and spends a lot of time curating and cleaning data. And we think depending on your size and your budget, you should
still be able to customize a model, maybe use the open source at the low budget, and then you can come and talk to us so that we can build a model for
you at the high budget, but then depends on really uh the ROI that you have in mind for your models.
One of the things I think people have been talking about a lot both for enterprise and actually for consumer is um fine-tuning versus image
editing. I actually think they don't necessarily have to be competitive. Like some people use image editing as a way of fine-tuning. Like they say,
"Take this image and put it into this style." Others think it's much more efficient and consistent to just
fine-tune a model to generate in that style. I know you alluded to wanting to release an editing model further down the line. So would love to hear
how you guys are thinking about that. >> Yeah, I agree. I mean, I think editing is very powerful. We both agree. We all agree when editing launched
last year, so many new possibilities opened up. And the nice thing about editing is it's quick. You
don't have to train a model. You just take a style or an existing image and you make some changes. And it's part of your iterative workflow cuz with
every creative that we worked with, it's never one-shot prompting. It's often okay, you get something and then you're going to go and fix uh certain
details after the first generation. Now, that's for editing. But then customization gives you really
freedom to not prompt at all, right? Cuz sometimes it's very hard for you to say, "Oh, I want to get inspired by this single image and edit that one."
Right. >> You may have some general style in mind and want to ideate in the context of that general style or you may have a character that has many
detailed degrees of freedom or characteristics. The left side may be different from the right side
and like certain outfit and it's very hard to really put all of those images as the input to your editing model and it often fails. So, we think
customization can give you uh a lot more powerful adherence to your characters and allows for an easier iteration and ideation. So, I I agree with
you. I don't think they're mutually exclusive, but they're both very powerful. >> With the JSON prompting
and editing and model fine-tuning, the composability aspect of the model is just huge. There's so many ways you could customize it. Um one hot topic
we, you know, in the industry, in the research community is agentic loop for creative tools, right? So, it used to be the creativity tools, the
consumption layer is always a UI. As a human, I look at it and then I make modification. Now, so much of
that may become like an API request uh like the agent makes. How do you see that? What would the API intel uh compared to how humans use it? >> And
your earlier point, I I want to say something which is we seem to compare image generation with language models a lot. And for example, in the
language model space, uh even the customization exists, it's not that like every company customize their
language models. But I think that actually misses the point. When you look at visual representation of a brand, you immediately recognize the
differences between brands. But if you look at the written communication, can you say, oh, is this Andreessen Horowitz or this is Sequoia? I mean, you
probably can tell, but >> so. >> [laughter] >> I get the point. >> But most people will not be able to
immediately look at the text >> That's true. >> and then say. So, there is a lot more diversity in the visual world, and that's very exciting for
customization. And there are a lot of unique ways of interacting with the models that kind of goes back to your earlier question, a lot of 3D
manipulation. So, for that kind of use case, the input will be some 3D representation of the joints or position
of the objects. Then you may have a completely different, you know, stylistic variation with the style being the input. So, there are a lot of
different types of interactions that you want to enable. And for that reason, it's much different from the language space where the input is always
text and like you can kind of more or less convert everything to text. We're so excited about agents. We we
have our own MCP. We use them a lot internally. What's really exciting is when you want to release a new feature, you can go into your agents and then
ask it to connect to the API and generate a bunch of images, and then you can go and find the best ones, and like in a couple hours you have your
landing page up and running. So, we are very, very excited about agenting workflows. We think there's
it's just at the beginning. To your point, we need evaluation as part of the loop. We don't want to be to have to look at every image. Uh and then
editing will be part of the uh agentic interaction, too. How we exactly want to compose these different pieces to accomplish a goal is it still to be
discussed, but we have API and we have MCP, and we really uh want to enable uh the agentic workflow,
and we think every company is trying to figure it out as well. So, we're really excited about that. >> I guess for the API business, it so much of a
design is iteration. It's a long tail of the design. It's no longer just you prompt something, you get an image, and call it a day, right? It's so
much of a call get an image, use the edit model to, you know, edit it, see if it works well. If it
doesn't, get another image with the JSON flop, which is, you know, easier for control. What are the net new use cases you have seen after launching
the model of how people do like compose different API calls on the uh Ideogram model? >> What's interesting is yes, we have these agents, a lot of
them live in a chat bot, and that's not enough for iteration, unfortunately. So, I think what's unique
about that you can really scale creativity, like kind of give it some high-level direction, and ask it to go and explore many different approaches,
and come back with hundreds of thousands of designs that can be easily uh you know, looked at, and then you get a better sense of okay, I want to
explore more in this direction. So, so uh the language model interaction allows for kind of very
large-scale exploration of creative possibilities. But once you know what you want, you need a UI, you need a UX to be able to go and edit, and
whether it's regional editing or text-based editing, I think at the at the end of the day, you want your canvas, you want to be able to point to
things, and then you want to also talk to it with natural language. And and it's actually very hard work
because models are changing, and now you're also designing the user interface at the same time. So, um kudos to the best designers who understand how
these models work and are trying to figure this part out. There's a still a lot of work to do. In terms of the things we've seen from the model, uh
again, design is something that's coming up a lot, and I saw a tweet um before coming here that
somebody said, "I I have no design training, and I got this design in uh 2 minutes." And it actually looked really nice. Uh that's that was one
example. And then uh people are really interest- excited about the art possibilities cuz this model was trained with very unique style description. As
part of training, we actually stripped that as part uh from the JSON prompt because it became too much.
But uh the model has a lot of artistic possibilities, and like many different styles are embedded into model. And if you've seen some of the uh
frontier models actually that is scoring very highly in the leaderboards, they don't have a lot of um kind of design variation. They they always
produce the same exact look. And I believe that's because they did a lot of reinforcement learning training.
They actually have done very little reinforcement learning. So, this is a very raw model. Now, with that, the outcome is you need to be much more
precise with your prompting, but you can get a lot of different styles from the model. And people seem to be very excited about that uh that aspect of
the model, especially in the art community. >> I think when you talk both about the design and the art,
it really brings back the taste point you made earlier because for so many sorts of designs, you're trying to communicate some sort of idea, whether
it's like an infographic or an ad or a logo or whatever, and um it kind of needs to stand out and like be distinct to you or have a unique style. And
I've totally noticed what you said, a lot of the frontier that like historical frontier image models
like you're scrolling the feed and you just see like I've now seen this style like 50 times. I've seen it 100 times. It doesn't like catch my eye
anymore. And it feels like um now when I prompt Ideogram 4, I often get something that like makes me stop and be like, "Wow, this is different than
anything I've seen before coming out of an image model. And like this is doing an amazing job of both
communicating what I want to communicate and also holding someone's attention." >> Yeah, so we we try to be enabling very different styles and that
was one of our goals. We still want to produce tasteful output. >> Yeah. >> But that doesn't mean we have to force a complex output to you. If you
want a minimalist design, uh you should be able to get that. Actually, our minimalist is too minimalist
in my opinion. >> [laughter] >> I was saying we should ban the word minimalist from the from the output. Uh but you get what I mean. It's like the
model can do many different things and that's that's by design. >> We know what's the first Ideogram model. >> We're going to post train will be A16Z
marketing brand style Art Deco. A16Z Art Deco. >> Yeah, that's good. That's very exciting. >> Yeah, so
I guess the question also becomes like I kind of asked you the representation question before, which is here's a JSON representation. As a artist,
obviously, if you abstract it out, you know, far enough, all the lines are pixels. So, you could say that the composability is on a pixel level, which
is actually different from the diffusion representation. It's like denoising and you know, they
operate in a different space. Where does this lead to if you travel down the JSON path granular enough? Does it lead to pixels? Does it lead to SVGs?
Does it lead to like language or something else? >> That's a very good question. So, in general, the recipe for building more powerful models in my
opinion is making the task as straightforward as possible for the diffusion model. That is specify the
exact details of the image. Uh and so now if you kind of make that extreme, then it becomes the pixels themselves. So, the diffusion model doesn't
have to do anything. Now, the catch is what we would like to do is to get the language model to produce that intermediate representation. And language
models as of now, they aren't very good with continuous output. They aren't very good with kind of
pixel values or uh very high-dimensional vector representations. So, I guess the constraint here is that the representation has to be you know, tokens
of like I mean, depending on your language model power, maybe it can be a million tokens. Maybe it can be like that's too extreme. For us, it's it's
about 4,000 tokens. And uh and that's where we still use natural language because these large
language models are trained with natural language. So, they are very good with natural language. But, it may become more close to HTML, for example.
That that's okay because again, large language models are trained with HTML and they know the tokens. Uh but, it's going to be >> your own version of
HTML or would you align to HTML? >> kind of alluding to the editable model that I'm talking about and
we've had a lot of back and forth. Are we going to have our own JSON for different um you know, text elements and buttons and stuff, or are we going
to use HTML? And seems like HTML makes more sense just because these large language models have already been trained on HTML as opposed to us
introducing a new JSON structure. Uh, but I would say, uh, to answer your question, that representation needs
to be easy for the language model with the particular design that we have right now, which is the language model does some uh, expansion of the ideas,
and then the image model takes those expanded, uh, descriptions and turn them into images. >> Would love to hear if there's anyone who's interested in
working at Ideogram or working with you guys as a customer or to fine-tune a model, what's the
best way for them to get in touch with you and the team? >> First of all, we would love to work with more engineers, like cracked engineers. We have
We have a very tiny team. You see what we were able to produce. It's such a tiny team, and if you want high agency, you know, if you want your work to
matter, and part of And you want to be part of the academic and, uh, open-source ecosystem, then
this is the perfect time to join us. Uh, now, in addition, enterprises see the potential and we would like to work with the most creative brands out
there to help them produce the best designs, uh, produce most provocative ads, uh, and, uh, also, we would like to partner with other startups or
other companies at different levels of the stack. This is open rate, so we can make it win-win, and we
would love to offer a different option to companies, uh, who want more control and, uh, data, uh, privacy and sovereignty. Uh, so, we would love to
work with other enterprises as well across the stack. And then the best way would be like we have this email partnerships at Ideogram. You can DM me
on Twitter or LinkedIn and I'm I'm very active on both platforms. >> Today, if I want to fine-tune my
own style, where should I go on Ideogram? >> Just come faster. >> [laughter] >> Like is there a call to action we should tell people like they should
There is actually model tab. If you log into Ideogram, there is a model tab and then you can go and upload your images and train your model. It's a
little more expensive. It's 60 bucks for two model training per month, but we think for
professionals that's totally worth it. >> Absolutely. And how many images do you need to get started? >> I think you need at least 15. >> So for
enterprise, if they want to fine-tune their own model on Ideogram, like what's your guidance on what they should upload on Ideogram to start
fine-tuning in the model? >> I think for enterprise, again, we have some sales forms they can fill and then come
talk to us because we see that there are many differences in what different companies want. Some companies care more about editing. Some companies
more want more like on the marketing side to automate some of their ads and we should we should talk first and then figure out what's the best
solution for them. >> Got it. Great. Awesome. Thanks so much for joining us. >> Thank you so much. >> [music]