Show HN: I trained a deep learning model to build sites for me | Hacker News Reader

Show HN: I trained a deep learning model to build sites for me

zecoda.com

397 pointsjohndamaia124 comments

Founder here. As a front end developer, I turn design files into front-end code regularly. But it's a repetitive and demotivating task. It takes time, lots of meetings and boring back and forth questions. I want to change that.

So I tried to automate this process. I trained a deep learning (DL) model to visually understand what elements are in a webpage (e.g. input, button, nav, etc) and build the front-end boiler plate automatically for me. I learned a lot in the process, the actual site builder is a mathematical interpretation of a Sketch file and the DL model outputs what elements to include in what section.

To be completely transparent I'm not quite there yet. I'm only able to have ~90% of the site right. The other ~10% are misinterpretations or incorrect positioning of the elements. In order for this to work properly, it requires manual validation - which for my own projects I’m more than happy to do.

Even at this stage I think this tool can help others as well, so please let me know if you have any questions.

We can be more productive without repetitive tasks.

[1] Demo: https://zecoda.com/#demo

[2] Request early access: https://zecoda.com/#subscribe

Interested to see this.

I've done something similar in the past (building models that translated paper drawings into HTML+css) and ultimately found it usually didn't save me much time.

Translating a look into code is much less labor intensive than it once was and code reuse and stylesheet preprocessers and the ilk (webpack, etc) save tons of time.

I'd get something reasonable out of the model but then the actual look and feel work came into play.

My next step was to be able to reference other sites for styling hints but ultimately I wasn't getting enough utility out of it to continue.

Do you find the same with this approach? If so, what plans do you have to take it the next few yards so it's a more comprehensive front end tool.

Actually my workflow improved drastically after getting the boiler plate to work on the fun stuff (interactivity, animations, etc). I use this tool to get started much faster.

Hey, why not give something for people to try. I know[1] it takes (GPU) resources and may give it a hug of death but its much better to let at least some people try it out and give this a feedback while its hot. People may lose interest later on, who knows.

1. Something I made and let people use but nobody uses: https://comic2gif.com

Good point. I'll definitely give it for people to try in the next couple weeks.

Eh, if anything maybe invite one or two people for usability testing. Otherwise, having that many data points this early on might hurt your business.

You may want to have an example comic gif on the homepage. I didn't really understand what it did until I tried it. Plus, it would be neat if panels with more text were visible longer.

> You may want to have an example comic gif on the homepage

There is one. If you click on "show today's" button. But I guess default should be changed to show instead of hidden.

> Plus, it would be neat if panels with more text were visible longer.

There is some logic in code to do that, amount of white pixels (in mask) is proportional to the time it appears. But maybe it needs tweaking.

What stack did you use?

It takes time, lots of meetings and boring back and forth questions. I want to change that.

So I tried to automate this process.

The reason for all the meetings and questions is that the designs you're working from are ambiguous, and you don't really have enough information to turn the design files in to working code. I can't quite see how a deep learning model solves that problem. By the sounds of it you've automated turning the original, ambiguous, wrong designs the you have at the start in to code and tried to remove the discussion process that takes those files and turns them in to the right solution to the customer's problem. This is not going to lead to better front end code.

Depending on the customer, that could still be helpful. I've worked with people where you can ask them very specific questions over and over and get nowhere - the only way they're capable of moving forward is if you give them something you know is wrong so they can look at it and say "No, it needs to X."

It's annoying to do, but if you can automate producing the initially wrong product for review it'd be easier.

Couldn't agree with you more

Cunningham's Law is one of the truest design patterns in all of mankind.

For everyone else: "Cunningham's Law states 'the best way to get the right answer on the internet is not to ask a question; it's to post the wrong answer.'" [0]

[0] https://meta.wikimedia.org/wiki/Cunningham%27s_Law

Wouldn't you achieve the same feedback with minimal dev effort by simply modifying the sketch files with the client until they are signed off on it, which is what designers would be doing before discussing the technical feasibility/performance tradeoffs with devs for implementation? Assuming you are working with a designer that is creating sketch files.

Have you worked with clients? If you have a good producer/leader yes you can get a lot done. But many times this just happens as A. Client approves design B. You build the Beta of the website and launch it for testing with client C. Client shows wife/boss/friend who says “I think we should change this”

End result: you end up having to redo the website. This tool helps get that Beta up even if just for the client to try it and show his friends/wife/boss

The point I made in my reply is that this tool won't save you from those meetings and rounds of changes. As it'll speed up the coding side of the iterative development process by making exactly what the client has described rather than the developer using their experience to build what the client actually wants, there'll have be more meetings in order to get to what the client really needs.

The benefit of automating the production of code is useful, and plenty of people would want that, but targeting the messaging at developers isn't going to work. I would have thought selling this tool to people who employ developers as a way to reduce the number of people they need is a far better route to market.

Depends on who he's meeting with and what industry. If it's an ad/service company where the design process is controlled by project managers or the clients... then I totally see where he's coming from in terms of meetings being a drag

There's almost always a reason why the person who organised a meeting wants one, and I doubt you can use automation to solve whatever their problem is without understanding it first.

If anything, by automating the process of producing code you're going to leave front end developers with nothing but meetings because the rest of the job is going to become a matter of pushing an AI-powered button. They won't thank you for that.

There is advantage here. Imagine being able to have a CI process for design files - just upload your sketch file's changes to the repo and this tool in your pipeline builds/deploys a version of this in HTML in your staging environment. Easy link sharing built in.

If designers have the ability to work closer to the web, more power to them. It should be our job as software engineers to make this as easy as possible and this tool could be one specific implementation.

Unless the machine learning model can infer what those discussions were going to lead to. That would be truly revolutionary.

For larger clients who are familiar with the design process that's true. But for smaller clients, I would 100% take a working, "wrong" version of a site based on a design file over Sketch iterations any day. You'll get immediate, actionable feedback much faster (ex. "I don't like button X, change it to Y"), often times with less effort and a happier client at the end.

How well do you think this concept will scale to bigger code bases and bigger production tools like photoshop?

The next step up has been pushing ML/DL/AI to do more sophisticated tasks and that means more sophisticated tools.

At this stage we are working hard to solve this simple problem of going from a design file to a front-end boiler plate. No animations or logic. We are looking forward to get as much feedback as possible from our users to understand where to go next.

Interesting idea but how maintainable is the code?

I've used a few tools over the years that automatically generate code, mostly in C, but I've usually had to rewrite it as it never fits in with any decent coding standard

That's our main goal. To always give you good code. Otherwise this tool doesn't make sense. Hopefully you'll find this code maintainable with close to zero rewrites.

Is the access to provide data tagging inputs and outputs or to the model too? Will either the data or the model be open sourced?

Seems like you’ e created the ultimate sketch plugin.

Wow, nice work! I used to work for Webydo on the part that translate the DB to static files, tons of work and the result was basically an "uneditable HTML blob".

I would really love to see the output zip file, any change you have something less complicated that goes responsive to show?

If you think it can help, I can send you the output code for the demo. Drop me a line at john [@] zecoda.com

Hi! I'm super curious about how you structured the input and output to the neural network, and which architecture you chose, if you're willing to share (I work at OpenAI)

If you do share I'm happy to give you feedback here if I have any!

I’m sorry but this feels really hand-wavy to me. What did you feed your model for it to learn how to build vuejs sites? Vuejs template language is total chaos if you haven’t put in quite a bit of time with vue before, there are 2 or 3 ways to do everything. I am having trouble believing this was auto-generated from a deeply learned model. How do you write css, are you writing sass, vue inline styles, or something vanilla? Are you bundling with with webpack, what custom modules does your webpack use for handling those inline svgs? So so many questions about how and if this was auto generated...

Great questions. And I also understand your skepticism. You're not the first and don't expect you to be the last :)

Trying to answer all your questions:

- Our models learnt website's components visually (e.g. what is a navbar, an input, a button, etc)

- The output is a .zip file with the .html files and directories for css, js and img

- All Vue components can be found in main.js inside the js directory

- Vue components are built when 'duplication' is recognized (e.g. a sequence of cards or several links for a navbar). Nothing fancy

- CSS is vanilla, trying to ditch as many inline style as possible (working on it as we speak :) )

- All SVGs are imported with the img element.

Let me know if you have more great questions.

Cool, thanks for the reply. For me making components is the hard work, anything that helps automate organizing / naming / and synchronizing a component library with sketch sounds intriguing to me. Is the output intended to be code-edited by a developer to add hrefs, interactivity? Is it intended to generate pseudo components for later finessing or is it intended to produce the final outcome of a site directly from sketch (making Sketch the new dreamweaver?)

It's intended to be code-edited by a developer. The goal is to have an advance starting point. You'll start with "everything" ready to work on the fun stuff. If you curious enough to see the output I'm happy to share the code from our demo. If it makes sense drop me a line at john [@] zecoda.com

Hi. I would like to have an option of using just the automated part of the service and do the developer part myself. If hypothetically you would include that option, how cheaper do you think it wold be compared to the current price ? Also one more option: convert design to just HTML and CSS, without javascript or framework of any kind. Would that make sense ?

Interesting feedback. Our main goal is to ditch the human verification, we want to be confident enough to provide great code automatically. Not there yet.

Regarding the output, it's great to know your thoughts on the frameworks. We actually went for a framework to make it simple for our users. In the future we want to allow our users to chose what output they want. So, it makes total sense.

Sounds good. Regarding the no-framework output, I'm not against frameworks, I was just thinking for a moment from the perspective of a more traditional web designer who only incorporates bits of pre-made JS code in their design. Such a person might not even know what Vue.js is but may still greatly benefit from your service if it allowed them to convert their designs to just html/css.

how is this different than working with a really good starter kit? I guess if you already have sketch designs this will give you a better starting point.

It's really great to see more and more companies emerging in the market from these initial ideas/PoC:

[1] Pix2Code: https://arxiv.org/abs/1705.07962

[2] Sketch2Code: https://airbnb.design/sketching-interfaces/

[3] ScreenShot2Code: https://blog.floydhub.com/turning-design-mockups-into-code-w...

Totally. The SharpieClassifier we built at Airbnb was super simple because we had a relatively fixed set of components that could be declaratively rendered with React or react-sketchapp as simply as <Marquee /> etc. I can't imagine how much more complex UIzard & Zecoda are - mad respect to the developers and for being able to commercialize these ideas.

Aside: I actually built react-sketchapp as a roundabout way of being able to build more interesting AI plugins for Sketch.

Sorry to be a bit meta here-- but a question about the design of the site.

This style of pastel colors, wavy svg graphics, and happy avatars seems super popular right now [1] [2] [3]. I'm curious if it has a name and where it originated from (if anywhere?)

Overall I like the aesthetic quite a bit, but it's such a distinctive look. I'm surprised I keep seeing it crop up nearly copy+pasted between all of these companies

[1] https://www.gusto.com [2] https://www.tandem.chat [3] https://www.notion.so

You'll see a similar thing in branding and advertising. For all the new yorkers on here, have you all seen the consumer brand ads on the subway? Almost every brand looks the same.

I'm imagining it has to do with the way designers work. They generally have stakeholders moodboard (basically creating a pinterest board with inspiration from different brands), and then they create a similar "feeling".

As humans, we're kind of evolved to recognize patterns, so when we start seeing the same branding art direction it stands out to us and we may try to conflate some deeper meaning to it. But in reality, it's just a few brands that share the same branding art direction probably because some execs liked the way it looked and felt and asked their designers to do something similar.

This style of web design is super in at the moment. Not sure it has a specific name, but I believe the previous Intercom site started it. Possibly Wealthfront as well. A lot of companies seem to be using it to take more risks with their aesthetic, particularly Dropbox and Mailchimp.

Another question about the design of the page: Does this kind of conversational / persona based pitch actually work to convert users? E.G. "Hi I'm Clara, my life's goal is... I struggle with..." "If you relate with Clara, try X now..."

At least for me the net emotional effect of the "For Desigers & Developers" section of the homepage was a strong impulse to bounce. (I'm in the target market.) Ultimately for the sake of research I clicked through to the ~/product page, where I found the diagrammatic graphics much more compelling.

Great feedback! I guess we just need to find our best target market. We believe the decision makers will be managers and not technical people (engineers that relate more with the product page instead of the emotional illustrations). But I could be wrong...

Right, how to be relevant to both the decision maker as well as the end-user is a key question.

I'm working on a product in a similar space, and struggling with the question for our own marketing materials. Amy Hoy's material [1] on audience "worldviews" has been helpful so far. (No personal affiliation.)

Best of luck with your launch!

[1] https://stackingthebricks.com/niches-are-for-suckers/

Not sure if there's a specific name for the avatars, but there's two near similar resources here:

[1] https://undraw.co/illustrations

[2] https://www.humaaans.com/

A lot of this looks like the next iteration on the semantic ui [1] look (particularly the avatars).

[1] https://semantic-ui.com/

This is very interesting. I love the approach to the problem that you are trying to solve. However, there are few questions that I will try to answer with respective to the product.

Q. Will I pay to use your problem?

A. Probably not, with the current version

Q. What problems I see with the platform?

A. I seriously don't know how you'd proceed forward with the platform. If it would only be used for creating static webpages then you'd end up creating a platform similar to Wix or other such website generation platforms.

Please let me know what you think about the suggestion.

Let me turn back some questions to you:

1. What's not enough in this current version for you to pay? What would make you be able to pay?

2. Do you have this problem recurrently? i.e. going from design to code.

It would be interesting to see the quality of the code. Way back in the 90s Macromedia had a product called Fireworks, which Adobe subsequently bought (along with Flash, Dreamweaver etc.). It could turn designs into HTML+CSS but the code was horrible and no-one used it professionally.

I know we are decades past that, and the machine learning algos may do a better job but without seeing some code examples it has limited appeal to me.

I don't really know the kind of person that has to turn around so many websites that the boilerplating exercise is a big enough problem to pay to solve. I can see the appeal to people who have no experience building websites and a company like Wix might be interested in the tech. But professionals who need a monthly subscription has to be a tiny niche.

Impressive all the same though, what you have learned in doing this will lead to other great things I am sure.

Thanks so much for your words. Happy to share our demo code, would love your feedback. If you think it makes sense, drop me a line at john [@] zecoda.com

It might be easier, since this is targeted at developers, to make a GitHub repo with both the Sketch files and the generated code so that we can determine whether this is something that we’d benefit from.

From my perspective, one of the difficulties I’d have is that we already have a pretty extensive Vue application, but we _still_ have problems with translating updates in place. So, we’d want to be able to compare one of the screens we have with one of the screens generated (and our designers tend to create multiple Sketches to illustrate different state possibilities).

The next question would be how updates might be flagged to minimize code churn after initial generation.

Wix still requires work to encode designs into html pages with interactivity. The use of this tool is to jump straight from designs to those initial pages with essentially no work. It also will be a lot more versatile with respect to the type of designs it accepts and can generate.

Even if you wouldn't pay for this I know a lot of high-revenue generating businesses that would. The time savings afforded by this tool (if it works) would be incredible.

Surprised it’s Vue and not React. Vue is great, but there’s a lot of activity in the Sketch <-> React space, for example AirBnb’s Lona [0]. This could fit in nicely.

[0] https://github.com/airbnb/Lona

We are part of its community for years. It's always good to start with tools you master. React is definitely on our roadmap. Are you more of a Vue or React dev?

Lona is After Effects -> web / native.

You might be thinking about react-sketchapp[1] or SharpieClassifier [2]

(I built both :P)

[1] https://airbnb.design/painting-with-code/ [2] https://airbnb.design/sketching-interfaces/

I thought Lottie was the After Effects to code plugin. Happy to be wrong though.

Anyway, it’s a pleasure to meet you - I’ve been impressed with both your projects and following them for a while now. Love to talk more about them.

It would make sense since react does the jsx thing and Vue uses straight html inside of the template tags.

Looks like Yotako...https://www.yotako.io/

Whats the difference? Yotako seems to be in the deadpool though...havn't seen anything new in ages. Hard to scale this stuff.

Didn't know them. The concept seems the same :)

Is this just a quick and short project or have you been working on this long term?

One month full time; around one year part-time.

Excuse my pessimism in this sea of optimism, but I see a .sketch-> vue.js converter which has proprietary elements of DL which later on gets checked with manual work anyway.

Where's the building of websites? What's the difference between this solution and something which simply converts .sketch into html with far simpler methods?

This is a really interesting project, thanks for sharing. I think it's going to be difficult to get the model to the point where it generalizes well for lots of different designs, but that's no reason not to try. Really hard problems sometimes lead to really big wins.

Hi. Wish you had one-off pricing. I'm sure there are agencies that do that kind of volume. I would love to use your service a few times a year.

Will an one-off monthly fee work? Or were you thinking on a per artboard approach?

I'm the same - I'd pay $30/40 for a single one, but yearly cost of $99 is too much

Can you please clarify what you mean by "single one"? You'd pay $30/40 for how many pages of one site?

As in using it a single time - I guess maybe at most 2/3 pages from one design?

Ok, cool. In fact you are right in our price point of $9 per artboard. The only think we need to evaluate is the one-off possibility which I'm sure we can try.

It's actually $99/month if paid yearly.

I think with a one off it’s better just to outsource it. But the benefit of the automatic approach is that you could iterate a design over the course of a few weeks, getting quick turn around on each tweak.

I'm honestly kind of surprised nothing like this has found its way into Sketch itself yet.

Looks cool so far.

I also always thought this is a natural thing for Sketch to have it. Webflow gives similar experience of Sketch but to build web pages.

This is really cool, it is really hard to evaluate the value of Zencoda without the code output of the demo. I get that the full product is not ready but being able to see the input and one good output would tell us a lot.

Thanks! Happy to share our demo code, would love your feedback. If you think it makes sense, drop me a line at john [@] zecoda.com

I built a thing like thi, and went quite a long way down the commercialization path.

The problem is that round-tripping is what almost everyone wants, and it's really hard to get to work.

Can you elaborate on what you mean by round-tripping please?

I assume they mean from web back to sketch.

Meaning going from front-end code to Sketch, right? Is this something you need?

I suppose most website html templates are later filled with code ( php, js, whatever). Then design change, you don’t want to restart inserting php code all over again.

At least that was one issue i thought about reading OP.

Is the home page itself built with this?

Can you share the input file and an estimate on how long it took to work with the output to make it production ready?

Most of the website was started with the model. However I was learning along the way, so most pages were fairly changed.

Regarding the time for the optimization of the output, we are estimating ~20/30 minutes per page. It could be very faraway from the reality or pretty close, not sure yet.

A dogfooding video, where you build your own site with it, would be much more convincing.

The next step is to allow voice/touch-driven repositioning of elements that the initial model didn't get right, e.g. telling it to "move Submit above Cancel" or just grabbing it and placing it there. That would clear most issues customers have and they could do that themselves, including changing styles/themes and seeing how it behaves right away.

Hi John, great idea but the demo didn't really convince me. How much of the system relies on engineers cleaning up the code?

And I can't get past the fact that it becomes a responsive site, is the logic in the responsive design in the Sketch file too?

Hey! No worries, I'm starting to like the skepticism, it makes the job funnier. Right now we are doing it right ~90% of the whole page. The responsiveness is done by the system automatically (it makes decisions on its own, therefore they could be wrong). No need to have it on the Sketch file.

I guess when you present magic, people get skeptical ;)

But seriously, I hope you guys do well, thanks for responding to all these skeptical comments

Hey john. Congrats on this, this is very interesting. From someone interested in ML/AI, can you explain just a bit more how the deep learning works?

Without going into detail, our model outputs the coordinates of the html elements it recognize. We are taking a visual approach to the problem.

Did you use a recurrent neural network? And/or reinforcement learning?

I imagine you do some kind of image recognition on the sketch files and then do the code generation.

I once built a tool that generates XAML from Sketch files and Zeplin projects.

I can now see how AI would have made my job easier.

I'm curious how you are handling ambiguity in the original design. Does the algorithm create multiple variants when something could be interpreted in multiple ways, and then the manual validation part of the process involves picking from among the different variants that the algorithm created?

It's simpler than that. The model makes only one choice based on what it learnt. However you gave an excellent idea :)

I predict that one day you'll be able to input your content, specify some site layouts you like and press a button to generate a completely ujique website based on your content, fully designed and coded. The first person to make that is going to be very rich.

Firedrop was doing that, but it seems they've pivoted to something else.

As I recall, it was literally upload your text and photo/video content, and their AI will design a static site for you, optimizing for SEO and engagement.

Seems a bit similar to what Avocode was/is trying to solve. Haven't followed them closely recently, but seems like they pivoted from pure design->code tool to design & collaboration tool that enables easy export to code...

Always a pleasure to see others end up with a project that I've working on myself for 2 years on my free time... Don't expect to have the market... I'll. (even if I have to set prices twice cheaper than yours). Sorry, I was working on this project for... 2 years on my free time. I can't fail. I don't matter if I "steal" your customers or what. Because during these 2 years, I was working for my firm, then I was going to workout for 1 hour, then I was working on this project for 3 hours, then bedtime. My weekends and holidays were dedicated-only to the project dev. I didn't have any social life for 2 years in order to work on a Websites generator. And YOU come with your own sites generator????? And you created it with a team, not me I was alone. I will have the market for me only. Too many sacrifices.

It would help if there were examples of input and output—images of Sketch files and HTML pages that were generated (without manual fixes, if that's an option in your offers).

Since it is completely automated and the code is doing all the processing, why is the price so high and why is there even a charge is per artboard?

Sketch is like 1/10th of that

Not completely automated at this stage. Check my last paragraph on my first comment here. As soon we can make it work 100% automatically we will have much more freedom to change our pricing model.

The fact that operational costs are low doesn't mean that development costs can now be ignored.

This looks promising, however, I'm not sure about the pricing model, $99 per month (IF billed annually) for 10 designs (10 artboards)?

Yes, one artboard usually correspond to one webpage. How much were you willing to pay? Is this a problem you have every week?

Your rates are higher than paying someone (offshore) to convert the artboards - why? Should be much cheaper if it's automated.

Wow, do you know developers that convert artboards to full webpage for $9? But I understand your point, we are intentionally pricing it at a high point. Hopefully we'll be able to decrease it in the future.

The definition of "full webpage" can be scaled out of that price point, but yes. I wish you best of luck.

wouldn't be surprised if this dude is bought out by Invision in the coming year or two.

How many FTE engineers can this replace? Do you have a case study or whitepaper?

Our goal is not to replace engineers. Is to increase their productivity and give them more time to create awesome products.

Sure, but every bit of automation that "increases productivity" and "gives more time for other things" naturally ends up replacing those that it was helping -- if not directly, at least replacing potential future hires in the same role.

There's a useless vertical scrollbar on desktop.

Which page?

All of them.

Can't replicate the issue... Can you show me a screen shot?

I have the same issue on Firefox. It's to the right of the webpage, next to Firefox's scrollbar.

#app { ... overflow-y: scroll !important; }

Is causing it and not needed.

Thanks!!!

From https://news.ycombinator.com/showhn.html

> Show HN is for something you've made that other people can play with. HN users can try it out, give you feedback, and ask questions in the thread.

This should not be a Show HN as there is nothing to try out, just a video + early access list that currently doesn't give you access.

Updated. Thanks for the heads up.

Sorry to be that guy, but fullstory, google tagmanager, and hotjar all on the same page, without any indication to the user that their every move is being tracked. This site is not too GDPR compatible.

Also, I'm working on something similar, but from a different perspective. My problem with this is that you have to design your website in Sketch, then turn it into code with a different service potentially waiting hours until you'll see the end result.

So my idea is that I can build an app where you can basically draw the page elements, similarly to how you'd do it in Sketch, and it automagically turns the whole thing into a responsive website right away (no position:absolute hacks involved). I still have a lot to do though... :D

Reply on news.ycombinator.com