Hacker News Reader

Comment by akersten | Hacker News Reader

I really don't have anything constructive to say. I think in general we're getting too soft on shitty things, so I'm going to be harsh.

I clicked through to the demo site ( http://gaugan.org/gaugan2/ ) and it was horrible.

The interface is clunky, slow, and confusing. I actually had to zoom out in my browser to see the whole thing. Had to click through a non-HTTPs warning. The onboarding tutorial is pretty bad.

I got a generic picture of the milky way for any prompt I tried ("rocks", "trees"). If you press Enter in the prompt field it refreshes the page.

This feels like a hackathon front-end hooked up to an intro to PyTorch webservice. It's only neat because, unlike the other 20 copies of this same project I've seen, it was the only one that didn't immediately throw its hands up and say "server overloaded, please wait."

If I'm meant to be impressed or fearful of "big data deepfake AI," this isn't it.

[0] https://imgur.com/BNLDt6A

View on HN

Your comment is a really good example of how startups can fail.

Here we have a web app that does something very clever, built by great devs, that fails because it doesn't work the way users expect it to.

So many highly intelligent, super technical founders have a belief that their amazing tech will sell itself, so they don't put the time and effort necessary in to design, UX, or marketing, and they fail because their UI didn't make it clear what to do. It probably works brilliantly when they demo to people, with the authors driving it or helping users get the most from it. But when users have to use it without that help... It fails.

The lesson for founders here is simple - test your UX, because you won't get a second chance with most customers.

The app I used worked nothing like the slick demo in the video. In fact, the UX and UI are some of the worst I have used in recent memory.

No matter what some backend folks believe, there will always need to be highly skilled front end engineers who can put together web apps in a way where the interface just 'gets out of the way' so you can focus on the actual utility.

"The app I used worked nothing like the slick demo in the video."

This point is very important and I hope to not do the same misstake.

Because I also watched the video, saw what was happening there, looked nice - but trying it for myself absolutely did not work as expected.

In other words, if there is a simple video with features shown - then trying it out needs to be as simple as the video, or it causes lots of frustration.

First time users do not want to deal with setup configs etc. first. This is something you want later.

After a sibling commenter very patiently pointed out that I was holding it wrong - I would encourage taking another look at this project if only to try out the Painting mode.

It does produce some very cool results: https://imgur.com/LQuo4UM

Had the press release/tutorial emphasized this angle instead of the wonky text-to-image thing, my initial impression would have been a lot better. This is genuinely a really neat feature. All my UI and discoverability criticism stands though!

The text-to-image thing is cool as well, as long as you can a) figure out how it works b) only enter landscape terms.

This comment explains how to use it: https://news.ycombinator.com/item?id=29338213

I managed to produce landscape - canada fall colors red yellow green and bright blue sunny sky.

No matter what variations I tried the sky was cloudy dark. Trees however were majestic. The UI does suck.

Your query worked for me on first try: https://imgur.com/a/iTKwZ3v

I tested it again and it worked. Not sure what was I doing wrong first few tries

Type in "cat" into the text box and see a wonderful variety of landscapes that look like they are straight from the surface of some sort of plane of hell. Or furry flowers. Or fruit with eyes. I was expecting "lion in the Savannah" type pics, not "a visualisation of my DND group's "baleful polymorph" spell"...

Yes, whatever I type I get some abomination in the general settings I asked for (hills, mountains, ocean.) And I agree that the UX is also horrible. I want the webapp they are showing in the video, not the one they have online.

>I got a generic picture of the milky way for any prompt I tried ("rocks", "trees").

The picture is just really zoomed out. That's how awesome it is - it shows you ALL rocks and trees.

What if this is done on purpose? If they make the UX too easy then since this tech is so impressive it will be shared across the general population and quickly the service will get overloaded. This way only those that truly have the patience to fiddle with controls have the ability to get it working.

It looks like you have selected for it to use a segmentation mask, and not to use text.

I have no idea what that means, but the fact that it both told me to enter a text prompt and actually let me do it while not being in whatever magical mode it should have been in in order to actually use the text prompt is another point that can be added to my above rant.

Alright, I've uttered the incantation for it to do the thing. I still don't get it. [0] https://imgur.com/4zbaiH0

I also tried another example prompt, which bared a striking similarity to the previous result. I don't know if it's persisting the result (It shouldn't - I didn't click the re-use image button), but the strange life-raft looking artifact is very persistent. [1] https://imgur.com/KTCM4xH

Yeah, the text-to-image seems to be highly dependent on whether the generator knows how to generate the specific objects the text model thinks should be in the image. I got much more consistent results using the semantic segmentation drawing as input:

https://imgur.com/QC13zml

(and for what it's worth, you're totally right that the UI is just an absolute disaster)

The picture you drew and had it turn into rocks is actually really cool!

I think I would have been more generous to the project had I known it could do that. Maybe I X'd out of the frustrating tutorial too early? :)

Literally the first step after accepting the TOS:

https://i.imgur.com/oH4P1Xc.png

Yes but the selection is lost if you press enter out of habit. I had the same frustration until I realized what happened which turned that into anger :)

Maybe I'm just too old for tech demos.

Nah, the UI is not very intuitive, so there's plenty of blame to go around :-)

Wow, you weren't kidding. What an awful interface.

Hackernews has never gone soft on anything. Quite the opposite.

Yeah Hackernews is almost universally critical. Even his comment is needlessly critical. nVidia isn't poising this as a polished consumer app. It's a demo! The URL is https://www.nvidia.com/en-us/research/ai-demos/

For a research demo the UI is extremely polished. It even has an interactive tutorial!! Some people...

I worked in research in human computer interactions, this is a very poor UI even in a research setting. The technology is impressive though.

Your parent comment didn't mean that this was a good UI for human-computer interaction research, he meant that it was a good UI considering it was probably just a quick demo built by the scientists who did the research to showcase their work.

I work in a research group and most of my colleagues would have a hard making an interface like this one, no matter how confusing. They are ML/DL scientists and usually have had absolutely 0 exposure to frontend.

Yes not every researcher is good in frontend development. But my research group wouldn't promote such a broken UI until someone fix it at least a little bit.

<<If I'm meant to be impressed or fearful of "big data deepfake AI," this isn't it.

Let me put my tinfoil on. Maybe the idea is to normalize and make it less scary. Look at the press releases. The general response was not a horrified 'omg, this tech has gone berserk', but 'oh, its a benign lil biddy'.

Tinfoil off.

The interface is absolutely atrocious.

I bailed on it also, with the same experience.

A barrage of obtrusive prompts. When it wanted me to enter a phrase I typed "sunset over a field" and got a picture of the Milky Way galaxy also, followed by more prompts.

I then closed the tab and came here to read the comments.

It just did absolutely nothing for me, I typed in words and hit enter, just a green screen.

I'll say something constructive. While the text to image functionality is terrible, the segmentation based on arbitrarily drawn images is kind of fun. They should have just gone with the latter.

This is a demo piece not a paid app. Nvidia has zero interest in selling web apps. This is just a marketing stunt to show off their capabilities so they can sell hardware.

> I got a generic picture of the milky way for any prompt I tried You need to uncheck "segmentation" and check "text"

I don’t know how to make sense of the tos for projects like this. They all have clauses that let them modify their terms later for example

> If you press Enter in the prompt field it refreshes the page.

That's just default browser behaviour for forms.

I'm still clicking on the next button.

Reply on news.ycombinator.com