5 ms·

Here's a image->html test. Gemini has always swung above its weight class for vision work, so I'm always eager to try it with this. Original images: https://im

by jjcm 1mo ago

Here's a image->html test. Gemini has always swung above its weight class for vision work, so I'm always eager to try it with this. Original images: https://image.non.io/neonRamenDesigns.webp https://image.non.io/neonRamenDesigns.webp Gemini 3.7 build: https://html.non.io/neonRamenGemini3.7 https://html.non.io/neonRamenGemini3.7 Opus 5 build for comparison: https://html.non.io/neonRamen https://html.non.io/neonRamen Opus is still best in class for this, but it's worth noting how well Gemini 3.7 does vs a more comparable LLM price wise, which is Grok 4.6: https://html.non.io/neonRamenGrok4.6 https://html.non.io/neonRamenGrok4.6 . I thought Gemini would blow Grok out of the water (it generally has in the past), but Grok has really caught up.

snissn 1mo ago
I'm curious how much the harness plays into this. I'm somewhat surprised by the gemini and grok results, they seem to have strongly deviated from the original images. I'm thinking maybe the harness has a big effect? It's possible to proxy in different models to claude code, if you're curious you might find it interesting to test!
- jjcm 1mo ago
  Harness could be a part of it, but worth noting both the Opus and Gemini 3.7 flash tests were both ran through opencode. The grok test was ran through the cursor cli agent however.
  - igravious 1mo ago
    why Grok not through `Grok Build` ?
jjcm 1mo ago
Other thoughts: I really think Google has fallen behind here. Even as a high speed offering (this build took ~7min, which is pretty good!), it wont be able to claim dominance for long with cerebras announcing the Sol preview today: https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultrafast-with-openai https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultraf... . It's not a bad model by any means, but I just don't know what situation I'd reach for 3.7 Flash first for. Google really needs a differentiator, especially given how hard it is to get an API key from them. They can't be high friction and non-pareto.
- basch 1mo ago
  Depends on the definition of friction. If someone is in the Google ecosystem, why would they reach out of it.
  - Ardon 1mo ago
    I already use GCP and Google for work, and getting an API key was so annoying that even I couldn't be bothered after a while of looking around. Maybe things there have improved some, but when I was looking it was a huge runaround.
    - beart 1mo ago
      Hmm. My company has an internal portal for generating Gemini API keys. I select a project from a drop down, enter a name, and press okay.
      - cameronh90 1mo ago
        That may be evidence the built-in Google experience is difficult or confusing.
- MrBuddyCasino 1mo ago
  I like using 3.5-flash-lite for doing cheap PDF and Image data extraction stuff. I don't think there is a better bang / buck model right now (3.1 is cheaper but a lot worse).
  - wahnfrieden 1mo ago
    Probably cheaper to run a Mac Mini with VisionKit (private APIs if you need bounding rects).
  - oh_no 1mo ago
    5.6 Luna costs far less and benchmarks far better, have you compared for this task?
codazoda 1mo ago
How are you doing this with Opus. Clearly I’m missing something. I always turn to ChatGPT when I need images because Opus typically refuses. I’ve tried Claude Code and Claude online in the past. I’m pretty sure neither created images for me and I thought this was because Anthropic was focused on code. I guess I need to try harder. :)
- jjcm 1mo ago
  Images and build step were generated with my own tool (https://news.ycombinator.com/item?id=48995754 https://news.ycombinator.com/item?id=48995754 - it's why I'm often running these img->html tests). Opus can't generate images since A\ doesn't have a diffusion model.
  - victor106 1mo ago
    did you build your own diffusion model?
    - jjcm 1mo ago
      I have a few custom ones (a post-trained flux 2 checkpoint for web design and a image->metalness map generator that the build step can call for more advanced lighting situations), but gpt-image-2 is better than my own for design, so it's weighted much more heavily in outputs my tool generates. I think gpt-image-2 currently generates 99%+ of the design outputs on diffui
  - joshbee 1mo ago
    This is pretty interesting take on design and your trick to make the surface maps is pretty slick! I did have one question about the tool, is it possible to set how many variations you want per step? I would rather be able to guide it manually at some steps where maybe I know pretty well what I want or just need minor tweaks, and then let it loose on others when really trying to experiment with an idea.
- tyre 1mo ago
  I believe they are testing giving it an image, which you can do in Claude code by dragging/dropping into the terminal or copy/pasting, and asking it to build the html equivalent.
mediumdeviation 1mo ago
I'm not sure what prompt you put in but did Gemini replace the all of the images in the original with its own? That would be really weird behavior unprompted.
- jjcm 1mo ago
  The prompt is a build step generated by my tool for image->html conversion, which includes APIs the model can call to generate images/patterns/svgs. https://image.non.io/12275ee8-71e9-4941-823b-e51fec157b4d.webp https://image.non.io/12275ee8-71e9-4941-823b-e51fec157b4d.we... The agent is told to generate assets as part of the buildout. It gets to decide what the prompt is for them / whether to do postprocessing like background removal / what type of asset to generate.
flockonus 1mo ago
There is some irony being a developer and reading along the lines of: "oh look at the comparison between these models executing a task for a few cents on a job i'd be charging 1k minimum"
- bushbaba 1mo ago
  FYI, developers are rarely given such a rich UX mock.
  - flockonus 1mo ago
    Depends who you work with, what's the intention, budget, etc. I'd agree this is a really good one. I'm used to incremental Figma wireframe -> final product and working together with a designer.
  - stronglikedan 1mo ago
    I don't think I'd say rarely. Companies rarely allocate the design resources to produce that, but the companies that do are typically much larger, so the actual number of individual developers that get rich mocks is probably closer to 40-50%.
    - prepend 1mo ago
      I’d say rarely in the sense that out of the 500 times (not really that much over 30 years) I’ve built similar things, I got UX mockups this detailed maybe 10% pf the time.
- ex-aws-dude 1mo ago
  meh I have no interest in that type of software dev anyway
  - DrewADesign 1mo ago
    Unfortunately, as soon as they can’t find work, everybody interested in the more easily automated dev work will suddenly become very interested in up-skilling into all other kinds of dev work. So then you have an excess supply, which means little job security and littler salaries. Developers were in the cool kid club in SV because it was more painful to fill dev roles than to treat developers with kid gloves. Without high labor demand, there is no leverage for developers. Increasingly, management has the leverage. Oh well.
orliesaurus 1mo ago
I think both outputs are really good. I don't see a lot of differences. So what exactly should be looking at and notice that one model did worse or better than the other one. EDIT: OKAY I see it's mostly the "image" generation, not so much the HTML... Noticeable in the food photos and the foodtruck/cart photo
- basch 1mo ago
  clicking the add buttons and scrolling the menu is just much better in Opus 5. It feels like an actual website vs a simulation of one.
  - zuzululu 1mo ago
    i do feel like opus here is the most natural. theres something off about 3.7 and grok while its an improvement feels flat and not complete i do wonder why gpt sol was not compared here but honestly it's not really known to be the best at UI a fable 5 comparison would've been also interesting and likely the best.
XCSme 1mo ago
They both have horizontal scroll on mobile...
- jjcm 1mo ago
  Oh yea, as a disclaimer the models didn't have any instructions to do a mobile version. I haven't tested them on mobile at all.
xyzsparetimexyz 1mo ago
This has so much less character than the pelican smdh.. Plus is the ramen in HK even any good?
- jjcm 1mo ago
  I think they're both testing very different things. The pelican test is testing if a LLM can come up with visuals on its own via writing bezier curves directly. This is testing if it can match visuals that have already been established, and represent them with all the tools available to a web developer. The ramen example was chosen in particular because there are a lot of things that aren't easy to do with CSS, and require creative strategies: 45deg button cuts, angular repeating pattern elements, blending of raster art and svgs, microglyphs, low contrast subtle elements, etc. Don't ask yourself whether it's a good design, as yourself whether it's a good test.
t3hTao 1mo ago
[dead]
giarc 1mo ago
Can you share what your prompt was for that?
- jjcm 1mo ago
  The original prompt includes auth codes for API gen, but here's a redacted version: https://non.io/prompt-for-ramen https://non.io/prompt-for-ramen This is the build step generated by my diffusion-based ui tool's copy-for-agent action.
  - roysting 1mo ago
    [dead]
pphysch 1mo ago
Was the original concept generated by Claude somehow? It gives me Claude UI vibes with all the extraneous small-caps text elements.
joshmn 1mo ago
What’s the prompt you used for this? Edit: oh wow, diffui looks nice!
keyle 1mo ago
FYI Gemini's version is less broken than Opus' in Safari...
a2ff6eeb0 1mo ago
These look exactly like all of the low budget bodega signs near me. They also look like a bunch of cheap ads for parties that I keep seeing. The sameness of style is uncanny. (I don't have the Bodega signs, but I'm thinking of shit like this, from a quick google: https://linkstub.com/en/wet-wild-foam-party https://linkstub.com/en/wet-wild-foam-party)
- Barbing 1mo ago
  Sigh. You might be interested in this piece that came to mind from Doctorow: “Let me explain: on average, illustrators don't make any money. They are already one of the most immiserated, precarized groups of workers out there. They suffer from a pathology called "vocational awe." That's a term coined by the librarian Fobazi Ettarh, and it refers to workers who are vulnerable to workplace exploitation because they actually care about their jobs – nurses, librarians, teachers, and artists. If AI image generators put every illustrator working today out of a job, the resulting wage-bill savings would be undetectable as a proportion of all the costs associated with training and operating image-generators. The total wage bill for commercial illustrators is less than the kombucha bill for the company cafeteria at just one of Open AI's campuses. The purpose of AI art – and the story of AI art as a death-knell for artists – is to convince the broad public that AI is amazing and will do amazing things. It's to create buzz. Which is not to say that it's not disgusting that former OpenAI CTO Mira Murati told a conference audience that "some creative jobs shouldn't have been there in the first place," and that it's not especially disgusting that she and her colleagues boast about using the work of artists to ruin those artists' livelihoods.” https://pluralistic.net/2025/12/05/pop-that-bubble/ https://pluralistic.net/2025/12/05/pop-that-bubble/ OK so after all that, THIS (bad foam party posters) is what we get!
  - paulluuk 1mo ago
    I like this concept of "vocational awe", and I do think it's why, for example, there is so much sexual and financial abuse in the movie, music and game industry: those people are willing to be treated poorly so they can work their craft - do the thing they love. Basically they are trading away good working conditions in exchange for actually doing work you like or find meaningful, the opposite of someone who does something they hate or find meaningless but pays a high salary and where you are treated really well. I don't know how well this actually translates to AI, though. We all understand that AI art has a pretty low quality, but sometimes low quality is enough. If I want an image for my D&D character that only I and my DM will likely ever see, I am fine with a 7 cent AI-generated image, but I'm not willing to pay 150$ for an artist to do it - not that I don't value their time, I just don't value the image that much. Before AI this was the same, I'd just have used an image from Pinterest and thought "Well, this isn't exactly a good match, but I can't find anything better". But I assume real illustrators do things like illustrations for children's books? I'd like to believe that those are still done by actual people, not AI.
jzemeocala 1mo ago
One of my favorite image tests with AI models is schematic analysis...I build and repair tube amps for a living, and use AI for such work a LOT. so far, IMHO, the best has been opus and fable\mythos.
- fumeux_fume 1mo ago
  Nice! I test if models know which tubes I can use for an amp given the power and number of pins.
cechmaster 1mo ago
I'm not sure what you consider good design, but if it's subjective, then I see it differently from your examples. Gemini 3.7 looks the best. Opus 5 looks almost as good as Gemini. Grok 4.6 looks pretty terrible.
getnormality 1mo ago
IMO Gemini's is better than all the others, including the original.
butlike 1mo ago
There's one specific aspect I like better with the grok version: The prices are above the fold. ON the Opus and Gemini versions, I have to scroll to see the full menu item showcase.
aagha 1mo ago
How are you prompting it with the original images to create such sites?

by jjcm 1mo ago

snissn 1mo ago

I'm curious how much the harness plays into this. I'm somewhat surprised by the gemini and grok results, they seem to have strongly deviated from the original images. I'm thinking maybe the harness has a big effect? It's possible to proxy in different models to claude code, if you're curious you might find it interesting to test!

jjcm 1mo ago
Harness could be a part of it, but worth noting both the Opus and Gemini 3.7 flash tests were both ran through opencode. The grok test was ran through the cursor cli agent however.
- igravious 1mo ago
  why Grok not through `Grok Build` ?

jjcm 1mo ago

Other thoughts: I really think Google has fallen behind here. Even as a high speed offering (this build took ~7min, which is pretty good!), it wont be able to claim dominance for long with cerebras announcing the Sol preview today: https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultrafast-with-openai https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultraf... . It's not a bad model by any means, but I just don't know what situation I'd reach for 3.7 Flash first for. Google really needs a differentiator, especially given how hard it is to get an API key from them. They can't be high friction and non-pareto.

basch 1mo ago
Depends on the definition of friction. If someone is in the Google ecosystem, why would they reach out of it.
- Ardon 1mo ago
  I already use GCP and Google for work, and getting an API key was so annoying that even I couldn't be bothered after a while of looking around. Maybe things there have improved some, but when I was looking it was a huge runaround.
  - beart 1mo ago
    Hmm. My company has an internal portal for generating Gemini API keys. I select a project from a drop down, enter a name, and press okay.
    - cameronh90 1mo ago
      That may be evidence the built-in Google experience is difficult or confusing.
MrBuddyCasino 1mo ago
I like using 3.5-flash-lite for doing cheap PDF and Image data extraction stuff. I don't think there is a better bang / buck model right now (3.1 is cheaper but a lot worse).
- wahnfrieden 1mo ago
  Probably cheaper to run a Mac Mini with VisionKit (private APIs if you need bounding rects).
- oh_no 1mo ago
  5.6 Luna costs far less and benchmarks far better, have you compared for this task?

codazoda 1mo ago

How are you doing this with Opus. Clearly I’m missing something. I always turn to ChatGPT when I need images because Opus typically refuses. I’ve tried Claude Code and Claude online in the past. I’m pretty sure neither created images for me and I thought this was because Anthropic was focused on code. I guess I need to try harder. :)

jjcm 1mo ago
Images and build step were generated with my own tool (https://news.ycombinator.com/item?id=48995754 https://news.ycombinator.com/item?id=48995754 - it's why I'm often running these img->html tests). Opus can't generate images since A\ doesn't have a diffusion model.
- victor106 1mo ago
  did you build your own diffusion model?
  - jjcm 1mo ago
    I have a few custom ones (a post-trained flux 2 checkpoint for web design and a image->metalness map generator that the build step can call for more advanced lighting situations), but gpt-image-2 is better than my own for design, so it's weighted much more heavily in outputs my tool generates. I think gpt-image-2 currently generates 99%+ of the design outputs on diffui
- joshbee 1mo ago
  This is pretty interesting take on design and your trick to make the surface maps is pretty slick! I did have one question about the tool, is it possible to set how many variations you want per step? I would rather be able to guide it manually at some steps where maybe I know pretty well what I want or just need minor tweaks, and then let it loose on others when really trying to experiment with an idea.
tyre 1mo ago
I believe they are testing giving it an image, which you can do in Claude code by dragging/dropping into the terminal or copy/pasting, and asking it to build the html equivalent.

mediumdeviation 1mo ago

I'm not sure what prompt you put in but did Gemini replace the all of the images in the original with its own? That would be really weird behavior unprompted.

jjcm 1mo ago
The prompt is a build step generated by my tool for image->html conversion, which includes APIs the model can call to generate images/patterns/svgs. https://image.non.io/12275ee8-71e9-4941-823b-e51fec157b4d.webp https://image.non.io/12275ee8-71e9-4941-823b-e51fec157b4d.we... The agent is told to generate assets as part of the buildout. It gets to decide what the prompt is for them / whether to do postprocessing like background removal / what type of asset to generate.

flockonus 1mo ago

There is some irony being a developer and reading along the lines of: "oh look at the comparison between these models executing a task for a few cents on a job i'd be charging 1k minimum"

bushbaba 1mo ago
FYI, developers are rarely given such a rich UX mock.
- flockonus 1mo ago
  Depends who you work with, what's the intention, budget, etc. I'd agree this is a really good one. I'm used to incremental Figma wireframe -> final product and working together with a designer.
- stronglikedan 1mo ago
  I don't think I'd say rarely. Companies rarely allocate the design resources to produce that, but the companies that do are typically much larger, so the actual number of individual developers that get rich mocks is probably closer to 40-50%.
  - prepend 1mo ago
    I’d say rarely in the sense that out of the 500 times (not really that much over 30 years) I’ve built similar things, I got UX mockups this detailed maybe 10% pf the time.
ex-aws-dude 1mo ago
meh I have no interest in that type of software dev anyway
- DrewADesign 1mo ago
  Unfortunately, as soon as they can’t find work, everybody interested in the more easily automated dev work will suddenly become very interested in up-skilling into all other kinds of dev work. So then you have an excess supply, which means little job security and littler salaries. Developers were in the cool kid club in SV because it was more painful to fill dev roles than to treat developers with kid gloves. Without high labor demand, there is no leverage for developers. Increasingly, management has the leverage. Oh well.

orliesaurus 1mo ago

I think both outputs are really good. I don't see a lot of differences. So what exactly should be looking at and notice that one model did worse or better than the other one. EDIT: OKAY I see it's mostly the "image" generation, not so much the HTML... Noticeable in the food photos and the foodtruck/cart photo

basch 1mo ago
clicking the add buttons and scrolling the menu is just much better in Opus 5. It feels like an actual website vs a simulation of one.
- zuzululu 1mo ago
  i do feel like opus here is the most natural. theres something off about 3.7 and grok while its an improvement feels flat and not complete i do wonder why gpt sol was not compared here but honestly it's not really known to be the best at UI a fable 5 comparison would've been also interesting and likely the best.

XCSme 1mo ago

They both have horizontal scroll on mobile...

jjcm 1mo ago
Oh yea, as a disclaimer the models didn't have any instructions to do a mobile version. I haven't tested them on mobile at all.

xyzsparetimexyz 1mo ago

This has so much less character than the pelican smdh.. Plus is the ramen in HK even any good?

jjcm 1mo ago
I think they're both testing very different things. The pelican test is testing if a LLM can come up with visuals on its own via writing bezier curves directly. This is testing if it can match visuals that have already been established, and represent them with all the tools available to a web developer. The ramen example was chosen in particular because there are a lot of things that aren't easy to do with CSS, and require creative strategies: 45deg button cuts, angular repeating pattern elements, blending of raster art and svgs, microglyphs, low contrast subtle elements, etc. Don't ask yourself whether it's a good design, as yourself whether it's a good test.

t3hTao 1mo ago

[dead]

giarc 1mo ago

Can you share what your prompt was for that?

jjcm 1mo ago
The original prompt includes auth codes for API gen, but here's a redacted version: https://non.io/prompt-for-ramen https://non.io/prompt-for-ramen This is the build step generated by my diffusion-based ui tool's copy-for-agent action.
- roysting 1mo ago
  [dead]

pphysch 1mo ago

Was the original concept generated by Claude somehow? It gives me Claude UI vibes with all the extraneous small-caps text elements.

joshmn 1mo ago

What’s the prompt you used for this? Edit: oh wow, diffui looks nice!

keyle 1mo ago

FYI Gemini's version is less broken than Opus' in Safari...

a2ff6eeb0 1mo ago

These look exactly like all of the low budget bodega signs near me. They also look like a bunch of cheap ads for parties that I keep seeing. The sameness of style is uncanny. (I don't have the Bodega signs, but I'm thinking of shit like this, from a quick google: https://linkstub.com/en/wet-wild-foam-party https://linkstub.com/en/wet-wild-foam-party)

Barbing 1mo ago
Sigh. You might be interested in this piece that came to mind from Doctorow: “Let me explain: on average, illustrators don't make any money. They are already one of the most immiserated, precarized groups of workers out there. They suffer from a pathology called "vocational awe." That's a term coined by the librarian Fobazi Ettarh, and it refers to workers who are vulnerable to workplace exploitation because they actually care about their jobs – nurses, librarians, teachers, and artists. If AI image generators put every illustrator working today out of a job, the resulting wage-bill savings would be undetectable as a proportion of all the costs associated with training and operating image-generators. The total wage bill for commercial illustrators is less than the kombucha bill for the company cafeteria at just one of Open AI's campuses. The purpose of AI art – and the story of AI art as a death-knell for artists – is to convince the broad public that AI is amazing and will do amazing things. It's to create buzz. Which is not to say that it's not disgusting that former OpenAI CTO Mira Murati told a conference audience that "some creative jobs shouldn't have been there in the first place," and that it's not especially disgusting that she and her colleagues boast about using the work of artists to ruin those artists' livelihoods.” https://pluralistic.net/2025/12/05/pop-that-bubble/ https://pluralistic.net/2025/12/05/pop-that-bubble/ OK so after all that, THIS (bad foam party posters) is what we get!
- paulluuk 1mo ago
  I like this concept of "vocational awe", and I do think it's why, for example, there is so much sexual and financial abuse in the movie, music and game industry: those people are willing to be treated poorly so they can work their craft - do the thing they love. Basically they are trading away good working conditions in exchange for actually doing work you like or find meaningful, the opposite of someone who does something they hate or find meaningless but pays a high salary and where you are treated really well. I don't know how well this actually translates to AI, though. We all understand that AI art has a pretty low quality, but sometimes low quality is enough. If I want an image for my D&D character that only I and my DM will likely ever see, I am fine with a 7 cent AI-generated image, but I'm not willing to pay 150$ for an artist to do it - not that I don't value their time, I just don't value the image that much. Before AI this was the same, I'd just have used an image from Pinterest and thought "Well, this isn't exactly a good match, but I can't find anything better". But I assume real illustrators do things like illustrations for children's books? I'd like to believe that those are still done by actual people, not AI.

jzemeocala 1mo ago

One of my favorite image tests with AI models is schematic analysis...I build and repair tube amps for a living, and use AI for such work a LOT. so far, IMHO, the best has been opus and fable\mythos.

fumeux_fume 1mo ago
Nice! I test if models know which tubes I can use for an amp given the power and number of pins.

cechmaster 1mo ago

I'm not sure what you consider good design, but if it's subjective, then I see it differently from your examples. Gemini 3.7 looks the best. Opus 5 looks almost as good as Gemini. Grok 4.6 looks pretty terrible.

getnormality 1mo ago

IMO Gemini's is better than all the others, including the original.

butlike 1mo ago

There's one specific aspect I like better with the grok version: The prices are above the fold. ON the Opus and Gemini versions, I have to scroll to see the full menu item showcase.

aagha 1mo ago

How are you prompting it with the original images to create such sites?