"Drawing" the Mona Lisa with GPT-5.6, Claude, Gemini, and Grok

https://www.tryai.dev/blog/ai-drawing-arena-colored-pencils-claude-gpt-grok

by hershyb_ • 1 day ago

245 102 1 day ago

tryai.dev

ms7892• 1 day ago

Fix this signup error please

https://www.tryai.dev/?error=server_error&error_code=unexpec...

pavel_lishin• 1 day ago

I think this is just an ad.

noodlescb• 1 day ago

If it is, it's a horrible ad. My main takeaway is that all of them were god awful at image generation and Fable was 20x more expensive and still awful.

Legend24402• 1 day ago

Well, your takeaway is bad. This isn't image generation, it's tool use with a digital paintbrush.

KolinFirz• 1 day ago

Tell them to draw LeBron James or Heisenberg. Everyone will refuse.

geroge_kyaw• 1 day ago

Useless They are not image generation models.

samuelknight• 1 day ago

Not useless. LLMs are the most general purpose computer algorithms ever created. They are getting smarter and cheaper at a geometric rate. What is a bad idea today could have useful applications tomorrow.

munk-a• 1 day ago

cheaper at a geometric rate

Citation needed - my company is paying more than ever for code generation. I have no reason to believe (given anecdotes) that anyone finds themselves in the opposite situation.

Legend24402• 1 day ago

That's what makes it an interesting challenge.

ksd482• 1 day ago

Grok! LOL!

Seriously, what's going on there ? Why is it so different from others? Is it just behind technologically/training wise or it's using something fundamentally different?

Ekshef• 1 day ago

Did Elon tell some poor engineer to give grok a prompt injection for drawing,"make it look like one of my childhood drawings!" Just like the Tesla truck?

jorl173• 24 hours ago

Grok 4.5 is...something else.

It performs much better than composer2.5 (while being as fast). It's not Opus, but I think it's not that far off. Definitely better than sonnet for what I've been doing.

On the other hand, I think they probably heavily adapted the training data so that it really is extremely focused on code. I just recently ran my personal "poetry benchmark" on it (where I give it ~850 poems I've written over my life and ask it to comment the corpus as a whole), and it's whack. It tries to write in portuguese (most of the poems are portuguese) and code-switches constantly and mixes up words to the point of making what it writes almost unreadable (e.g. it writes stuff like "You can't QoS that that look for beast poems", in portuguese, all messed up). The quality of the analysis is also quite bad (I'd say it's definitely behind Sonnet).

So I really think they either threw away data that wasn't tied to coding so that they could fine-tune it to that, or somehow they've got such an unbalanced dataset that coding ends up dominating either way. To me, its disastrous performance in this drawing "competition" fits this narrative.

userbinator2• 22 hours ago

"You can't QoS that" sounds like the title of a nerdcore rap song.

giancarlostoro• 24 hours ago

They focused a little too much on Grok Imagine.

eth0up6• 23 hours ago

The razor-wire at the bottom for Starry Night was clever, and very Grok. Really shows its military spirit.

Edit: I just don't see the point of redacting the Mona Lisa

yojo• 23 hours ago

I read the balls as “houses”, though the phallic spire emerging from them dead-center is also very on-brand

vidarh4• 14 hours ago

Redacting? Is that not just the model using the smudge tool it was given.

visiondude• 22 hours ago

the starry night one is soo funny

eth0up6• 24 hours ago

Claude was clearly 'pushing back' on the coziness of the cabin. But I think it did best with the cat. Grok, I fear, is making a case for euthanasia. It's suffering and I think it would be cruel to let it continue. Someone pull the plug. .

AlienRobot• 24 hours ago

The difference in cost is pretty incredible.

elcritch3• 22 hours ago

Rather the difference of cost of Claude is pretty incredible. Then again, I guess they gotta make it rain while they can.

usef-• 15 hours ago

He mentions they used quite different methods to get to the end-result. I'd love to know how much changing the prompt impacts things: like telling claude to only draw and blend, similar to sol, etc.

There's likely not training data for this and Claude seemed to spend a lot of time blending the background.

NichoPaolucci2• 24 hours ago

As I looked through the images I was unimpressed entirely, at first. But, then I started thinking, these look a little... "childish" to me.

Childish as in... A newish artist who is drawing a concept rather than light / forms (Which is something artists typically do as they understand drawing more and more).

The rose in the vase specifically - some models understood that there was supposed to be shading, reflections, the concept of refraction - others just drew "blue = glass" and "green = stem" and "red = rose".

Really odd to look at, considering if I saw any of these drawings from a human kid, I would say "good job buddy" and put it on the fridge. I'm expecting these to get better as models improve, and perhaps the artistic progression will be there along with it...

ralusek• 23 hours ago

What's interesting is that the way in which they're childish is actually extremely human. In fact, one of the ways that you're often taught to draw more realistically is to stop thinking of the concepts as icons you're drawing the outlines of, and instead sort of blur your eyes and see things as they are: hues and values. In other words, become a camera or a printer that has no idea what it's capturing or printing other than a grid of values. That is how you achieve realism.

The fact that it has clearly iconified these concepts in its mind and is tracing the outlines of the things it thinks/expects to go where is very human.

cj4• 23 hours ago

Parrots can also sound extremely human, but it’s only mimicry.

How do we know the difference?

NichoPaolucci2• 6 hours ago

Yes - precisely what I was getting at with the childish comment.

It felt a bit surreal to see a machines rendition of something that very closely maps to a younger human as they explore + understand more about the world. Scary, even (to me). This was the first time I’ve seen LLM output and thought “wow, maybe it is learning.”

userbinator2• 22 hours ago

Really odd to look at, considering if I saw any of these drawings from a human kid, I would say "good job buddy" and put it on the fridge.

The Grok ones in particular gave me that thought. Most of them really look like what a kid would do when given the same tools, while the other models' output have distinctly more "AI-ness" to them (for lack of a better term.)

MarcScott• 14 hours ago

Grok consistently draws images that, if I were its parent, I would refuse to put on the fridge door.

jvanderbot• 9 hours ago

I call this "symbol drawing". Beginning artists do this. They think "This is a head, a head is round. This is where eyes go, eyes are shaped like this", and the whole thing ends up being a collage of symbols vs a representation of the space and experience of viewing a face. When I used to be OK at drawing, it was because I forced myself to use touch instead of sight to compose images. So weird to explain, but I'd feel the 3d to get the lighting and such better.

A lot of art that someone smarter than me told me to appreciate seems to follow the pattern of hitting the space and/or experience while minimizing the use of symbols. Impressionistic paintings esp avoid symbols IMHO, while bizzaro picassos abuse symbols outright and still hit the experience they are going for.

danesparza• 9 hours ago

I took the same prompts to Gemini and was stunned by the results. The are completely different from the images shown in the article (and genuinely good art pieces that were generated).

jstanley• 9 hours ago

Were you using the Gemini image generator though? The article is about using LLMs to do drawing via tool calls.

taf2• 23 hours ago

Would love to see the results if they had used /goal or similar

SXX• 23 hours ago

Now add Deepseek, GLM and Kimi :-)

felixfurtak• 21 hours ago

I just tried with Deepseek and got this for the Mona Lisa

           .-""-.
          /      \
         |  _  _  |
         | (o)(o) |
         \   /\   /
          |  --  |
          |  \/  |
          |      |
         /   --   \
        /  /    \  \
       /  /      \  \
      (__/        \__)

It might not have been the most scientific test

fph• 10 hours ago

Looks more like Munch's The Scream to me.

ahoka• 13 hours ago

Any of those would be more relevant than Grok.

tylerrobinson• 23 hours ago

The Grok ones are so weird that they cross into uncanny and surreal. Truly bizarre and capable of eliciting feelings from me, if only bad feelings…

Razengan• 23 hours ago

Now this is a cool test. Sol's the best in each one, while Grok looks like the work of that person who "fixed" that Jesus painting..

LastTrain• 22 hours ago

Why include Grok when it is clearly not even in the same category as the other three?

dizzard• 22 hours ago

I wonder how much better a harness could get for drawing

thomasfromcdnjs• 20 hours ago

I think a lot better

- the best image related stuff I've seen is where the harness is constantly cropping and looking closer at things (likely helps a lot for computer vision in general) - also it would be interesting because the harness could almost have its own "palette" as if it could play with blank squares and different strokes over lapping or blending before applying to the main canvas

aqfamnzc• 7 hours ago

I'm surprised that the harness wasn't better described here. Is the draw tool really just asking LLMs to provide a set of coordinates/parameters for generating shapes? Honestly, if so, the results are shockingly good. Think of how well a typical human artist would do with such a crude method of applying brushstrokes.

fwip• 22 hours ago

Article is AI-written.

hombre_fatal2• 21 hours ago

GPT-5.6 Sol is the knock out here. Some of those results are really human/charming.

The rectangular smudge tool is a weird tool in the first place, but it's cute to see the models try to use it.

js82• 19 hours ago

I really like the cabin from Sol. It could be an art piece.

hombre_fatal2• 7 hours ago

Its rose is also insane. Not just the shading of the flower but its clever use of the smudge tool to soften the background and then soften the vase so it looked more like glass.

And I like how its starry night decided to accentuate the church/steeple in the village backdrop.

Pretty cool how it did these with a scrappy little tool palette. What a fun idea.

jnathsf• 21 hours ago

GPT 5.6 Sol had the best two drawings (rose and starry nights) but even more impressive was how efficient it was RE cost/time/tokens vs Fable (3.4M vs 14.6M / $7.74 vs $161!). OpenAI has quietly innovated around inference - this is will be a growing differentiator even against open models.

lubujackson• 21 hours ago

I do think they are going to stretch their lead in value if Anthropic doesn't wake up and stop YOLOing tokens. Kimi is an amazing achievement, but it has the same (or worse) kitchen sink approach as Fable.

At work, even if Fable is technically better I much prefer Sol because it is so much faster and concise.

theplumber• 20 hours ago

To me GPT is quite slow. I use it with max reasoning though but Fable at max as well

reactordev• 21 hours ago

Should have compared against FLUX Klein and Z-Image Turbo...

maxall4• 20 hours ago

I very much enjoyed Grok 4.5's rendition of the Mona Lisa as Elon Musk with tentacles.

bdcravens• 20 hours ago

The Grok ones are amusing, almost comically bad. However, whenever I've tried to pass an image creation request to any of the Opus models, it's been far worse, like first week of using Microsoft Paint bad (while ChatGPT would create social media quality images using the same prompts)

vidarh4• 14 hours ago

Opus doesn't have access to an image model, while ChatGPT does (it's not multi-modal, but it can prompt OpenAI's image model). So Opus by default is forced to write code to generate images, or generate SVG's as text. That's putting it in a worse situation than the linked article, where all the models were given rudimentary tools.

Give Opus access to an image model, and it will use that to create images just fine just like ChatGPT.

MitziMoto• 19 hours ago

Grok's look like a truly disturbed child. Like the drawings from that kid in "The Ring".

wolttam• 18 hours ago

Would be great to see Kimi here.

roguedemon• 17 hours ago

That was a good testcase. Pretty cool.

sashank_1509• 16 hours ago

Where do these models even have this data to learn from. There must be massive computer use datasets? I have a hard time believing it’s emergent if it’s able to do something this good.

fastball• 16 hours ago

The most interesting result for me is that they apparently prompted the models to optimize for SSIM, but many of the models trend worse over time. I suppose because viewing the canvas always comes after drawing, and they didn't give "revert to previous" capability as part of the toolkit.

Which in turn kinda jives with my experience of using these models for code: to some extent they only seem to have a concept of "forward", which invariably leads to "write more code to fix previous problems created", rather than taking a step back and removing broken things entirely.

d1sxeyes• 15 hours ago

Stopping problem at work. I wonder if results would improve if we gave the agents an “undo” or “revert” tool.

NiloCK• 16 hours ago

For capabilities reference:

I made a lower effort but similar scaffold for LLMs to do iterative drawing in Nov 2024, with Sonnet 3.5 as the artist: https://paritybits.me/llm-drawing-with-eyes-open/

Quite a difference.

Athanase000• 15 hours ago

The way the LLM describes its own work at the end is hilarious.

mbmbn• 15 hours ago

Shows about the same level of technical mastery that present days painters.

tulio_ribeiro• 15 hours ago

Weird choice of SSIM/RMSE. By feeding these back into the model, the agent is actively degrading its artistic output. This is only valid for the target reference I think?

Alternatively, I think much better results could be had by computing the cosine similarly using DINOv2 ONNX (@xenova/transformers).

My guess is that the results are greatly limited by relying on the current metrics.

Even better, attach toModelOutput directly to drawTool so it returns the rendered canvas image immediately upon drawing.

leobg• 15 hours ago

I'm missing some context here.

First of all, this is a marketing piece, made for engagement.

Second, in production, you would adjust your approach to the model's strengths and weaknesses. For example, if you found that a model does not respond to the tools you are offering, then you would find a way to make it more responsive. Also, we know that different models handle prompts and system prompts differently. So comparing several models from different providers against exactly the same prompt is quite a naive way of exploring the capabilities of those models or comparing them to each other. You may have just happened to “speak the language“ of one model better than that of another. Why would that be the model’s fault and not yours?

Third, none of these results mean anything if you do not run each process on each model several times to check for consistency. Your model might just have gotten lucky. This point is even more true when you're talking about a progressive task where one step builds on the previous one. The model might just have painted itself into a corner.

traes• 14 hours ago

Perhaps there was something in the prompt or settings preventing this, but I'm surprised (and slightly disappointed) that none of the models approached this the way I would: download an image of the Mona Lisa, run all of the drawing functions many times to construct a forward model of the drawing implements, and attempt to solve some kind of explicit inverse problem through either ML or a classical algorithm to minimize some difference metric. Were they just restricted from running code or are the models unindustrious without a very specific prompt?

bibin765• 8 hours ago

A few months back, we had a use case to create some SVGs and added this skill: https://github.com/upbrew-tech/svg-creator-skill

I think using this skill, Claude drew it better than the examples given. https://claude.ai/share/bc0867dc-034f-4865-bb71-acb83bfa36a1

raybb• 8 hours ago

A year or so ago I was trying to get llms to make pixel art by asking them to create a 2d array of color codes. I'm no prompt guru but they all failed at anything beyond a basic shape. I tried at many resolutions from 16x16, 32x32, 64x64, and maybe a little larger but basically only then quit small ones were even recognizeable.

vunderba• 8 hours ago

There’s a related benchmark along these lines that uses an LLM with a discrete “instructive graphic language” to get models to generate Minecraft voxel structures with varying degrees of success.

https://github.com/Ammaar-Alam/minebench

dehrmann• 7 hours ago

This is picking the wrong tool for the job twice over. Half of it is seeing if text-to-image mechanics exist in general purpose LLMs, the other half is seeing if they can reason their way about generating the 2d array. Imagine asking this of yourself and doing it through just a train of thought. Yes, I could probably draw a basic smiley face in the color of your choice in a 16x16 without antialiasing in 15 minutes.

"Drawing" the Mona Lisa with GPT-5.6, Claude, Gemini, and Grok

Related Stories