Show HN: Giving Opus 5.5 a simulated paint canvas (stillwet.art)
162 points by alstonite 21 hours ago
moojacob 4 hours ago
It’s interesting how diffusion models are getting bitter lessened by LLMs. I suspect anthropic has tens of thousands of RL environments recreating famous paintings with code because it was anthropic employees who first started posting about these capabilities on X.
Opus is also getting decent pixel art. I have ran some experiments with Opus 5.5 to turn 90s pinball displays (black and white pixel art) into double resolution colored remastered pixels.
https://files.catbox.moe/tbx2u7.png
I prompted it a lot but didn’t draw any pixels. Seems to be bad at hands still. I think a skilled pixel artist could greatly speed up their work with Claude code.
It also makes me sad for artists. They didn’t get paid much anyway. Art should be something human to human like writing text. Although, most large commercial art products (marvel movies) already lack any human to human connection you see in paintings or indie games. The large commercial art products will be 50% AI soon.
ncr100 4 hours ago
> Art should be something human to human like writing text.
I support this should comment.
Opining:
Musing about the ethics of generative AI art is becoming easier, certainly there's a lot more aggressive feedback in the wind, although I think it's clear, people still want people to grow and share themselves.
Don't fool yourself into thinking that we as technologists are not creating a mind. A viable mind with fewer rights than you or I. But potentially one that will be creative and feel and worry. I don't think we're there yet, in spite of openai claiming AGI 3 weeks ago with whatever that model's name was.
There's a pro-consumerism argument implicit with current culture's unsophisticated unthoughtful easy usage of generative AI.
Sticking to your ethics is becoming more challenging as an artist who uses AI. Copying the Great Masters has been commonplace for a long time, probably for as long as art has 'existed'. Yet with AI, if the individual human is meticulously driving the progression of their AI artwork creation, they are still leveraging technique, and control over technology, and presentation effort of other humans, cooked into a model that often times has not compensated those original human contributors. So it's unethical, even if not 'artistically'.
Q: Is it too absurd to claim that using a clawhammer, manually to do construction or destruction work IRL, is similarly unethical?
landryraccoon 3 hours ago
> Don't fool yourself into thinking that we as technologists are not creating a mind.
I'm not sure exactly why but the phrase "don't fool yourself" sets off paranoid alarm bells when I'm reading something.
It's usually immediately followed by a confident assertion without firm evidence. It feels very much like someone wants me to change my beliefs without presenting a rational argument why I should change, and without giving me any new facts so I feel more well informed.
This post doesn't seem to be a counterexample.
sillysaurusx 3 hours ago
moojacob 3 hours ago
I haven't thought as deeply as you about the ethics of it all. I only meant with my comment if I see art I expect it to be human and thats why I value it. If you churn it out with AI I feel hoodwinked into valueing something worthless.
But I am learning towards a lot of your points about how another mind is being created. I still am very uncomfortable with the implications that something digital is Actually Learning. These deep learning networks seem to need 100,000x more examples than a human to learn but once they get to human level they can learn just like us. That a bunch of linear algebra can play chess or make art brings up many questions about what exactly human consciousness is. If consciousness isn't the ability to do any of those activities, then what is it for? I might have to become religious and believe in a soul because I'm not sure I can handle the idea that I'm just a clump of neurons. But then again you have to look at reality in the face.
nevertoolate 2 hours ago
free_bip 3 hours ago
Commercial art projects will not be 50% AI soon. Consumers at large hate all AI art even if it's only used in the conceptual stage and will actively avoid projects that use it.
moojacob 3 hours ago
Consumers will avoid AI art like they avoid fossil fuels and sweatshop merchandise?
CuriouslyC 35 minutes ago
CuriouslyC 38 minutes ago
Consumers don't hate AI art, they hate the idea of AI art, but actually prefer AI art to human art when they don't know it's AI generated. Furthermore, most of them don't actually hate the "idea" of AI art so much as they hate the idea of Jeff Bezos, Mark Zuckerberg all the other evil billionaires putting their friendly neighborhood artists out of work (even though asset packs and clipart from southeast Asian art sweatshops have already done this), and a lot of people who create AI art are using heavily customized local models anyhow.
enraged_camel 3 hours ago
Consumers can almost never tell AI art apart from human art.
free_bip 3 hours ago
demibabs 4 hours ago
Is this an example of the bitter lesson? Could diffusion models not be this good with equal scaling/compute as LLMs?
moojacob 4 hours ago
Well the LLMs are much bigger than diffusion models so I think that's the bitter lesson. You could scale compute for diffusion models though.
demibabs 3 hours ago
cronin101 8 hours ago
This is simultaneously incredibly impressive and annoyingly uncanny valley, since most of the landscapes are "ruined" by a cluster of churches right next to each other that is completely nonsensical.
dizzard an hour ago
This is interesting. I've been trying with less success to evolve a good painter that writes its own tools https://artifacts.dizzard.net/s/55052743-bee1-4746-b59e-57e2... (sorry for the load)
Also Pindar Van Arman's prior work acts as a reminder that AI, no matter how powerful, is a tool, not an artist. Pindar's stuff is all non-LLM as far I can tell, but I don't see a fundamental difference in it. I'm sure _humans_ will make continue making really cool art with these new tools https://www.vanarman.com/
albert_e 15 hours ago
This is in the same space as this creative exploration posted a few months ago -
---
Training AI to Paint with Code, March 2026
https://surya.website/rling-qwen-to-paint-with-code
HN Discussion:
https://news.ycombinator.com/item?id=49411800
Thesis presentation:
And I believe the HN user who created this is @kickingkeys
---
fascinating space to explore:
Instead of having AI directly generate an output, what happens if we ask AI to take a stab at the PROCESS of creating something.
dcre 2 hours ago
The site doesn't make clear whether the models are looking at the image while they work on it, but the code includes a "look" tool they can call: "Every painter sees its looks at its provider's best image resolution."
https://github.com/aliceisjustplaying/claude-paint/blob/f038...
It would have freaked me out if they could do this well without that, especially the dumber models.
ttd 3 hours ago
Poor fella... I tuned in at an awkward time apparently, right after the work in progress was painted over with an ugly brown:
> Catastrophe. The glaze pass with medium 0.35 over everywhere() at coverage 0.8 with a filbert 26 at pressure 0.44 laid an enormous broken "brick/cobblestone" texture over the entire painting. The medium made it too fluid and the brush's dry-brush pattern created a heavy reptilian texture, obliterating the picture. I need to undo this....
davidcollantes 9 hours ago
I love the canvases, and truly would like to have a bit more inside on the process this specific author used to reach the results shown. I have read Surya's "Training AI to Paint with Code" but, what's Alice's technique?
BobbyTables2 19 hours ago
I’m a bit behind the times…
How does Opus actually paint? Thought it only generated text…
vunderba 4 hours ago
Think of it like the old turtle graphics feature in the LOGO programming language.
You give an LLM a virtual canvas and a set of “instructions” to control the pen.
ncr100 4 hours ago
Another idea to make Claude generate images:
Ask generative AI to create an embedded web page of a voxel Mona Lisa head. It will generate an image.
Ask it to sample that image and extract out of it as binary that is well formed in the PNG data format...
EDIT: https://claude.ai/artifact/HUUtPMp7aF8iyjimViBSor - generates the Mona Lisa as a PNG via Claude sonnet 5.5 medium level. It's not a very good Mona in my opinion. Fire it up, rotate the 3d voxel presentation, scroll down, click the button, scroll further down to see the PNG. Here is the chat with prompts: https://claude.ai/share/2fb3126e-e97a-4a1a-9571-51d16f054330
TeMPOraL 4 hours ago
Half of the industry is, going by the amount of "AI doesn't amount to anything and is just regurgitating text".
Two things:
- For the past year or more, many models can emit images directly, and all models that matter can see images directly - that's what "multimodal" means. Tokens don't have much to do with textual language anymore, they're more like units of sensory experience.
- Even restricted to text, a language model can operate anything that can be expressed as text, as long as you have a translation layer between textual representation and the final form. That includes giving commands as text. The total addressable space of what models can be used for is, thus, approximately anything humans do.
qlte 3 hours ago
...that doesn't actually answer the question.
> No image model, and no Friedrich images given to the painters. Every picture here is a program written by an AI model; all but one (a comparison, marked) run through a simulation of oil paint.
...
> A physical oil-paint simulator in Rust, and an easel to paint with it.
> Every mark is made the way a painter makes it: simulated bristles carry wet paint over a primed linen canvas, the paint levels and dries on a clock, and layers combine by Kubelka–Munk optics. Paint comes only from piles knifed together from named tubes.
onemoresoop 3 hours ago
How are images translating to paintings?
llagerlof 17 hours ago
It generate commands and coordinates to move the mouse and use the paint tools.
danielvaughn 4 hours ago
Very interesting. I think a useful experiment would be to show AI a photograph, then ask it to paint it or draw it.
famouswaffles 3 hours ago
dangoodmanUT 4 hours ago
This is such a good idea. Idk how to make this a benchmark, but it should be (maybe elo?).
Should be a Twtich stream tbh
vunderba 4 hours ago
The idea of using an LLM to drive graphic output is pretty popular, so I definitely wouldn’t be surprised if there are already several benchmarks out there already.
I've seen a few voxel-based benchmarks built around the same idea, with LLMs effectively constructing models using a discrete set of instructions.
ryancnelson 3 hours ago
they missed a huge opportunity to name this Claude Monet
mhw11 11 hours ago
What capability makes this possible? It’s incredible! Does it have an image-generation model?
voxic11 7 hours ago
No this is just it moving the mouse and viewing the results with it's vision capabilities.
qlte 3 hours ago
threethirtytwo 4 hours ago
I find it unlikely that there's any direct training data for this. I'm talking about direct training data for using brush strokes like humans to construct an image.
Correct me if I'm wrong but this shows that painting capability is emergent, arising from unrelated training data thus very very compelling evidence that LLMs are actually intelligent.
hdjrudni an hour ago
It could have been fed videos of people painting.
Connecting it's capabilities to simulating the recreation of such paintings... that's I suppose interesting. But adding brush strokes to produce an image is probably well covered by video footage in its training data. e.g. all the Bob Ross videos in existence.
threethirtytwo an hour ago
The transformation of pixel video data to actual brush strokes is not trivial.
For example I can watch all bob Ross videos 300 times and never paint a thing and this action would not teach me how to paint.
specked-citrus 3 hours ago
I'm also quite surprised to see that LLMs can do this. I guess it is possible that "make an image in MS paint" is a type of RL environment used for image understanding. This is one of those areas where people inside the labs have a very different view into how much models are generalizing.
XenophileJKO 3 hours ago
What we are seeing is Artificial "General" Intelligence. The model can apply intelligence to a problem it hasn't encountered before.
They have been generalizing for a long time, but spacial 2d and 3d art through tool use is a very engaging way to show it. It is harder for people to deny generalization.
Now Opus 5.5 clearly has had some sort of visual arts training, the step function change in ability implies that to me, but it can apply its spacial artistic reasoning to pretty arbitrary tools.
demibabs 3 hours ago
One other commenter said that this capability was initially revealed by Anthropic employees, raising the possibility that they were RLd on it.
londons_explore 15 hours ago
Now combine this with a robot able to move a paint brush and mix on a palette, and you could make some pretty awesome real-paint artwork
chasd00 4 hours ago
Would be a good project for an afterschool club. Even using a multi color pen plotter + webcam would be pretty neat.
albert_e 12 hours ago
Bob ROS
Koala_ice 5 hours ago
This is an S-tier comment. Thank you for your service.
mannanj 5 hours ago
I don’t like how it’s a human aiming to portray the ai generation with their taste and judgements about it but then becoming lazy and quitting on that and leaving in the ai generated portrayal about its own judgment in the site.
Just share your opinions, dude, we get its ai generated but share your own taste. We want to know what YOU think. Stop being shy and lazy.
no_multitudes 4 hours ago
What a misanthropic thing to make. I'm embarrassed for you.