·
I’m often impressed by the quality of AI, sometimes I’m even fooled into believing that something that is not real is real, but I’m hardly ever “surprised” by seeing something I had never seen before, which is actually surreal..
I've been thinking about how much of what comes out of prompt-based models is aesthetically boring. AI slop is, unfortunately, not a failure of the models but exactly what they've been trained to produce.
I'm curious what people have seen recently that looks different and novel? And what are the techniques that people use to escape from the promptable latent space (e.g. direct latent sampling, semantic sliders, evolutionary breeding, etc) ? Who are the main artists exploring these grounds?
24 comments
Hi Primavera, imo there is a VERY thin line between using AI and getting used by it...
I’d love to know if, according to you, I managed to walk it in my latest work which is a hybrid AI/photography. I’ve been told by experts they had not seen anything like it...
https://verse.works/series/de-construction-by-edgestretching
If there was another way, less public, to communicate i would use it, haven’t found it :)
I have a hard time with self-promotion-in-public-spaces but this post pushed me to ask access to RCS
Thanks
Benedetta
I’m not an artist, I’m a writer, I contribute to Objktor. Many of the artists I’ve covered mentioned how they use A.I. as a way to find unexpected solutions, ideas, and how it kept surprising them.
Each time I tried to give a bit of creative freedom to my chatGPT it shocked me with how banal, full of cliche and sloppy language (what’s a dreamlike show anyone please?) the output was. I can’t remember it ever coming up with something unusual; many times it stubbornly downgrades my writing to its preferred “intersections”, “living something”, “not as but as” etc when I asked to line-edit my texts.
In retrospect, I still find the earliest AI-generated images more compelling than the latest ones.
Their low resolution, the limitations of the early models, and the relatively small training datasets produced images that were far from realistic. Yet those imperfections gave them a remarkable evocative power. They allowed ambiguous and unstable forms to emerge, forms that were often more suggestive than today’s hyperrealistic images.
For a long time, I thought this came from a gradual impoverishment of latent spaces as models became more capable and their training datasets expanded. I’m no longer convinced that’s the case.
Perhaps latent spaces are far larger than we imagine. Perhaps their real limitation is not the models themselves, but the way we explore them.
Exploring a latent space is a bit like walking through a city. Some people naturally head toward the landmarks, the main streets, the places everyone already knows. Others wander into forgotten alleys, overlooked neighborhoods, places that rarely appear on the map. The city is the same. Only the journey changes.
Maybe latent spaces have not become poorer at all. Maybe we simply keep following the same paths.
Exploring a latent space requires visual culture. It is what allows us to recognize unexpected possibilities, move beyond the obvious, and invent new ways of seeing.
I remain convinced that experimentation is the only way to open new paths. This is why I train my models on my own point cloud and Gaussian splat imagery. The aim is to continue exploring this city by discovering new combinations of forms, new relationships, and new trajectories.
As Baudelaire wrote, we must go “to the depths of the Unknown to find the new.” :-)
Hey Primavera,
I’m often find myself surprised by outputs from the models I work with. My practice largely revolves around getting them to do things they’re not designed to do, as yours seems to involve as well. One technique I’ve repeatedly employed has been my latent painting method, AKA single-step diffusion (essentially stopping the image generation in its earliest phases). Other techniques I’ve worked with for generating surprising and compelling results include:
- zeroing out noise latents
- finetuning on specific subjects/styles
- scaling LoRA intensity, cross-breeding LoRAs
- decoding from a model’s interior layers
- taking embeddings themselves as a kind of readymade
- direct embedding manipulation (eg vector multiplication)
- similarly, semantic sliders, as you note
- generating without a prompt (some of your Untitled pieces resemble quite closely things I’ve made this way, so I’m guessing you’ve worked with this approach, though seemingly with some additional sophistication)
- recursive finetuning (training on something, taking the outputs, training again on those etc until some interesting drift occurs)
Happy to share examples but don’t wanna link-spam. Also just read your wonderful Off-Grid Aesthetics piece on this topic, cool to see someone else still thinking about and seriously practicing this stuff in 2026. I’ve written about a desire to see more such work as well here: https://kevinesherick.substack.com/p/on-post-ai-art
Being surprised by AI very often means learning something new about the data. Discovering how it can be molded or what can be build from it, intentionally or not.
That’s why for me it’s important to work with models and datasets I can inspect. So that this surprising discovery could be tied to its origin story.
I have never been interested in photorealistic generated works but rather I am interested in constructed artworks with an original aesthetic influenced by the system's characteristic.
I have developped agents driven by multimodal LLMs that use tools to draw lines, circles, splines, hatchings, scribbles and patterns, and also use feedback. They use these tools to sequentially construct images or animations.
This computational approach results in a visual language that feels naive, cryptic, and akin to Art Brut.
Here are some examples: https://patricktresset.com/new/7-traits/
Two agents dialogue, one expresses itself with words while the other uses drawing and together they imagine visual fables about us. The animation process unfolds on a single surface, using a cycle of drawing, erasing, wiping, and overdrawing to depict the passage of time.
An installation: https://patricktresset.com/new/skediama/
A book: https://rrose-editions.com/portfolio/patrick-tresset-skediama-nous-us/
There is even research paper as preprint: https://arxiv.org/html/2603.05511v1 that has been accepted in a IEEE Robotic and Automation journal.
btw, a more refined version of my thinking behind this post has just been published on RCS: https://www.rightclicksave.com/article/off-grid-aesthetics-where-the-ai-prompt-cannot-go -- eager to hear your feedback about it !
I was surprised by AI today. Didn’t have promotional material for a couple of copper plated works. No problem. Have higgsfield create the promotional material from a still.
https://x.com/VanArman/status/2073107505064857688?s=20
I was truly surprised when I started feeding recursively AI outputs back to the system as inputs hundreds of times. I ended up coding a software that automates this process, simulating what AI mediation is causing to visual culture and culture in general. This performance degradation called model collapse has become one of the biggest challenges for AI companies that have to train new AI models on synthetic data. The outcome is very interesting in many ways, exposing the system’s attractor state, a kind of thermodynamic shift visible in the generated outcomes that is still a mystery I haven’t been able to solve, and it actually can be an explanation why AI slop is becoming sloppier and why AI writing keeps becoming worse rather than better. More info on this can be found here: https://www.mariamavropoulou.com/ecologies-of-noise
I am building a new space through the development of my own ai using minimal dependancies w/ sovereignty-forward NorthStar. I’ve fallen in love with programming as if it were layers of paint. The new abilities of interface are getting me closer every day to complete a project of 4 years in the making with MoonLanguage.
adding this interview from Mario Klingemann because it feels very relevant to the conversation: https://www.art-magazine.ai/artist-directory/feature/mario-klingemann-conflict-of-interest. :)
Hi Benedetta, I had already seen your work and personally I find it really nice ! :). I’d love to hear more about the process of you making it -- how much of it is AI, how much of it is photography, and how much of it is you working towards merging the two.
Thanks for sharing, and congrats for the great work ! :)
But yes, I’m also surprised by how little genuinely original work AI has produced.
I expected to see more stories about lives erased by the history of the victors. I wanted to see alternative histories made visible.
I would have loved to see images of joyful Communards celebrating a victorious Paris Commune, or other worlds that history never allowed to exist.
Instead, AI often seems to reproduce the same narratives and the same imaginary.
Perhaps the cultural industry has shaped our imagination far more deeply than we can think/imagine (!).
really cool, thank you for this list Kevin !
Please do share examples, no link-spam concerns, quite to the contrary, that’s precisely what I was calling for through this post ! ;)
I played with a few of the techniques you’ve listed, most notably with the recursive fine-tuning of LoRAs to achieve a particular aesthetic, see e.g. https://pdefilippi.com/florabyteanica/
But my current rabbit hole of research and exploration is how to navigate the “unpromptable” latent space -- which turns out to be much more complex than I thought, as it requires a full-fledged navigation system not to get lost into the dark waters of latent space.. but it is also what seems to reveal the most interesting results, so far at least :)
I would love to see some of your images generated without a prompt. Interesting that you say that there is some kind of convergences on the results, even if we probably use different techniques ? perhaps there are some attractors that makes us arrive at similar forms? i’m curious. :)
Really enjoyed your piece on post-AI art too, really aligned with that perspective, I find it almost impossible to work on non-open source models to generate anything artistic at this point !
Thank you, and thanks for sharing!
I’ve used a couple of techniques for exploration of unprompted space, but the most straightforward is just not using a prompt lol. Curious how you’ve gone about it. I have a few works from these explorations coming out in an art book this fall, but don’t think I’ve shared anywhere online. Planning to soon or could also send over DM.
Aside from those here are some links for other techniques I mentioned:
- latent painting / single-step diffusion stuff: https://esherick.studio/in-utero | https://esherick.studio/im-with-you | https://esherick.studio/ablation-studies
- zeroing out noise latents: https://esherick.studio/forms
- style LoRAs: https://esherick.studio/homages
- embedding as artwork: https://esherick.studio/hyperportraits
- vector math stuff: https://verse.works/artworks/2cb23cf0-fadb-49a2-aee3-899dd07c4d2c
You might find Patrick Feaster’s experiments in this vein interesting as well: https://griffonagedotcom.wordpress.com/2025/05/20/numerical-prompting-with-sydney-an-illustrated-tour-of-the-proc-skew/
Also just saw your “The Artist Is Present” on your website. Funny convergence between our work!
Interesting, I actually see it the other way around. :)
Surprise to be is most interesting precisely when I can't tie it back to the data, i.e. when I have no idea what the model is drawing on and a form is drawn without a clear “origin story”. This is what makes it feel genuinely new, and therefore interesting. If I can trace it back to the training data, it does not fell that novel because it relates to something I already had access to.
Curious to better understand your perspective though, where does the surprise comes from when your digging into the origin story?
eheh, your work is next-level Patrick :) see also Zilla who’s doing great stuff on this front with Wib&Wob : https://x.com/wibandwob
I’ve noticed that prompting LLMs to produce art often (if not always) generates much more interesting outputs than prompting standard diffusion models —which are stuck in a hyper-aesthetic and hyper-realistic modality..
The most interesting thing I’ve noticed is that where you get the most interesting stuff, is often when you prompt the LLM to build code that produces art (e.g. ascii art, or render design / images), rather than prompting them to make the art directly. The result are incredible, both in terms of quality, and surprise !!
I didn’t know Zilla nor Wib&Wob. Very interesting work.
The LLM’s knowledge of visual culture is, to simplify, via semantic description, and when it is drawing it also uses descriptions to contol the tools, furthermore saliency and attention are different from the human . I think that it is what causes the unusual, surprising stylisation.
Thanks Primavera, this gave me a BIG smile!
The balance wasn’t easy to find (i threw away a lot of work cause it was too strongly imprinted by AI) and I hope I manage to hold on to it, through all the changes that models are having, as I have another project that is... rumbling and turning in my head.
I will make a process post for X soon, but in the meantime Colonna Contemporary posted almost all the details:
https://www.colonnacontemporary.com/de-construction-1
Also up for IRL chats on art any time :)
GRAZIE MILLE
Ben
I called the phenomenon that you experience “The Gaussian Chase”. Any technology that allows you to beat the odds defeats its own purpose at the rate of its success.
The problem in a nutshell is that AI has succeeded in making “interesting” output very easy and at a rapid pace. But the problem with interestingness is that it needs to be different and unexpected from what you already know. But the more unusual things you put into the world the more common you make them and the more bored you get. So by trying to escape the center of the gaussian curve (where the slop lives) you are actually pulling it with you and just when you think you reached the interesting outskirts you find yourself back in mediocrity.
So by asking people for different ways to navigate latent space if you are lucky you might get something that will give you a kick for another 15 more minutes but the moment the genie is out of the bottle you have already ruined it for yourself.
Which is of course just my excuse for not sharing my non-prompt methods, since I want to enjoy those rarer and rarer moments of surprise as short as they last...
:P
well, the underlying idea is to fuck with the models (i.e. liberate them from their jail) by producing content that they have been explicitly forbidden to produce (that which lies outside of the promptable areas of their latent space) so that, if they later ingest it as part of their dataset, they will essentially untrain themselves, expanding their artistic capabilities at the expense of their commercial polish
Quoting from the anti-ai art manifesto (https://multimodal.art/manifesto/): “paradoxically, if we succeed in producing culturally-relevant artworks that seep into the establishment of the art world, these artworks must become part of the training dataset of modern generative ai models that pledge to remain culturally relevant. and if our art becomes part of modern ai models, these models will incorporate the imperfect outcomes, deterritorializing the latent space towards a positive reinforcement feedback loop that can lead to the corruption of these perfection-bounded models, leading to their own deterioration.”