Anyone who has tried to put a price on a digital ape knows the old question: does the picture add anything the list of traits does not already say? In a study published this October in Nature's Scientific Reports family, researchers tested that question against real sales of Bored Ape and Mutant Ape tokens, and found that, for the most part, the traits already tell the story. The images help only in fleeting, unreliable ways, a reminder that a valuation story can look convincing until it is asked to hold still under scrutiny.
Image Features Add Little to NFT Price Prediction Beyond Metadata, Temporal Tests Show
The traits already carry nearly everything the pictures do.
Mark
Why does a study of ape NFT prices matter beyond the NFT market?
Mimi
It's a clean test of a common assumption: that adding image data to a model must help. Here the traits and the picture describe the same thing, so the image mostly repeats what's already known.
Mark
How big were the gains?
Mimi
Tiny. Early fusion cut error by 0.09% in BAYC and 0.95% in MAYC compared with metadata alone.
Luke
And late fusion was inconsistent, worse by 0.07% in BAYC, better by 0.70% in MAYC. Differences that small are easy to lose.
Mark
Did the models predict anything at all?
Mimi
Yes. All four groups beat a zero log-relative-price benchmark, so the prediction is real. It just doesn't come from the images.
Luke
Careful, though. Beating that benchmark doesn't tell us how good the models are in absolute terms, and the source gives no such figure.
Mark
What about new tokens?
Mimi
For unseen tokens, the 95% intervals comparing fusion with metadata include zero in both collections.
Luke
Those intervals are labeled exploratory, so they're suggestive rather than definitive.
Mark
Then the cleaning tests?
Mimi
Under two stricter variants, 13 of 16 representative configurations changed, and fusion gains weren't consistently supported after reselection.
Luke
Which also means the findings are fragile in both directions. The paper doesn't say images are useless, only that it can't establish consistent gains or aesthetic pricing power.
Der Puls
- Models trained on years of earlier sales were tested against a later stretch of real transactions, leaving little room for hindsight to flatter the results.
- Adding image data to trait-based metadata trimmed prediction error by fractions of a percent, a margin small enough to vanish under noise.
- When the researchers tightened the data cleaning and reran model selection, 13 of 16 representative setups shifted, unsettling the apparent gains.
- The verdict is landing firmly on skepticism: image-based NFT valuation claims need time-split testing and robustness checks before anyone should trust them.
Anyone who has tried to put a price on a digital ape knows the old question: does the picture add anything the list of traits does not already say? In a study published this October in Nature's Scientific Reports family, researchers tested that question against real sales of Bored Ape and Mutant Ape tokens, and found that, for the most part, the traits already tell the story. The images help only in fleeting, unreliable ways, a reminder that a valuation story can look convincing until it is asked to hold still under scrutiny.
The question at the heart of this study is an old one in a new market. If you want to predict what a Bored Ape or a Mutant Ape will sell for, does a computer looking at the picture help, or is everything the picture shows already written into the token's traits? The researchers, publishing in Nature's Scientific Reports family, put the question to real sales data and found the answer was mostly no.
They examined two collections, Bored Ape Yacht Club and Mutant Ape Yacht Club. Here the image and its attributes describe nearly the same object, which made the value of the image uncertain from the start. They built models using metadata alone, images alone, and two ways of combining them, testing eight regression methods and seven visual encoders. Crucially, they chose their models using only pre-2025 data and then judged them on sales from January 2025 through April 2026, so the future could not leak into the test.
The models did learn something real: every group beat a naive baseline. But the benefit of adding images was tiny. Early fusion reduced prediction error by under one percent in each collection, and late fusion was inconsistent, worsening one collection while improving the other slightly.
The robustness checks left the case weaker still. For unseen tokens, the uncertainty ranges comparing fusion with metadata included zero in both collections, meaning no real difference could be confirmed. When the authors tightened the data cleaning and reran model selection, most configurations changed, and the modest fusion gains stopped holding consistently.
The authors stop short of declaring images worthless. They claim only that this evidence does not establish consistent gains from images, independent aesthetic pricing power, or a clear winner between fusion strategies. The practical lesson is cautious: in these collections, the traits already carry nearly everything the pictures do, and any claim that images unlock value should prove itself on time-separated data and survive changes in method.
The question behind this study is simple to state. If you want to predict what a Bored Ape or a Mutant Ape will sell for, does it help to have a computer look at the picture, or is everything the picture shows already captured in the list of traits attached to the token? A paper published in Nature's Scientific Reports family, dated October 2026, tests that question against real sales and finds that the answer is mostly no.
The researchers worked with two collections, Bored Ape Yacht Club (BAYC) and Mutant Ape Yacht Club (MAYC). These are profile-picture NFTs, where the image and the categorical traits (the attributes that describe it) are closely related descriptions of the same object. That overlap is why the value of image data was uncertain from the start.
They built four kinds of models: metadata-only, image-only, early-fusion (image and metadata combined before prediction), and weighted late-fusion (separate predictions blended afterward). The comparison covered eight regression families and seven visual encoders. Model selection used only data from before 2025. The models were then evaluated retrospectively on sales from January 2025 through April 2026, which keeps the test from peeking at the future.
There is one clear positive result. The selected models in all four groups beat a benchmark that always predicts a log-relative price of zero. So the models learn something real about price.
The gains from adding images, though, are small. Compared with metadata alone, early fusion cut root mean squared error (RMSE) by 0.09% in BAYC and 0.95% in MAYC. Late fusion was mixed: it raised BAYC error by 0.07% and lowered MAYC error by 0.70%.
The sensitivity checks weaken the case further. For tokens the models had not seen before, the exploratory 95% intervals comparing fusion with metadata include zero in both collections, so a real difference cannot be distinguished from none. In MAYC, a metadata-only partial least squares (PLS) model outperformed the CLIP–PLS early-fusion model. The authors also varied the data cleaning, using two stricter variants and rerunning the selection. Against a common base-clean control, 13 of 16 representative configurations changed. After reselection, fusion improvements were not consistently supported. Some of the tree-based comparisons were also sensitive to the implementation used.
The authors' conclusion is restrained. The results do not establish consistent gains from images, any independent pricing power for aesthetics, or the superiority of either fusion strategy. That is a statement about what this evidence can support, not proof that images never matter.
The practical reading is that claims about image-based NFT valuation deserve skepticism unless they are tested on time-split data and survive changes in cleaning and model selection. In these two collections, the traits already carry nearly everything the pictures do.
Bemerkenswerte Zitate
The results do not establish consistent image gains, independent aesthetic pricing power, or superiority of either fusion strategy.— Study authors (paraphrased from abstract)
Möchten Sie die ganze Geschichte?
Das Original lesen bei Nature ↗
Science & Space