Adriano D'Alessandro
@adrian-dalessandro.bsky.social
| Computer vision researcher
| Computer science PhD candidate @ SFU
| More: https://dalessandro.dev/
I like to count things and periodically I work on applications in plant agriculture + ecology.
Follow for stale political hot takes.
Free Palestine 🇵🇸
| Computer science PhD candidate @ SFU
| More: https://dalessandro.dev/
I like to count things and periodically I work on applications in plant agriculture + ecology.
Follow for stale political hot takes.
Free Palestine 🇵🇸
This is an interesting failure mode. It's obvious with context that the segmented object is a pistachio. Yet it doesn't use that surrounding context and just generates a featureless bean.
November 24, 2025 at 5:22 PM
This is an interesting failure mode. It's obvious with context that the segmented object is a pistachio. Yet it doesn't use that surrounding context and just generates a featureless bean.
This one is fun, though! The desired outcome is unclear because the billboard itself is flat, but people are not. It generates a (very oddly shaped) 3D person.
November 24, 2025 at 5:17 PM
This one is fun, though! The desired outcome is unclear because the billboard itself is flat, but people are not. It generates a (very oddly shaped) 3D person.
I was shocked it was able to generate something reasonable, here. The source image is an aerial view of pelicans. It generated something reasonably pelican-esque!
November 24, 2025 at 5:16 PM
I was shocked it was able to generate something reasonable, here. The source image is an aerial view of pelicans. It generated something reasonably pelican-esque!
That's interesting! I sometimes use this puzzle as a litmus test for mental rotation in multimodal language models. I didn't expect that they could learn this skill spontaneously but it's interesting to see the information is there in some models!
September 22, 2025 at 2:53 PM
That's interesting! I sometimes use this puzzle as a litmus test for mental rotation in multimodal language models. I didn't expect that they could learn this skill spontaneously but it's interesting to see the information is there in some models!
GPT5 still can't solve this visual puzzle, which is interesting!
August 20, 2025 at 6:32 AM
GPT5 still can't solve this visual puzzle, which is interesting!
I guess we know what's left for GPT6
August 8, 2025 at 1:13 AM
I guess we know what's left for GPT6
Paper idea: Adversarial images but for people. When shown this image of a real sand sculpture, organic intelligence tends to identify it as fake.
August 2, 2025 at 6:27 AM
Paper idea: Adversarial images but for people. When shown this image of a real sand sculpture, organic intelligence tends to identify it as fake.
o4-mini-high doesn't know about "Loss"
June 3, 2025 at 9:09 PM
o4-mini-high doesn't know about "Loss"
This is an interesting visual puzzle, which o4-mini-high fails to solve. It never considers the orientation of the letters with respect to each other and thus that there must be one letter that exists twice on the cube.
May 29, 2025 at 9:36 PM
This is an interesting visual puzzle, which o4-mini-high fails to solve. It never considers the orientation of the letters with respect to each other and thus that there must be one letter that exists twice on the cube.
Did gpt-4o get a bunk update recently? I gave it this riddle and it fell through the floor.
May 25, 2025 at 6:25 AM
Did gpt-4o get a bunk update recently? I gave it this riddle and it fell through the floor.
Asking ChatGPT to convert the children's drawing (left) to an anime character (middle) vs. a human attempt (right).
May 6, 2025 at 3:04 AM
Asking ChatGPT to convert the children's drawing (left) to an anime character (middle) vs. a human attempt (right).
I gave it this synthetic image (capital of Regina with mountains), and the models been cycling through a loop of web searchs for like 10ish minutes. It's stuck zooming in on key details, making some proposals, searching the web for matches, not finding anything, and then starting over again.
May 1, 2025 at 7:55 PM
I gave it this synthetic image (capital of Regina with mountains), and the models been cycling through a loop of web searchs for like 10ish minutes. It's stuck zooming in on key details, making some proposals, searching the web for matches, not finding anything, and then starting over again.
I waited 2 hours to vote in the Vancouver byelection yesterday. I'm absolutely ecstatic to see a landslide for socialist and progressive candidates Sean Orr and Lucy Maloney.
April 6, 2025 at 2:54 PM
I waited 2 hours to vote in the Vancouver byelection yesterday. I'm absolutely ecstatic to see a landslide for socialist and progressive candidates Sean Orr and Lucy Maloney.
I tried to get it to produce "an instance segmentation map for this image that only segments the individual figs". These aren't figs, they are Cashew fruit. Interestingly, it neither correctly identifies that there are no fig fruit nor produces an instance segmentation map.
March 28, 2025 at 9:26 PM
I tried to get it to produce "an instance segmentation map for this image that only segments the individual figs". These aren't figs, they are Cashew fruit. Interestingly, it neither correctly identifies that there are no fig fruit nor produces an instance segmentation map.
Hmph 😏 looks like my paper on semantic segmentation of occluded road signs and posts is still safe
March 28, 2025 at 3:47 PM
Hmph 😏 looks like my paper on semantic segmentation of occluded road signs and posts is still safe
Me, when I'm reviewing and actually everyone else is wrong.
March 25, 2025 at 9:07 PM
Me, when I'm reviewing and actually everyone else is wrong.
Idea: use SDXL to generate a bunch of images of reflective surfaces and extract the attention maps as pseudo-labels to train a reflective surface segmentation model and use that to constrain the depth estimation to treat this as a flat surface.
(also, I'm now realizing SDXL struggles with mirrors!)
(also, I'm now realizing SDXL struggles with mirrors!)
March 19, 2025 at 9:10 PM
Idea: use SDXL to generate a bunch of images of reflective surfaces and extract the attention maps as pseudo-labels to train a reflective surface segmentation model and use that to constrain the depth estimation to treat this as a flat surface.
(also, I'm now realizing SDXL struggles with mirrors!)
(also, I'm now realizing SDXL struggles with mirrors!)
I had this one kicking around from March 2016! Before a global pandemic, two Trump presidencies, and the "Attention is All You Need" paper.
March 17, 2025 at 7:37 PM
I had this one kicking around from March 2016! Before a global pandemic, two Trump presidencies, and the "Attention is All You Need" paper.
but why do these look like those scratched out Naruto headbands?? They look like ninjas who just abandoned the hidden village of pedophilia
March 2, 2025 at 7:34 PM
but why do these look like those scratched out Naruto headbands?? They look like ninjas who just abandoned the hidden village of pedophilia
I would like to remind the AI jury that I am in fact wearing a suit, and thus could not have committed tax fraud.
February 21, 2025 at 10:25 PM
I would like to remind the AI jury that I am in fact wearing a suit, and thus could not have committed tax fraud.
I had received an email inviting me to be a reviewer and it appears as though there was an accept option and a decline option. If one person on a paper declines the reviewer invitation is the whole paper at risk of being rejected (even if the other authors produced more reviews than they received?)
February 1, 2025 at 5:37 AM
I had received an email inviting me to be a reviewer and it appears as though there was an accept option and a decline option. If one person on a paper declines the reviewer invitation is the whole paper at risk of being rejected (even if the other authors produced more reviews than they received?)
This is the closest I got, as it has juice outside of the cup and you can see a sudden burst leave the bottom of the cup. But honestly, I'm curious how Sora is going to improve significantly when simple text-based reasoning alone requires substantial resources.
January 10, 2025 at 11:24 PM
This is the closest I got, as it has juice outside of the cup and you can see a sudden burst leave the bottom of the cup. But honestly, I'm curious how Sora is going to improve significantly when simple text-based reasoning alone requires substantial resources.