Elorian Glimpse: Visual Thinking That Stays Grounded

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Elorian Glimpse: Visual Thinking That Stays Grounded · Channel: Fahd Mirza · 09:57 · Watch on YouTube

Topics: AI evaluation and benchmarks, Multimodal models

Summary

The video tests Glimpse, a visual-thinking model from Illorian, on an AI-generated video, real footage, and five AI-generated images. The presenter explains that the model examines regions and moments over time, then assesses its performance alongside the company’s claims about benchmarks and lower pricing. Glimpse identifies several visual inconsistencies and gives detailed observations, while also showing some errors and uncertainty. The presenter concludes that it is strong at visual reasoning, though its answers about legal details are less reliable.

Key points

Glimpse is presented as a visual-thinking model that revisits regions and short moments, rather than answering from a single look.
The presenter notes that the benchmark results are mixed: Glimpse wins or ties on some tests but does not lead across the board.
The company’s reported pricing is a central selling point, with Glimpse described as much cheaper than the larger models compared.
On the train video, Glimpse flags that the trains do not move relative to the man or track and identifies the clip as likely AI-generated.
The presenter also notes flaws in that response, including a text glitch, excessive length, and claims that seemed uncertain.
On real footage, Glimpse tracks the presenter’s right hand, counts two screens and a microphone, and acknowledges uncertainty about a desk object.
The model says it cannot confirm what the presenter is saying because the audio quality is insufficient.
Across five country-specific images, Glimpse matches the images and points out visual details, but contradicts itself on some seatbelt observations.
The presenter considers Glimpse’s visual reasoning stronger than its handling of legal details, which it often marks as uncertain.

Chapters

Start	Chapter	Summary
00:01	Introducing Glimpse and visual thinking	The presenter introduces the model and explains its approach of revisiting video regions and tracking details over time.
03:09	Testing an AI-generated video	Glimpse identifies signs that the train clip is AI-generated, though the presenter notes weaknesses in its response.
04:45	Testing real footage	The model gives detailed observations about a real video, tracks hand movement, and states where it is uncertain.
06:50	Comparing five images and final assessment	The presenter tests cross-image reasoning and seatbelt observations, finding useful visual detail alongside contradictions and uncertainty about legal information.

Quotes

It zooms into regions, rewatches short moments at a higher frame rate, and tracks one feature through time before it responds. — 01:03

Its conclusion was that the trains never move relative to the man or the track, which suggests the clip is AI-generated. — 03:09

So, very honest approach, no hallucination, flying um, nothing there. — 06:20

A pretty good release, the first release with limited resources. — 09:57

Concepts and entities

Glimpse, Illorian, visual thinking, GPT-6 Astra, Claude Opus 5.5, Happy Horse, Hermes Agent

Full transcript (original language)

00:01 Watch this video first. This is a man doing the epic split across two moving trains. An AI-generated video looks cool, but is it physically real? Let’s find out with this new model, which is Glimpse. This is the playground for it, a new model from a company called as Illorian. I’m pasting a hard prompt here, as you

00:31 can see. It asks the model to track the man and both trains, compare their speed, check the physics, predict what happens if one train breaks, and flag anything that looks AI-generated with timestamps. Let me run this. It is working through it, and I will play it again for you. So, I’ll just show you the reasoning because they have said it is bit different. How is it different? Let’s talk bit more around it.

01:03 So, look. We are now testing that whether it describes the video or not, because what they say is that they have come up with a new framework called as visual thinking, and this new Glimpse model is a visual thinking model. The idea is that it does not look once and answers. It zooms into regions, rewatches short moments at a higher frame rate, and tracks one feature through time before it responds.

01:34 If you go through this report which they have shared, and there are a lot of examples, benchmarks, and stuff. For example, look at this. What they are saying here is that it was built by a small team on a few hundred GPUs, post-trained on an open weights base model, and priced much lower than the bigger frontier models. But, these are their claims, not my test, and that is what exactly I am doing here. It is still working through it, watching the video. So, I will wait for it to finish. Meanwhile, let me

02:06 quickly talk about this benchmark table. So, what this statistics show me is that it is not winning everywhere. And that is exactly what makes it interesting. On the big expensive model, GPT-6 Astra still takes the crown on several tests like fine-grained visual puzzles and dense video captioning. Claude Opus 5.5 edges ahead on reading scientific charts, but Glimpse wins or ties on the robotics test, short-form video

02:38 captioning and counting objects, and it stays close behind on the rest. I think the main punchline which they are trying to convey is the price. Glimpse costs about 94% less than GPT-6 Astra and 85% less than Opus 5.5. So, you’re getting near frontier vision at a fraction of the cost, and that is the whole premise here, which is quite attractive. And that is what we are exactly testing right now on their playground, and I will drop the link in video’s

03:09 description, too. And this is the response of the model, as you can see. I will quickly take you to the end of the response first. Look at this. Instead of just describing the video, it pulled out the frames and cited at timestamps. Now, check this response out from the start. Look at this. Its conclusion was that the trains never move relative to the man or the track, which suggests the clip is AI-generated. Totally spot-on. It also pointed to the

03:42 happy horse watermark, which is of course from Alibaba, which is quite correct. There were weak spots, too, a lot of them. A small glitch leaked into its text. The answer was long. Um and I think some of them are correct. Some of them it is just I think slipping away and fluttering. So, I think so 50/50 I would say it is quite good that in terms of some of the um you know, limitations it has already admitted like

04:13 this, which is quite good. And then it says that it is optically inconsistent. So, that is the main thing. That it is identifying all the limitations and flaws in Happy Horses video, which is AI generated. So, which is quite good. Then it talks about that dust plume at the left horizon, which is quite good. And then some of the haze patches it has identified. So, all in all I think it is quite a decent balance between performance and price. And now let’s try out a real footage. As

04:45 you can see here, it’s not an AI generated, no editing whatsoever. So, let’s see if it is able to do um whatever claim it has made because I’m asking it lot of questions around how many monitors, track the person, and describe the lighting and the window blinds. I will run this and let’s see what it comes up with. Let me also play it a bit for you. >> Hello guys. I’m just creating this sample short video for this deep live cam demo. And again, this is just for educational

05:17 purposes. And don’t judge my >> So, you see there are some hand gestures. The quality is not good enough around the lighting. Plus um it’s a bit of an older video where some light is coming through the blind. So, let’s see what the model comes back with. And meanwhile, please subscribe to this free weekly AI newsletter at fawadmirza.substack.com. I promise I’m not going to spam you with it. And if you want to help out the channel, please become a member by clicking on

05:49 this link or in the first pinned comment. And here is a fair deal. On the real footage and clip clips was very detailed as you can see from this response. It broke my 29 seconds into time detections, tracked only my right hand entering and leaving the frame, which is correct. Counted two screens and one microphone and measured the blinds staying constant by sampling the brightness across the clip. It also flagged its own uncertainty on a few items like

06:20 a small device on the desk it could not identify. And it pushed back on a claim that my gaze lifts to the camera at the end. Um, there were few weak spots too, but not many, seriously. Because it only worked from the frame to frame. And but it said that it could not confirm what I was saying because the audio was not good enough. So, very honest approach, no hallucination, flying um, nothing there. So, I think it has done fairly well and you can see these are all the frames it

06:50 has gone with. Really good stuff, you see. This is a hand movement it was talking about frame by frame. For the next test I’m going to use this agentic harness with these five images, each from a different country. So, this is Australia, all AI generated by the way. This is Canada. And this is Germany. This is India. And this is Nigeria. So, I have I’m trying to be fair to each

07:21 continent. Now, what we are trying to test here is that three things. Whether a model can see small details, like exactly how each seatbelt is routed. Can it reason across several images at once? And does it admit when it does not show a country’s law? These images AI generated and they’re photographed from inside the car. So, a good model should also point out that a real enforcement camera would see something very different. So, what I

07:53 have done here, I have already attached the images in Hermès agent and I’m also using it through their API at um Illodion Glimpse, as you can see. Let me paste that prompt. So, this is a prompt where I’m testing what I what I just told you. So, let’s run this. It is thinking pretty nicely again, frame by frame, and then it’s reasoning across it. It is also trying to map all of these images as how it should go and how they

08:26 relate to each other. And this prompt is really stretching the model. It is getting confused at some points, but if you look at the reasoning, it is trying its best across the countries. Also checking its own data to see if it is able to find anything, but uh having a hard time at the moment. Let’s see what it produces. And here you go. So, the response looks pretty good. It has matched all five images from what I can see. The missing

08:56 sash on the cream sweater, the underarm belt on the uh blouse, and all that stuff. So, it stayed honest on the camera question, too, by the way. That these are interior shots and that contrast, like hair and glare, would decide detection. And it also marked many legal details as uncertain. So, the training data set doesn’t have them, it seems, which is fine. Um but there are few things where it slipped. It contra- dicts at several places itself. For

09:26 example, on India, the strap is visible low across the abdomen, yet it says the lap portion is absent. The fines and camera rollouts are um doesn’t seem that correct to me, but overall I think what I would say in general that it is quite strong on the visual side, which is the model’s real selling point anyway in all fairness. It reasoned carefully across five images in one pass. So it treated the legal half as a starting point to check.

09:57 So but other than that it has got the fine visual details accurately, which is exactly what we wanted to test. A pretty good release, the first release with limited resources. Let me know your thoughts. Thank you for all the support.