DeepSeek Vision Is Actually Impressive.... And Incredibly Cheap

English
BBetter Stack
Computing/SoftwareSmall Business/Startups

Transcript

00:00:00DeepSeq just released a dedicated vision model and it's incredibly cheap,
00:00:04coming in as low as 0.00008 cents per photo, meaning you can run this at scale behind your
00:00:11own apps and services. Today we'll test out how capable this really is in real world conditions,
00:00:17we'll snap foods, labels, random objects outside to see how well the model performs.
00:00:22I talk about 14 photos as part of the test and it cost me less than 1p,
00:00:26so let's see how good DeepSeq Vision really is.
00:00:34So just a quick rundown of how this works before we jump into the tests.
00:00:39Vision is available under the model name DeepSeq v4 Flash Vision EXP, the EXP being experimental.
00:00:47DeepSeq themselves say the model accepts images alongside text, so you can ask the model to
00:00:52describe pictures, read text from screenshots, analyse charts and more.
00:00:56We can send images in three ways. First, as a base 64 encoded image as part of the standard request
00:01:01body. However, the encoded data counts towards the 48 megabyte request body limit, so this will only
00:01:07work for smaller images. The second option is you can provide an external image URL, but here the image
00:01:13can be max 32 megabytes and the download must complete within 60 seconds. And the final option
00:01:18is file upload via the files API. This lets you send larger images and it's likely the option you'd want,
00:01:25and what I'll be using for my demos today. Now, it's also worth noting smaller images are scaled up to
00:01:30reach roughly 394 by 394 pixels and larger images are scaled down to match roughly 800 by 800 pixels. I say
00:01:39roughly because the aspect ratio is always preserved. This can mean that small details and texts could be
00:01:45lost before the model even had a chance to read them. So just something to keep in mind. And with that said,
00:01:50let's go take some photos. But before we do, you do me a huge favor by subscribing to the channel so we can
00:01:55keep bringing you great content every single day. Okay, so I'm inside my food cupboard now because I want
00:02:00to see how well DeepSeek can do it detecting random items like this apple here. And then I also want to be able
00:02:06to scan things like this food label. So the front should be fairly easy because the text is so big
00:02:11but I also want to see how well it does at detecting the label on the back as well because the text there
00:02:18is so small. And because DeepSeek shrinks a lot of these images down, I want to see how well it can
00:02:23actually process tiny text like the back of a label. So I'll take a bunch of these photos and then we'll
00:02:28come back and look at the results in a second. Okay, so let's look at the first image of this peanut butter here.
00:02:33You can see DeepSeek and Claude have both brought back the text as expected, although DeepSeek has
00:02:38made a few mistakes here. The per measurement is 15 grams and DeepSeek seems to think that's 30 grams.
00:02:45Claude correctly got that right at per 15 grams, but most of the other text here is actually very accurate.
00:02:50So you can see with the clear photo, DeepSeek still does pretty well. Now, if you look at the second one,
00:02:55DeepSeek again did pretty well, though it did make a few mistakes. For the main clear bold text,
00:03:01meridian, fully roasted and smooth. It got all of that correct. But Claude was also able to determine
00:03:08that this little badge here says no palm oil ever. DeepSeek was not able to analyze that at all. So you
00:03:14can see if things are slightly blurred or at the wrong angle, DeepSeek does start to struggle.
00:03:19It's a similar thing again with Yorkshire tea. You can see it gets all of the bold text. Yorkshire tea,
00:03:25decaf. Let's have a proper brew. But again, Claude was able to correctly determine that this says plant
00:03:30based composable tea bags, plant based composable tea bags. DeepSeek wasn't able to see that. I also
00:03:36wanted to see how DeepSeek could perform with different fonts and branding. So we've got a
00:03:40little pack of jammy dodgers here and DeepSeek correctly analyzed all of the text here. Interestingly,
00:03:45though, it did also hallucinate. So if we look down here, we see a trip to Florida,
00:03:49Minions and monsters. And then randomly, DeepSeek just decided to add the rise of Gru,
00:03:55which isn't mentioned anywhere on the label. Now, this is strange because Gru is obviously
00:04:00a character inside Minions. So this means DeepSeek is hallucinating based on its training data
00:04:06and just inventing things when it struggles to understand the text. Next up is just a photo of an
00:04:11apple. DeepSeek actually thought this was a yellow peach. It's just completely wrong on all counts.
00:04:17Claude detected it was an apple and also detected that it was in fact a garla apple. So it was able
00:04:23to also find the correct species of apple, not just the fruit itself. And I thought that was super
00:04:28impressive. All right, guys, we're actually going to head outside for once to touch some grass.
00:04:32I'll grab a quick coffee on the way and then we can take some more photos.
00:04:37I took a photo of this leaf here and DeepSeek correctly detected it was a leaf,
00:04:42but believes it's a leaf from an American sycamore tree, which is not correct. This was
00:04:46taken in the UK. Claude though was super accurate and he said a plane tree leaf in a UK park. There's
00:04:53almost certainly a London plane. Then we just got a photo here of a street lamp. Both DeepSeek and
00:04:58Claude were able to detect a Victorian style street lamp. Again, for this bench as well,
00:05:03this is a traditional metal slatted park bench. And finally, I just took a picture of this car,
00:05:07which is a Tesla Model 3. DeepSeek was able to detect that this is in fact a Tesla Model 3.
00:05:13So you can see the model is still slightly rough around the edges, but due to the extremely cheap
00:05:18price, I do think this is still a viable option to run inside your own services. The cost is just
00:05:24astronomically low and DeepSeek have been releasing like crazy recently. They also dropped their own
00:05:29harness which exploded on GitHub and you can find a video on that right here.

Key Takeaway

DeepSeek v4 Flash Vision EXP provides an extremely affordable visual recognition option at 0.00008 cents per photo, though accuracy suffers on small text, subtle details, and fine-grained categorizations compared to competing models.

Highlights

  • DeepSeek v4 Flash Vision EXP costs as low as 0.00008 cents per photo for API usage.

  • The model scales images down to roughly 800 by 800 pixels or up to 394 by 394 pixels while preserving aspect ratio.

  • DeepSeek misidentified a Gala apple as a yellow peach and incorrectly calculated nutritional portion measurements on product labels.

  • Hallucinations occur during text processing, such as inventing characters from training data when struggling with blurry fonts.

  • Objects with clear, bold formatting like Tesla Model 3 vehicles and Victorian street lamps are detected accurately.

Timeline

Model Specifications and Image Input Methods

  • DeepSeek v4 Flash Vision EXP accepts images via base64 encoding, external URLs, or the files API.
  • Image uploads have strict size caps, with base64 restricted by a 48 megabyte request body limit and URLs capped at 32 megabytes.
  • Automatic image scaling adjusts dimensions to roughly 394 by 394 pixels for smaller images and 800 by 800 pixels for larger ones.

The experimental vision model operates at an exceptionally low cost of 0.00008 cents per photo, making high-scale deployment financially viable. Input mechanisms cater to various data pipeline designs, though the automatic resizing process risks stripping away fine textual details before the model evaluates them.

Food Product and Label Analysis

  • Bold, prominent brand text on items like peanut butter, tea, and biscuits is accurately extracted by the model.
  • Nutritional data and smaller badge details frequently trigger errors, such as misinterpreting 15-gram measurements as 30 grams.
  • Hallucinations emerge on complex labels, causing the model to invent movie characters not present on the packaging.

Real-world testing against consumer goods reveals strengths in bold typography recognition alongside weaknesses in nuance. While primary titles are transcribed correctly, small warning labels or secondary badges are missed entirely, and unrecognized text leads the model to fabricate related terms from its training data.

Outdoor Object and Environment Recognition

  • Nature photography tests show mixed results, misidentifying a Gala apple as a yellow peach and misclassifying tree leaf species.
  • Man-made infrastructure objects such as Victorian street lamps, metal slatted park benches, and Tesla Model 3 cars are successfully detected.
  • Rough edges in performance remain, but low operational costs preserve utility for specific application environments.

Outdoor evaluations demonstrate reliable identification of common manufactured objects while highlighting classification errors in botanical subjects. Despite these accuracy limitations, the microscopic pricing structure establishes the model as a viable candidate for budget-conscious service infrastructure.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video