screen 3 \u2192 confirmation<\/li>\n<\/ul>\n\n\n\nThe idea was to have it analyze UX across transitions.<\/p>\n\n\n\n
This is where things got messy.<\/p>\n\n\n\n
It doesn\u2019t maintain a strong sense of continuity between screens unless you force it.<\/p>\n\n\n\n
So it evaluates each screen in isolation.<\/p>\n\n\n\n
Which misses the whole point of UX.<\/p>\n\n\n\n
We tried stitching context together manually:<\/p>\n\n\n\n
\u201cGiven previous screen X, analyze current screen Y.\u201d<\/p>\n\n\n\n
It helped a bit, but the model still treated each step as a fresh interpretation.<\/p>\n\n\n\n
Not a continuous experience.<\/p>\n\n\n\n
\n\n\n\nThere\u2019s also a weird issue with overconfidence.<\/p>\n\n\n\n
DeepSeek VL will confidently critique spacing, alignment, hierarchy\u2014even when it\u2019s misreading the layout.<\/p>\n\n\n\n
So you get feedback like:<\/p>\n\n\n\n
\u201cButton alignment is inconsistent\u201d<\/p>\n\n\n\n
Except the buttons were aligned\u2014it just grouped them incorrectly.<\/p>\n\n\n\n
That means you can\u2019t trust feedback at face value.<\/p>\n\n\n\n
You have to verify everything.<\/p>\n\n\n\n
Which reduces the time savings.<\/p>\n\n\n\n
\n\n\n\nWhere it does<\/em> work well is pattern detection at scale.<\/p>\n\n\n\nWe ran batches of UI screenshots through it to find recurring issues.<\/p>\n\n\n\n
Not detailed critiques. Just patterns:<\/p>\n\n\n\n
\n- repeated layout inconsistencies<\/li>\n\n\n\n
- common color misuse<\/li>\n\n\n\n
- duplicated components with slight variations<\/li>\n<\/ul>\n\n\n\n
At that level, it\u2019s useful.<\/p>\n\n\n\n
Because even if it misreads individual elements, aggregate patterns still emerge.<\/p>\n\n\n\n
It\u2019s less about precision, more about signal.<\/p>\n\n\n\n
\n\n\n\nWe also tried combining DeepSeek VL with agent workflows.<\/p>\n\n\n\n
Idea:<\/p>\n\n\n\n
\n- VL model analyzes screen<\/li>\n\n\n\n
- agent extracts structured issues<\/li>\n\n\n\n
- another agent prioritizes them<\/li>\n<\/ul>\n\n\n\n
On paper, this is exactly what you\u2019d want.<\/p>\n\n\n\n
In practice, small visual misinterpretations cascade through the system.<\/p>\n\n\n\n
If the first step mislabels something, every downstream step builds on that error.<\/p>\n\n\n\n
So instead of amplifying accuracy, the pipeline amplifies mistakes.<\/p>\n\n\n\n
We had cases where a minor UI quirk turned into a \u201chigh-priority usability issue\u201d because the chain reinforced it.<\/p>\n\n\n\n
\n\n\n\nMemory 2.0 doesn\u2019t help much here either.<\/p>\n\n\n\n
You\u2019d think it could learn design patterns over time.<\/p>\n\n\n\n
But what it actually does is store surface-level preferences.<\/p>\n\n\n\n
Like:<\/p>\n\n\n\n
\n- \u201cprefers minimal design\u201d<\/li>\n\n\n\n
- \u201cuses blue primary buttons\u201d<\/li>\n<\/ul>\n\n\n\n
Not actionable UX insight.<\/p>\n\n\n\n
And sometimes it applies those preferences where they don\u2019t belong.<\/p>\n\n\n\n
So analysis becomes biased.<\/p>\n\n\n\n
We ended up disabling memory for most UX tasks.<\/p>\n\n\n\n
\n\n\n\nAnother friction point is resolution and clarity.<\/p>\n\n\n\n
DeepSeek VL handles standard screenshots fine.<\/p>\n\n\n\n
But once you introduce:<\/p>\n\n\n\n
\n- low-res captures<\/li>\n\n\n\n
- cropped elements<\/li>\n\n\n\n
- mobile screenshots with overlays<\/li>\n<\/ul>\n\n\n\n
Accuracy drops.<\/p>\n\n\n\n
Not dramatically, but enough to introduce ambiguity.<\/p>\n\n\n\n
And again, the model doesn\u2019t express uncertainty clearly.<\/p>\n\n\n\n
It just\u2026 guesses.<\/p>\n\n\n\n
\n\n\n\nWe compared it with other vision models, including GPT-5.5\u2019s multimodal capabilities.<\/p>\n\n\n\n
GPT-5.5 felt more conservative.<\/p>\n\n\n\n
Less detailed, but also less likely to overinterpret.<\/p>\n\n\n\n
DeepSeek VL is more aggressive in analysis.<\/p>\n\n\n\n
Which is useful when it\u2019s right.<\/p>\n\n\n\n
Problematic when it\u2019s not.<\/p>\n\n\n\n
\n\n\n\nOne unexpected use case that worked better than UX critique was UI documentation<\/em>.<\/p>\n\n\n\nInstead of asking \u201cwhat\u2019s wrong with this,\u201d we asked:<\/p>\n\n\n\n
\u201cDescribe this interface for documentation.\u201d<\/p>\n\n\n\n
That produced more reliable outputs.<\/p>\n\n\n\n
Because it\u2019s descriptive, not evaluative.<\/p>\n\n\n\n
Less room for misinterpretation.<\/p>\n\n\n\n
We ended up using it more for:<\/p>\n\n\n\n
\n- onboarding docs<\/li>\n\n\n\n
- internal UI explanations<\/li>\n\n\n\n
- quick interface summaries<\/li>\n<\/ul>\n\n\n\n
Than for actual UX decisions.<\/p>\n\n\n\n
\n\n\n\nThere\u2019s also a latency issue when you scale this.<\/p>\n\n\n\n
Single image analysis is fast enough.<\/p>\n\n\n\n
Batch processing hundreds of screens?<\/p>\n\n\n\n
Not as smooth.<\/p>\n\n\n\n
And when you combine it with agents, latency stacks up.<\/p>\n\n\n\n
Not unusable, but noticeable.<\/p>\n\n\n\n
\n\n\n\nOne thing that kept coming up was expectation mismatch.<\/p>\n\n\n\n
DeepSeek VL feels like it should understand interfaces the way a designer does.<\/p>\n\n\n\n
It doesn\u2019t.<\/p>\n\n\n\n
It understands visual patterns.<\/p>\n\n\n\n
Not intent.<\/p>\n\n\n\n
Not user behavior.<\/p>\n\n\n\n
So when you ask it to critique UX, you\u2019re asking it to simulate something it doesn\u2019t fully model.<\/p>\n\n\n\n
Sometimes it approximates well.<\/p>\n\n\n\n
Sometimes it doesn\u2019t.<\/p>\n\n\n\n
\n\n\n\nIf I had to reframe how to use DeepSeek VL for UI\/UX work:<\/p>\n\n\n\n
Don\u2019t treat it as a UX expert.<\/p>\n\n\n\n
Treat it as a pattern scanner.<\/p>\n\n\n\n
\u53ef\u4ee5\uff1a<\/p>\n\n\n\n
\n- surface inconsistencies<\/li>\n\n\n\n
- describe layouts<\/li>\n\n\n\n
- highlight obvious issues<\/li>\n<\/ul>\n\n\n\n
But it can\u2019t reliably judge experience quality.<\/p>\n\n\n\n
At least not yet.<\/p>\n\n\n\n
\n\n\n\nWe still use it.<\/p>\n\n\n\n
Just differently than we expected.<\/p>\n\n\n\n
Less \u201ctell us what\u2019s wrong with this design\u201d
More \u201chelp us process large volumes of UI data faster\u201d<\/p>\n\n\n\n
That shift made it useful again.<\/p>\n\n\n\n
\n\n\n\nSome of the questions that came up while working with it:<\/p>\n\n\n\n
Can DeepSeek VL replace manual UX audits?
No. It can assist, but not replace.<\/p>\n\n\n\n
Why does it misinterpret layout relationships?
Likely because it prioritizes visual proximity over functional grouping.<\/p>\n\n\n\n
Is it better than other vision models?
In some cases, yes\u2014especially with complex visuals. But less reliable in interpretation.<\/p>\n\n\n\n
Can agents fix its mistakes?
Not really. They usually amplify them unless tightly controlled.<\/p>\n\n\n\n
Is this a limitation of DeepSeek or vision models in general?
Feels like a broader limitation. DeepSeek VL just exposes it more because it attempts deeper analysis.<\/p>\n\n\n\n
\n\n\n\nThis isn\u2019t a \u201cuse it or don\u2019t\u201d situation.<\/p>\n\n\n\n
It\u2019s more about using it in the right layer of your workflow.<\/p>\n\n\n\n
If you expect it to think like a UX designer, you\u2019ll be disappointed.<\/p>\n\n\n\n
If you use it to handle visual overload and extract rough structure, it\u2019s actually pretty helpful.<\/p>\n\n\n\n
\n\n\n\nAdvanced Visual Reasoning with DeepSeek-VL and InternVL3<\/a><\/p>\n\n\n\nAnd like most things in this stack right now, the gap isn\u2019t capability.<\/p>\n\n\n\n
It\u2019s consistency.<\/p>\n\n\n\n
Sometimes it gets it exactly right.<\/p>\n\n\n\n
Other times, it confidently explains something that isn\u2019t even there.<\/p>\n\n\n\n
And you don\u2019t know which one you\u2019re getting until you check.<\/p>","protected":false},"excerpt":{"rendered":"
DeepSeek VL can \u201cunderstand\u201d interfaces\u2014until it doesn\u2019t. This is what happens when you try to use it for real UI and UX analysis across messy, inconsistent designs.<\/p>","protected":false},"author":91,"featured_media":1372,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","iawp_total_views":0,"footnotes":""},"categories":[73],"tags":[88],"class_list":["post-3433","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-deepseek-vl","tag-breaking"],"blocksy_meta":[],"_links":{"self":[{"href":"https:\/\/deepseek.international\/zh\/wp-json\/wp\/v2\/posts\/3433","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/deepseek.international\/zh\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/deepseek.international\/zh\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/deepseek.international\/zh\/wp-json\/wp\/v2\/users\/91"}],"replies":[{"embeddable":true,"href":"https:\/\/deepseek.international\/zh\/wp-json\/wp\/v2\/comments?post=3433"}],"version-history":[{"count":3,"href":"https:\/\/deepseek.international\/zh\/wp-json\/wp\/v2\/posts\/3433\/revisions"}],"predecessor-version":[{"id":3436,"href":"https:\/\/deepseek.international\/zh\/wp-json\/wp\/v2\/posts\/3433\/revisions\/3436"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/deepseek.international\/zh\/wp-json\/wp\/v2\/media\/1372"}],"wp:attachment":[{"href":"https:\/\/deepseek.international\/zh\/wp-json\/wp\/v2\/media?parent=3433"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/deepseek.international\/zh\/wp-json\/wp\/v2\/categories?post=3433"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/deepseek.international\/zh\/wp-json\/wp\/v2\/tags?post=3433"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}