Qwen output UX: images, video, search & selection

What works

• The assistant reply is a full art-direction brief (yarn doll, lighting, texture) instead of jumping straight to pixels.
• User sees the plan in chat before spend. Easier to edit text than regenerate blindly.
• Create Image chip and Qwen-Image 2.0 / 16:9 stay pinned in the composer through the reply.

What we would push on

• The prompt is long and single-block. No highlight of what changed from the upload.
• Send stays greyed while mode chips are set. Unclear if user must approve the text prompt first.

Product bet

Showing prompt expansion builds trust for image spend. Users feel the model understood the reference photo before GPU time runs.

Tradeoff

Takeaway

For reference-image gen, show the rewritten prompt in-thread and let users edit it before render.

What works

• Image renders inline at readable size with share, regenerate, and overflow on the message.
• Create Video and Edit float on the image. Next steps are on the artifact, not buried in +.
• Thumbs, share, and download also sit top-right on hover for quick feedback and export.

What we would push on

• Composer still shows 16:9 while the output image is square. Format chip and result can disagree.
• AI-generated content may not be accurate disclaimer is easy to miss under the bar.

Tradeoff

Takeaway

Put chain actions on the image itself. Match aspect label to what actually rendered.

What works

• Source thumbnail appears in the bar above animate it. Context is visible before send.
• Create Video chip replaces Create Image automatically. Mode follows intent.
• Short verb prompt works. User did not re-describe the doll.

What we would push on

• No motion controls (duration, camera move) in the bar yet, only 16:9 later.
• Chip swap is smart but silent. A one-line "Switching to video from this image" would teach the pattern.

Product bet

Image-to-video is the upsell path. Reference thumbnail plus auto chip swap keeps users in one thread instead of opening a video app.

Tradeoff

Takeaway

When a reply references prior media, pin the thumbnail and swap mode chips to match the next job.

What works

• Gradient card with spark icon and 4% label. Users know video is rendering, not stuck.
• Create Video chip and 16:9 stay in the composer so settings survive the wait.
• User bubble animate it stays minimal. Focus is on the job card.

What we would push on

• No ETA or stage name beyond percent. Long video jobs will feel opaque.
• Send button greys out. Cannot queue a follow-up until render finishes.

Takeaway

Show percent for video gen, but add stage copy or ETA when renders run longer than a few seconds.

What works

• Modal player with 00:00 / 00:05 timer. Short clip length is obvious before download.
• Download and Publish are equal primary actions under the player.
• Create Video chip and 16:9 remain in the bottom bar for another pass without closing context.

What we would push on

• Publish destination is not named on the button. Trust requires knowing where it goes.
• Square video in a 16:9 modal adds letterboxing. Same aspect mismatch as image output.

Tradeoff

Takeaway

Treat video like a deliverable: preview, duration, download, and name the publish target on the button.

What works

• Lead sentence states intent: I'll search for information about the best budget laptops…
• Searched the web row is tappable. Signals retrieval happened.
• Inline grey domain chips after claims (pcworld.com, pcmag.com, cnet.com) mirror Perplexity-style trust chrome.
• Structured list with bold product names and price line reads scannable.

What we would push on

• Answer cites 2026 reviews and an Apple MacBook Neo at $599. That product does not exist. The UI still looks authoritative.
• Domains are not numbered footnotes. Hard to match a claim to a specific source click.
• Disclaimer under the bar is tiny compared to the confident list above it.

Product bet

Search mode competes on Perplexity-shaped answers. Citations and Searched the web sell freshness even when the model fills gaps with fiction.

Tradeoff

Takeaway

If you show domain chips, tie each claim to a numbered source and surface when retrieval failed or dates look wrong.

What works

• Sidebar title Thinking and Search sets expectation: reasoning plus retrieval, not chat history.
• Source pills (wired.com, cnet.com, pcworld.com) and +14 show breadth at a glance.
• Numbered results include page titles and snippets so users can sanity-check before trusting the summary.

What we would push on

• Snippets also reference 2026 and MacBook Neo. Transparency reveals the same hallucination problem, not a fix.
• Panel is optional. Many users will never open it and only see the polished main answer.

Tradeoff

Takeaway

A source drawer helps only if snippets are real. Pair it with date checks and a low-confidence state in the main answer.

What works

• Floating menu: Copy, Ask Qwen, Explain, Translate with a language submenu.
• Translate lists many locales in one scroll (English US, Arabic, German, Spanish, Persian, French, Hindi, Italian, Japanese, Korean…).
• Selection works on search answers, not only plain chat prose.

What we would push on

• Ask Qwen vs Explain overlap. Users will not know which to pick.
• No Check sources or Cite selection action despite search mode being on.

Tradeoff

Takeaway

Extend selection menus for search threads with source-check actions, not only translate and explain.

What works

• Three follow-up pills continue the research thread (under $500, battery life, what to look for).
• Message actions row includes copy, feedback, share, regenerate, and globe badges with +17.
• Web search chip persists in the composer so the next turn stays in search mode.

What we would push on

• Follow-ups repeat the 2026 framing. Bad suggestions reinforce bad facts.
• +17 globes are not clickable domains in this view. Count impresses more than it informs.

Tradeoff

Takeaway

Follow-ups should deepen verification, not repeat the same shaky premise. Make source badges open the drawer.