What's New in Visuals
The live run card from 10.14 now stays put, keeps every card, and tells the truth about how the run went.
Press Run on a Ready card and that same card becomes the run card, right where it is. It holds the whole plan from start to finish, and it catches up if a result lands late.
The short version:
The Ready card becomes the live run card, in place
Every planned card stays listed, and finished rows open the result
More "run only this card" requests now run, and nothing else changes
Queued Agent messages come back after a reload
Sign up shows the password rules before you submit
Up to date ChatGPT setup steps in Settings
A better live run card
It stays in one place. Press Run on a Ready to run card and it turns into the live run card, in the same spot.
Every card stays listed, in order. Cards that haven't started say Waiting, and the layout stays the same for the whole run.
Open any result. Click a finished row to see it in the full viewer. A card being made shows a soft glowing tile, and images fade in as they load.
2 views after a run. The chip above the message box shows Last run, for example stopped at 3 of 5, and Full flow. Full flow lists every card, shows the total cost, and offers Run all and Make test video.
No false stops. If a result lands just after the run stops reporting, the card catches up and the Run stopped label goes away.
Runs that finished while you were away show up. If every tab was closed, reopening the flow shows that run and its results.
The Agent sticks to what you asked
More requests now run. "Run only Video Card 1", "rerun this" on an attached card, "fix this card but keep the other one unchanged" and "run this saved flow without changing it" used to stop without running. Now the Agent runs just those cards and keeps every other card and connection as it was.
It builds only the cards you described. It no longer adds a treatment, writer or reference card you didn't ask for, which also cost credits.
Audio only means that card. Asking for audio only on one video card no longer changes what the other cards need.
Edits stay in sync. When the Agent edits a text card's prompt, the card's output follows the new instructions.
Queued messages survive a reload
Messages you queued while the Agent was busy were lost if you left or reloaded the page. Now they come back when you return to the same conversation in the same browser tab. They come back paused, marked Queued · waiting for you to send, so nothing goes out by surprise.
Sign up, clearer
The sign up form said to use at least 6 characters, but the real rule was stricter, so many people were turned away again and again. Now sign up and password reset show a live checklist as you type:
At least 10 characters
A lowercase letter
An uppercase letter
A number
It also asks you to avoid common or leaked passwords. If you try to sign in before confirming your email, you now see Confirm your email to sign in with a Resend confirmation email button.
Connect Visuals to ChatGPT and other AI apps
Current ChatGPT steps. Settings, API & MCP shows the steps that match ChatGPT today.
Results are kept. If a finished generation can't be read right away, the AI app is told to check the same run again instead of reconnecting.
Real reasons. Failed generations say why, and whether to try again.
Faster with big libraries. Browsing and searching a very large library from an AI app is much quicker, and starting several generations at once is more reliable.
Also fixed
New chat starts a new chat. Your next message could land in the old conversation.
Stop waits for the run. Right after you press Run, the button says Starting… until the run really begins, and a stopped run stays stopped.
Lyric videos from an image in My Library. Choosing Remix, then As Lyric Video, keeps your image after you pick a song, and Back takes you where you expect.
Seedance 2.0 reference audio needs a picture. Seedance 2.0, 2.0 Fast and 2.0 Mini need a reference image or video alongside reference audio. You now see that before the run, not as an error after it. Seedance 2.5 and Gemini Omni 1.1 Flash still take audio on its own.
Trim video is exact. The clip sent to the model starts and ends where you trimmed it.
ZIP downloads no longer hang, and an interrupted download no longer says it finished.
More accurate price estimates. The price you see before a run now matches what you pay in more cases, including images with saved references and requests from AI apps for several results at once.
Folders and collections create every time in My Library, and the page stays clickable after menus close.
Render details no longer show Audio: Off on videos that have sound.
Plus assorted polish on the Agent chat, notices and loading states.