What's New in Visuals
Save a character once, and use them in every shot.
Keeping the same person, product or place looking the same across a whole video used to mean wiring the same reference images into card after card and hoping. Now you save that subject once as an Element, tag it in a prompt, and every video model that takes reference images knows who you mean.
The short version:
Save a person, product or place once as an Element
Tag an Element by name in any video prompt on the visual canvas
Type @ to tag any image, clip, audio or text connected to a card
Connected text is used automatically, no more typing a placeholder
Many image models and Runway video now cost fewer credits
Kling V3 Turbo, a faster Kling model with sound built in
Veo respects your start frame again
Elements: keep a character the same in every shot
This is the part to read.
An Element is a named, reusable subject: a person, a product, a place. You make it once, and from then on you can drop it into any video card by name.
Making one
There are 2 ways to define an Element:
From images. A front view, plus up to 3 more angles.
From a clip. A 3 to 8 second video of the subject, 200MB or smaller.
Give it a short name, up to 20 characters, and optionally a line describing it. The name is what you will type later, so pick something you would actually say, like Zoe.
You can start one from scratch, or pick Set as element on something you already have: hover an image or clip in the assets panel on the visual canvas, or right click an image or video card. Either way, your library is right there to choose from, and you can upload too.
Using one
Attach an Element to a video card, then tag it in the prompt by name. It shows up in the card's reference strip with a badge, so you can always see what is feeding the shot, even though there is no connection line on the canvas.
It works on every video model that takes reference images, not just one. Kling V3 Video, Kling V3 Omni, Kling V3 Motion Control and Kling O1 Edit build the character straight from the Element, which gives the strongest likeness. On every other model, its images go in as references and the name in your prompt is pointed at the right one for you.
On a Kling card you can attach up to 3 Elements, or 1 on Motion Control.
What Visuals does so you do not waste a generation
It waits instead of guessing. The first time an Element is used on Kling it takes a moment to get ready. While it does, the run waits rather than going ahead with a weaker result. Open the Element to see where it is on Kling: processing, ready, or not on Kling yet. If Kling setup fails, the Element still works everywhere else through its images.
It never makes the video without your character. If a Kling card with an Element attached cannot run right now, it stops and tells you, and nothing is charged. Run it again shortly, or remove the Element if you would rather go ahead without it.
The Visuals Agent can use them too
Ask for "Zoe" and the Agent finds the Element you saved by that name instead of guessing. It can attach Elements to the cards it builds, and at the end of a character workflow it can save the images it made as a new Element, so the character you just designed is ready for the next video.
You can keep up to 500 Elements on your account.
Tag anything connected to a card
Every video model has its own way of pointing at an image in a prompt. One wants a number in brackets, another a code, another plain words. Before, getting that right was on you.
Now, type @ in a video card's prompt and you get a list of everything wired into that card: reference images, frames, source video, audio, and connected text. Pick one and Visuals writes it the way that model expects. If an input is not actually part of the run, the tag is cleaned up before anything is sent, so a stray code never ends up inside a paid generation.
You can tag an input before it exists. If an earlier card has not generated yet, you can still tag its output, and it holds its place.
Connected text just works. Wiring a text card into another card used to pop up a warning saying the text would be ignored unless you typed a placeholder by hand. Connecting it is the point, so the text is now added for you. It shows as a pill in the prompt you can edit or undo, it appears in the @ list, and it sits in the reference strip as a T tile.
Kling, steadier
Busy moments wait instead of failing. When you run a lot of Kling clips at once, the extra cards wait their turn rather than erroring out.
Content blocks are reported straight away. If a prompt is blocked on content grounds, you hear about it immediately instead of waiting while it tries again.
Time estimates match real runs. A 10 second clip on Kling V3 Omni used to promise about 4 minutes and take nearer 6. The estimate now says so. Kling's estimates are still being tuned as more runs come in.
Kling V3 Turbo
A new, faster Kling model. Text or a start frame to video, 3 to 15 seconds, at 720p on Standard or 1080p on Pro, with multi shot prompts and a negative prompt. Sound is always generated, so there is no audio switch.
It does not take reference images or Elements. For a character you have saved, use Kling V3 Video or Kling V3 Omni.
Kling O1 Edit tells you before, not after
Kling O1 Edit will not take reference images or an Element alongside an end frame. The card now catches that before the run and offers both ways out, instead of letting it fail after you press Run.
Image models and Runway cost fewer credits
Most image models now cost less, some by a lot, and so does Runway video. Your plan and your monthly credits have not changed, so they go further. As always, the exact cost shows before you generate.
GPT Image 2: 1 credit at Low quality, 2 to 3 at Medium, and 8 to 10 at High, depending on the shape. It was 20.
GPT Image 1.5: 1 credit at Low, 2 to 3 at Medium, and 7 to 10 at High. It was 20.
Nano Banana 2: 4 credits at 1K, 6 at 2K and 8 at 4K. It was 5, 8 and 12.
Nano Banana Pro: 7 credits at 1K and 2K, and 12 at 4K. It was 10 and 20.
FLUX Kontext Max: 4 credits. It was 10.
Runway Gen-4.5: 30 credits for 5 seconds and 60 for 10. It was 50 and 100.
Runway Gen-4 Turbo: 12 credits for 5 seconds and 24 for 10. It was 20 and 40.
Two things worth knowing:
With reference images, GPT Image can still cost 20 credits. The lower prices are for images made from a prompt alone.
FLUX 2 Pro depends on your references. On its own it is 2 credits, down from 8. With reference images it stays at 8, and it can cost more when you give it several large ones.
You are never charged more than the price you saw. If an image would cost more by the time you generate, it stops and shows you the new price first.
Also fixed
Veo respects your start frame again. Veo cards with both a start frame and a reference image connected were failing, sometimes with a "rate limited" error, and the clips that did come back ignored the start frame. Veo cannot take both in the same clip, so now the start frame always wins and the reference image is left out.
The Agent stopped steering people away from the default model. It was told Gemini Omni 1.1 Flash could not handle recognizable people and was quietly sending those flows to Veo 3.1 instead. That was out of date. It now picks by what the shot needs.
Double clicking a finished video opens that video. It could open another card's image instead.
The version strip shows which version is selected. It showed none.
Plus assorted polish on prompt pills, previews and error messages.