Grok rewrites this into a full cinematic brief before it reaches the model.
One paragraph asks the model to guess where a move happens. A shot list tells it. Durations must add up to the total.
Aspect ratio
Resolution 4K costs double
1K and 2K are billed the same, so 2K is the floor worth using.
Duration 5 seconds
Billed per second, so this is the price dial. 3s is the cheapest real test at $0.34.
Kling writes the audio from your words. Name the song, the line, or the room tone or it comes back near-silent.
End on closes the loop
Ending on the opening frame makes a clip that loops with no cut. Free.
Creativity
How far Krea may drift from your words. Raw obeys literally, High reinterprets. Medium is the model's own default.
Krea cannot compose your photos together — it borrows their look. For combining pictures, use Compose.
Style strength 0.5
Every source in the tray is merged into one picture. Up to 14, and the count is free.
Seed lock it to iterate on the same look
Same seed and same prompt gives the same picture. Lock it, change one word, and whatever moves is what that word did.
Enlarge by 2.0×
Priced on the picture that comes out, so this is the price dial — and it is not limited to whole numbers.
Model
Uses the first source, at full resolution.
fal does not publish a price for the generative models. Treat the meter as a floor, not a quote.
What to bring back Redefine only
Invention 3 of 6
Texture 3 of 5
Advanced
Prompt adherence model default
Higher sticks closer to your words and further from the source image. The untried lever for a camera move Kling keeps ignoring.
Keep out
Leave empty to use Kling's own default, which is exactly that placeholder.
Shot structure
This is about how the clip is cut into shots, not about camera moves.
Build a person or object from a front-on image plus up to 3 other angles, then call it @Element1 in the prompt. A voice id makes it speak — and moves the whole clip to $0.196/s.
Content filter model default (4)
1 blocks the most, 6 the least. Every refusal we have ever had ran at the default 4 and it has never been tried higher.
Standing instruction
Steers style across the whole request, separately from the prompt.
For a real place, product or person it may not know. Costs a penny and a half.
On by default in Topaz. Turn it off for a subject with no face — it will otherwise find one.
Face strength model default
Focus on
Sharpen off
Denoise off
Fix compression off
All three are off unless you move them, which leaves Topaz on its own defaults.
Nothing extra for this one. Krea's controls are all above.
Cost this run
$0.56
Kling 3.0 Pro 5s, no audio
This session $0.00 spent
Renders collect here. Tap one to put it back on the stage.
Made with this
Every one of these started as a sentence. Tap any picture to read the
sentence that made it, then make your own from the same starting point.
Start here
You describe something, and a few minutes later it exists. A photo of a place that was never photographed. A short film of your dog on the moon. An old picture from a shoebox, sharp again.
There are four things this can do, and you pick one before you start:
Video
A short clip, with sound, made from a sentence.
Image
A single picture, made from a sentence.
Compose
Hand it several photos, get one picture with all of them in it.
Upscale
Take a picture you already have and make it bigger and clearer.
Each one uses a different piece of software behind the scenes, and they are good at different things — but you never have to choose between them. Pick what you want to make, and the right one is already loaded.
Just write like you are talking to a person. Your words get expanded into a longer, more detailed brief before anything is made, so you do not need special phrasing or a list of keywords. Say what you want to see.
Every run costs real money — a few cents for a picture, more for video. The exact price is on the button before you press it, and it is never a surprise. If nothing comes back, you are not charged. Refusals, crashes and runs you give up on are all free.
Words you will see
Prompt
What you type. The description of what you want.
Render
One attempt. You press Generate, it thinks, a picture or clip comes back.
Source
A photo you supply, to be combined, enlarged, or used as a starting frame.
Aspect ratio
The shape of the result. Tall like a phone, wide like a TV, or square.
Seed
A number that decides all the small stuff. Explained properly below.
Your first four runs
In this order. It takes about ten minutes and costs well under a dollar, and by the end you will know how the whole thing behaves.
Pick Upscale and give it any photo from your phone. $0.08, nothing to write, back in under a minute. Start here because there is nothing to get wrong — it just proves everything works.
Now Image. One honest sentence about something you would like to see. $0.06. When it arrives, look at the bottom of the picture: there is a number labelled seed. Tap it.
Change one word of your sentence and run it again. Because you tapped the seed, everything else stays put — so whatever changes in the picture is what that word did. This is the trick that teaches you the most, and it is the reason the seed exists.
Finally Video, set to 3 seconds with sound off — $0.336, the cheapest real clip. Video is the expensive one: fifteen seconds with sound is $2.52, so learn on short ones first.
The four jobs
Video
$0.336+
Kling 3.0 Pro. 3 to 15 seconds, billed per second. Sound written from your words.
Image
$0.06
Krea 2 Large. Words to picture. Draws real people and copyrighted characters without complaint.
Compose
$0.15
Nano Banana Pro. Merges the photos you hand it into one picture. Refuses likenesses.
Upscale
$0.08
Topaz. Enlarges something you already have. No prompt.
Image and Compose are not a cheap and an expensive version of the same thing. Krea is the one that will draw anyone; Nano Banana Pro is the one that can combine your photos. Pick on what you need, not on price.
If Compose refuses to draw your subject, switch to Image — it will draw almost anything, and it costs less. There is no third option hiding somewhere that does both; this pair is as good as it currently gets.
Seed
You will never write a sentence detailed enough to describe a whole picture. You might ask for a man in a red coat on a beach — but which way is he facing? Morning or evening? Is the beach empty? Something has to decide all of that, and that something is a random number called the seed.
Think of it as which shuffle of the deck you were dealt. The same words with the same seed always give the same picture. The same words with a different seed give a completely different take on the same idea — still a man in a red coat on a beach, but a new one.
Here is why you should care. Say a picture comes back that you nearly love, and you want the coat to be blue. Change the word and you also get a new random seed — so you get a different beach, a different man, a different everything. The thing you liked is gone.
So keep the seed and change the word. Now the only thing that can move is the coat. That is how you steer a picture instead of gambling on it, and it is the most useful habit here.
In practice: when a picture arrives, its seed is shown just underneath it. Tap that number and it drops into the Seed box, ready for your next run. The padlock holds it there, the dice rolls a fresh one, and an empty box means a new random seed every time — which is what you want while you are still exploring.
Only Image and Compose have a seed. Video and Upscale do not, so the box is not shown for those two.
What each job does with the tray
Compose
Merges every source into a single picture.
Image
Borrows their look, if you tick style references. Never copies them in.
Video
First source is the opening frame. Any source can be the last one.
Upscale
First source, at full resolution.
Photos you add get shrunk a little before being sent, so adding several does not mean a long upload. Upscale is the exception and always sends your original at full size — shrinking a picture you are about to enlarge would rather defeat the point.
Image — Krea
Creativity is how far Krea may drift from your words. Raw takes them literally, High reinterprets freely. Medium is the model's own default and the place to start.
Eight aspect ratios, including 2.35:1 — anamorphic cinema, the widest shape in the rig. Only Krea offers it.
Image uses the seed, and it can use your photos too — though not in the way you might expect. See the next line.
It treats sources as style references, not as content — it borrows their look, it does not put them in the picture. Tick Use sources as style references and set the strength. That is the only thing on Krea that costs extra, and it is half a cent. To actually combine photos, use Compose.
Compose — Nano Banana Pro
Every source in the tray is merged into one picture. Up to 14 of them, and the number you put in is free — you are billed per picture that comes out. Give it a face, a jacket and a room, and ask for all three together.
Order matters to the model and the tray is numbered, so describe them the same way: the man in image 1 wearing the coat from image 2.
Always use 2K. There is no 1K button because 1K costs exactly the same and gives you less. 4K costs double and is only worth it for printing, or a very big screen.
Under Advanced: Content filter runs from 1 to 6 and sits at 4. Every refusal so far happened at that setting, and nobody has tried loosening it yet. Standing instruction steers style across the whole request, separately from the prompt. Let it check the web is for a real place, product or person the model may not know, and costs a penny and a half.
Video — Kling
Per second
$0.112
Any whole number of seconds, 3 to 15.
With audio
$0.168
Half again as much.
With a voice
$0.196
Giving a named character a voice moves the whole clip to this rate.
Expect about 44 seconds of waiting for every second of video, so a 10 second clip takes roughly 7 minutes. Go and do something else — it keeps going without you, and you can even close the tab and come back.
The sound is written from your words. A prompt with no sound cues comes back nearly silent. Name the song, the line of dialogue, the room tone — whatever you want to hear, say it.
Start from a photo and the clip copies that photo's shape. The aspect ratio buttons disappear, because Kling has no such setting when it starts from an image. If you want a 9:16 clip, start from a 9:16 picture.
Directing a clip
Direct it shot by shot. One paragraph asks the model to guess where a move happens; a shot list tells it. The durations have to add up to the clip length and the rig will not let you submit if they do not. This is the real answer to a camera move Kling keeps ignoring.
Name a character from the tray. A front-on picture plus up to three other angles becomes @Element1, and you call it that in the prompt. Use it to hold the same face, or the same jacket, across a whole clip. Giving it a voice makes it speak — and moves the clip to $0.196 a second.
End on picks which source is the last frame. Ending on the opening frame gives a clip that loops with no cut, and it is free.
Prompt adherence, under Advanced, decides how literally your words are followed. Turn it up and the clip sticks closer to what you asked for and drifts further from the starting photo. Leave it alone unless a camera move keeps getting ignored — then it is worth a try.
All of this is free. Shot lists, characters, end frames, adherence and keep-out words change the clip, never the bill.
Upscale — Topaz
Up to 24 MP
$0.08
A 2K picture at 2× lands here.
Up to 48 MP
$0.16
2K at 3×.
Up to 96 MP
$0.32
2K at 4×.
The price is set by what comes out, so the meter works out the exact figure from your source instead of guessing. Enlarge by is not limited to whole numbers — a 12 MP photo at 1.4× is $0.08, at 2× it is $0.16. Dial it down until the meter drops a tier.
Match the model to the source. Standard for photographs, High Fidelity when there is fine detail worth protecting, Low Resolution for something small or soft, CGI for renders and illustration, Text Refine for screenshots and type, Recovery for something damaged or heavily compressed.
The generative models — Redefine, Wonder, Wonder 3 and Standard MAX — invent new detail instead of rebuilding what is there, and fal publishes no price for them. The meter says "from" and means it. Redefine takes a prompt of its own, which is how you tell it what to bring back.
Faces are enhanced by default, at Topaz's own strength. Turn it off for a subject with no face or it will find one anyway.
What costs extra
Audio on a clip
+50%
$0.112 a second becomes $0.168.
A character voice
+75%
$0.196 a second, for the whole clip.
Style references
+$0.005
Image only.
Web search
+$0.015
Compose only.
4K
×2
Compose only. $0.15 becomes $0.30.
Everything else is free. Shot lists, characters without a voice, end frames, prompt adherence, keep-out words, seeds, aspect ratios, creativity, the content filter, a standing instruction — none of them move the price.
And more sources never costs more. You are billed for pictures that come out, not pictures you put in.
Working in the console
The stage keeps your last result while you set up the next run, so you can look at what you got while you change what you asked for.
Everything you make drops into the session strip. Tap one to bring it back to the stage; the small arrow sends it into the source tray, so you can build on it, upscale it, or make it a clip's opening frame.
The strip survives a refresh, a crash and a browser restart, because it is stored on this device. It does not follow you to another device, and Clear empties it.
You can close the tab mid-render. Come back within an hour and the rig picks the job back up where it left off, so a paid render is no longer lost by closing a tab.
The passphrase is remembered for two hours, then the rig asks again. It spends real money, so it should not stay unlocked forever.