The idea is simple and it is the reason this works so well. You take a view of a real place, draw a red line across it showing where you want the drone to fly, and hand both the image and a very specific prompt to a video model. The model reads the red line as the flight path and renders a continuous FPV shot racing along it.
The plan · 3 steps
- 1Draw Your Flight Path
- 2Connect Higgsfield To Claude
- 3The Master Prompt
The key word there is FPV. You are not making a slow, floaty real estate flyover, you are making the fast first person stuff that banks around buildings and dives between them, which is exactly the footage that performs on social. It is also a POV from the drone rather than a shot of a drone, and that distinction has to be spelled out in the prompt or the model will happily put a quadcopter in the middle of your frame.
You are running the video models through Claude rather than going to a separate site, which is what the Higgsfield connector is for. Claude handles the prompt, sends the job, and brings the finished clip back into your chat.
What You Need Before You Start
A Claude account, a Higgsfield account with an active paid plan (the connector will not generate on a free account), and any way to draw a line on a screenshot. Your phone's photo markup tool is completely fine, you do not need Photoshop.
Setup is a one time thing and takes about two minutes. After that each clip takes a few minutes to come back, sometimes longer when the queue is busy. Plan for a handful of attempts rather than one perfect shot first time, because the first render is usually the one that teaches you what to change in the prompt.
Heads up. Clips come out short. Gemini Omni, the model this guide uses, does 4 to 10 seconds and 8 is the sweet spot. If you want a longer sequence you generate several clips and cut them together, which is also how real FPV edits are made, so it is less of a limitation than it sounds.
Step One: Draw Your Flight Path
This is the part that decides how good the result is, and it takes about a minute.
1. Find your location. Open Google Earth and search for any spot on the planet. Zoom in until the frame shows roughly the area you want the drone to cover. For city FPV shots, tilt the view so you can see the buildings standing up rather than looking straight down, because that gives the model a much better read on what it is flying through.
2. Take a screenshot. Grab a clean screenshot of that view. Crop out the browser chrome, the search bar and any interface panels, because anything left in the frame is something the model might try to render into your video.
3. Draw the path in red. Open the screenshot in any markup tool and draw a single clear line showing where the drone flies. Use red, because that is what the master prompt refers to by name, and mark the start of the line with a dot so the direction is obvious.
4. Save it. Save that marked up image to your photos or downloads. That is your input.
Tip. Draw a path a real FPV pilot could actually fly. Sweeping curves between buildings, a dive down a street, a bank around a tower, all good. What does not work is a line that doubles back on itself or turns at a hard right angle, because no footage the model has ever seen moves like that and it will either ignore the turn or make something strange out of it.
The line is not a literal camera track the software follows frame by frame. It is a strong visual instruction that the model reads alongside your prompt, so the clearer and simpler the line, the closer the result lands. That is also why the prompt still has to describe the move in words.
Step Two: Connect Higgsfield To Claude
This is the one time setup. You are adding Higgsfield as a custom connector so Claude can send generation jobs to it directly.
1. In Claude, click Customize, then Connectors.
2. Click Add, then Add custom connector.
3. Paste this in as the connector URL: https://mcp.higgsfield.ai/mcp
4. Hit Connect. A Higgsfield window opens in your browser asking you to authorise the connection. Sign in and approve it, and if you are already logged in it will just go through on its own. Then head back to Claude.
5. Higgsfield should now be showing in your connector list as connected. If it is not, the usual cause is the authorisation window being closed before it finished, so run it once more and let it complete.
Heads up. The connector will attach fine on a free Higgsfield account and then fail when you actually try to generate, which reads like a broken setup when it is a billing thing. Video generation needs an active paid plan on the Higgsfield side.
Step Three: The Master Prompt
Open a new chat, make sure Fable 5 is selected in the model picker, upload your marked up screenshot, and paste this in underneath. That is the whole thing, no filling in blanks.
Here's a screenshot from Google Earth with a red line drawn on it. Generate a photorealistic, cinematic high-speed FPV drone video that follows that exact red flight path through the city - fast smooth motion, banking around buildings, golden-hour lighting, motion blur, 9:16 vertical. Make it look like real FPV drone footage. Use Higgsfield and the Gemini Omni video model. This should be a POV from the drone not of the drone itself.
Why Each Block Is In There
Every line in there is fixing a specific way these shots go wrong.
POV from the drone, not of the drone. Without this you get a shot of a quadcopter flying through the scene, which is the single most common failure and the one that ruins the illusion instantly.
One unbroken take. The model's instinct is to cut between angles like a trailer. Naming it stops the cut.
Bank into the turns, horizon rolls. This is the whole difference between FPV and a floaty real estate flyover. Without it the camera slides sideways like a video game.
Natural motion blur. Fast motion with perfectly sharp frames is one of the strongest tells that something was generated.
Carries momentum through turns. Stops the drone changing direction instantly, which no object with mass does.
Photorealistic and motion blur. Fast motion with perfectly sharp frames is one of the strongest tells that something was generated.
Tip. When a render comes back wrong, do not rewrite the whole prompt. Find the thing that went wrong, add one short line to the end of the prompt naming it directly, and run it again.
FPV Moves Worth Stealing
Add one of these to the end of the prompt if you want a specific move. They are the moves that read as real FPV rather than as a camera drifting through a scene.
The dive. Start above the rooftops, drop hard down between the buildings, level out at street height. Best for city openers, and the most reliably impressive one.
The weave. Bank left and right through a line of buildings, rolling into each turn. Best for dense downtown blocks.
The chase. Race low and fast along a street just above traffic. Best anywhere with strong lines leading somewhere.
The orbit break. Circle a landmark once, then break away and accelerate out. Best for one hero building.
The climb out. Race at low level then pull up hard and clear the skyline. Best for endings and anything you want to feel like a reveal.
Light Does Most Of The Work
Time of day changes the result more than any other single word in the prompt. Golden hour and blue hour both give you long shadows and warm or cool separation that reads as real footage, and golden hour raking between buildings is the reason so much of this stuff looks expensive. Midday sun flattens everything and is the fastest way to get something that looks synthetic.
The Model, And Where It Caps Out
The master prompt asks for Gemini Omni because it handles 9:16 vertical natively, generates its own audio, and reads reference images well, which is exactly this job. One thing to know is that it maxes out at 720p, which is completely fine for social but is worth knowing if you need something bigger.
Gemini Omni. 4 to 10 seconds, 720p. The default here. Vertical, native audio, strong reference handling.
Seedance 2.0. 4 to 15 seconds, 4K. When you need real resolution for a bigger screen.
Kling v3.0. 3 to 15 seconds, 4K mode. When you want several shots in one generation.
MiniMax H3. 4 to 15 seconds, 2K. When you want to control the first and last frame precisely.
Grok Video 1.5. 2 to 15 seconds, 1080p. When the camera move itself is the whole point of the shot.
Tip. Generate two or three variants of the same prompt before you change anything. These models are not deterministic, so the difference between attempts of the same prompt is often bigger than the difference between two prompts.
Where People Get This Wrong
Forgetting the POV line. Leave out the line about this being a POV from the drone and you will get a video of a drone flying around. It is one sentence and it is the most important one in the whole prompt.
Leaving interface elements in the screenshot. Search bars, compass widgets, copyright strips and browser tabs all get read as part of the scene, and you end up with strange floating artefacts in the sky. Crop hard before you draw anything.
Drawing a path nothing could fly. FPV paths can be aggressive, that is the point, but they still obey physics. Fast and sweeping is the target, not sharp and impossible.
Zooming in too far in Google Earth. Past a certain zoom the imagery goes soft and the 3D geometry starts melting, and the model faithfully reproduces that mush. Back off until the frame looks crisp, then let the prompt bring the drone lower instead.
Asking for one long clip. Pushing for the maximum duration usually gives you a shot that drifts or repeats near the end. Eight seconds holds together well, and cutting three good eight second runs together beats one shaky ten second one.
Judging it on the first render. The first attempt is reconnaissance. Look at what went wrong, name that thing explicitly in the DO NOT block, and run it again. Two or three passes is normal.
A good first test
Pick a city skyline you know well, draw one sweeping curve that starts above the rooftops and drops between two towers, set golden hour, ask for eight seconds at 9:16, and generate three variants. City dives are forgiving because the speed and motion blur hide small errors.
The Honest Part
Check the source imagery before client work. Google's own Geo Guidelines say Earth content may not be used for commercial or promotional purposes, that attribution is required where it is used, and that you should not use Earth output to create other content or products. I am not a lawyer and how that lands on an AI generated video built from a screenshot is genuinely unsettled, but if you are doing paid work for a client it is not a question you want to find out the hard way.
The easy workaround. The workflow does not care where the aerial view comes from. Use your own drone still, a photo you took, a properly licensed stock aerial, or skip the image entirely and describe the location in text.
It is not the same as real drone footage. Look closely and buildings morph slightly as the drone passes, small details like cars and windows are invented rather than reproduced, and anywhere you actually know well will look subtly wrong to you. It is superb for openers, mood pieces and concept work, and it is not a substitute for anything where the real geography has to be accurate.
The costs are real. Every generation spends Higgsfield credits, and since the workflow depends on running several attempts you should expect to spend meaningfully more than one clip's worth to get one clip you are happy with.
Be straight about what it is. If footage like this is going into something where the audience would reasonably assume a camera was actually there, say it is AI generated. That is increasingly just the expectation, and it costs you nothing when the work is good.
Quick Reference
1. Flight path. Google Earth, find location, tilt for buildings. Screenshot, crop out all interface. Draw one red line, dot the start, save it.
2. Connector (one time). Claude, Customize, Connectors, Add custom connector, https://mcp.higgsfield.ai/mcp, connect and authorise.
3. Generate. New chat, select Fable 5, upload the marked screenshot, paste the master prompt, fill the brackets, send.
The settings that matter most. Aspect 9:16 for social. Duration 8 seconds, long enough to read and short enough to hold together. Golden hour, because it flatters everything and hides small errors. A 14 to 18mm lens with slight barrel distortion. A fast, sweeping path with no hard angles. And three variants before you change anything.
Draw the line. Let the drone fly it.