Welcome to Nomad
Nomad is a visual workspace for architects. You drop in a screenshot of your model, say what you want in a sentence, and it writes the prompts, wires the canvas, prices the work and renders it — stills, camera moves, multi-shot films and reels — while your building stays exactly your building.
Your first render
You need one thing to start: a screenshot of your model. Any viewport, any modeller, any resolution. Nomad reads what it can see.
Drop the screenshot on the canvas
Drag the file in, paste it with Ctrl+V, or use the upload
button in the dock. It becomes image 1.
Add a note and say what you want
Press N or use the toolbar. Write it as you would to a colleague: "villa in Dubai, evening, from the courtyard". No syntax to learn.
Wire the reference into it
Drag from the right-hand port of the image to the note. The wire lights up when it connects.
Answer the question
Nomad asks about what it cannot see — usually the material, sometimes the place or the season. Three or four options, or type your own.
Press play
The prompt is written and the image renders. The cost is on the button before you press it, and the pre-flight check runs first, free.
The canvas replica is open to anyone. Same workspace, same panels, renders already in place — you can walk the whole thing before deciding anything.
The canvas
An infinite board. Everything lives on it: your references, the prompts, the renders, notes to yourself, and the slides you will present from.
- Pan — drag anywhere empty, or hold space.
- Zoom — scroll, or the buttons in the dock. Ctrl+0 fits everything.
- Select — click. Ctrl-click adds to the selection.
- Marquee — Ctrl-drag across empty canvas.
Nodes
They behave differently on purpose.
| Node | What it holds | What it does |
|---|---|---|
| image | A picture you dropped, pasted or rendered | Feeds anything downstream as a reference |
| prompt | Your note on top, the written prompt below | Turns a sentence into a proper prompt |
| slide | A sheet at A3, A4 or 16:9 | Composes images and text into something presentable |
| video | A clip, with its length and engine | Renders a camera move or a whole film from the pictures wired into it |
| note | Whatever you type | Nothing. It is for you. |
A rendered image stops being a machine the moment it exists — the settings disappear and it behaves like any other picture, so you can move it, resize it and place it on a slide. It still remembers what made it: select it and you can re-render or upscale.
Wires
Drag from a node's right-hand port to another node. Drag left to disconnect. A node can take several inputs; they stack on the left edge in the order you wired them, and that order is the order Nomad reads them in.
If your note says "the villa from image 1 in the image 3 context", Nomad assigns those roles by number. Image 1 becomes the locked source; image 3 becomes setting only. Rewire and the roles follow.
Groups and slides
Select several nodes and press Ctrl+G. They get a translucent box with a name and a live credit total, and moving the box moves everything inside. Ctrl+Shift+G releases it — the contents survive.
Slides are frames at real paper sizes. Select one and press Compose to have Nomad fill it, or drag images in yourself. Present full screen straight from the canvas with the arrow keys.
How Nomad writes
You type a line. What comes back is a blocked master prompt of 250 to 400 words — naming the geometry that must survive, the single thing that changes, the lens, the aperture, the light and the grade, in the order that survives contact with a model.
It asks before it writes, but only about things it genuinely cannot see. Material is the usual one: geometry tells it the shape, not whether that surface is concrete, travertine or corten.
Keeping your geometry
This is the rule everything else is built around. Every image-to-image prompt names the features that must survive, states the failure it is preventing, and repeats the lock at the start and the end. It is not a setting you can forget to switch on.
A model can still drift. What changes is that drift becomes rare and correctable. When it happens, say what is wrong in plain words — "the roof pitch changed" — and Nomad rewrites the prompt against the failed image, naming that exact failure so it cannot repeat.
Storyboards and sets
Press Storyboard with a reference selected. Choose six, twelve or eighteen shots. Nomad plans the sequence, writes every prompt, wires each one to your reference, sets the ratio per shot and quotes the total before spending anything.
The sequence follows how published project sets are actually built: the building small in its landscape, an establishing three-quarter, an honest oblique, a material study in flat light, a threshold, a principal interior, a corridor, a junction, people using it, and one dusk frame at the end.
A set of nothing but hero shots reads as a render pack. The corridors and junctions are what make the hero believable.
Slides and boards
Describe a slide in a sentence and Nomad composes four coordinated images, a title, a subtitle and a written concept, laid out inside the sheet. It renders them in order and the slide fills as they arrive.
Plans, sections and elevations are never generated. A board reserves the slot and labels it, because a drawing that looks authoritative and is wrong is worse than a blank space.
Upscaling
Select a rendered image and press Upscale. Only resolutions above the current one are offered, and you can keep the ratio or recompose while you are there.
The upscale brief is deliberately strict: this is an upscale, not a reinterpretation. Every element must survive — architecture, materials, composition, light, sky, people, colour grade. What gets added is resolution only.
Editing a picture
Double-click a render to open the picture editor. Exposure, contrast, colour and finish — clarity is a real unsharp mask and grain is turbulence, not an overlay image, so both hold up at 4K. Looks are starting points: pick one and the sliders move to match, nudge any of them and it becomes yours.
Edits are non-destructive and live on the node. Copy them from one picture and paste them onto another, or press Start again to return to how it arrived.
Replacing a picture
Right-click a picture and choose Replace picture…, or simply drag a new image straight onto the existing one. The node survives: same id, same wires, same number, same place on the canvas. Only the picture changes.
Because a reference wired into three prompts and a slide would orphan all four. Replacing keeps every connection, so a prompt that says image 3 still means image 3. If the new picture is a different shape the box keeps its width and re-derives its height, and says so.
Video, and the four engines
Wire a render into a video node, say what the camera does — "slow push through the court, hold on the west wall" — and it renders the move. Video bills by the second rather than by resolution, so the length control replaces the resolution one, and the price on the button follows.
There are four engines and they are genuinely different tools. Each one has its own prompting guide built in: the same sentence is rewritten in the grammar the chosen model was trained on, because a prompt written for one reads as noise to another.
| Engine | Good for | Lengths | Per clip |
|---|---|---|---|
| MiniMax H3 · 768P | Trying a move. The cheapest video on the canvas | 4–15s | 70c at 4s, 263c at 15s |
| MiniMax H3 · 2K | Finishing the move you kept | 4–15s | 114c at 4s, 427c at 15s |
| Omni Flash | Quick work with no relay needed | 4–10s | 90c at 4s, 220c at 10s |
| LTX Pro · 1080p | Fidelity, with sound | 6, 8, 10s | 197c at 6s, 329c at 10s |
| LTX Fast · 4K | Resolution, when the frame has to be big | 6, 8, 10s | 486c at 6s, 810c at 10s |
| FLUX 3 · HD | Long takes and multi-shot films | 5–20s | 186c at 5s, 745c at 20s |
| FLUX 3 · FHD | The finished piece | 5–20s | 318c at 5s, 1270c at 20s |
A twenty-second FHD clip is the single dearest act on the canvas — roughly what thirty-eight 4K stills cost. Find the move at 768P for 70 credits, then spend the real money once, on the take you kept.
MiniMax, LTX and FLUX all generate sound alongside the picture. The geometry rule carries over unchanged: the building does not change between frames. A camera move is a camera move, not a redesign in twenty-four parts.
Camera moves
Eight presets, named the way a film set names them.
| Move | What the camera does |
|---|---|
| Dolly In / Out | Travels forward or back on a track, about a tenth of the frame. Not a zoom — the parallax is the point |
| Dolly Left / Right | A short level track sideways. It does not pan or rotate |
| Jib Up / Down | Rises or falls on a jib, staying level. The camera travels; the view does not tilt |
| Static | Locked off. Nothing moves but the air |
| Focus Shift | The frame holds; focus racks from the foreground to the subject in one continuous move |
Every preset states its bound — how far, how fast, and what does not happen. A large move asks the model to invent more of the scene, and everything invented is a chance to drift. Where a shot genuinely needs to travel, split it across shots rather than accelerating.
Dense repeating texture — woven screens, ribbed facades, fine louvres — is where a moving camera crawls and swims. The fix is less secondary motion, not more description: hold the light steady, still the planting, and let parallax carry the shot.
A film in one clip
Several shots, cut inside a single generation. One charge, and the cuts land where the prompt says rather than where an editor guesses. Select a render, open Motion, and press A film in one clip.
You pick the length, answer three to five questions about what the still cannot settle, and it writes a multi-shot prompt into a text node you can read and change before spending anything. The usual arc: the reference view, a macro of a chosen detail, then a low-angle beauty shot.
Several pictures become the shots
Select several renders and the selection order becomes the shot order — the first is the opening shot, the last is the closing one. How many can actually be used depends on the engine, and this is what the APIs accept rather than a preference:
| Engine | Pictures | How they are used |
|---|---|---|
| FLUX 3 | Up to 10 | One per shot, each pinned to its own second in the clip |
| MiniMax H3 | Exactly 2 | First frame and last frame; the clip travels between them |
| LTX 2.5 | Up to 2 | An opening frame and a closing one |
Select three or more on MiniMax and it refuses before the length is even chosen, rather than quietly using two and ignoring the rest. Two pictures there is a different mode — the model interpolates a continuous path from the first to the last — so it prefers few or no cuts, and the final shot must land on the second picture. FLUX is the engine that pins every shot to its own picture.
Reels
A reel is the opposite shape: many separate clips, cut together afterwards. From one render, Nomad plans the shots, draws a storyboard, makes the stills, then renders a clip from each.
How long is the reel?
20, 30, 40 or 60 seconds.
How long is each cut?
5, 8 or 10 seconds — with the shot count and the total cost shown for each. Sixty seconds of five-second cuts is twelve shots; of ten-second cuts, six. Same length, a different film.
One question per shot
Twelve shots means twelve questions, each labelled for its shot and each about what the still cannot settle — time of day, weather, whether people are there, which detail the close shot lands on. Never about the architecture. Your answer reaches that shot's prompt and no other.
The storyboard
Every shot becomes an image node with its prompt already written, wired to your reference. Read them, change any of them, then render.
The clips
Each still becomes its own video node, at the cut length you chose. Nothing is spent on video until you have seen the frames.
Answering them is halfway through a job, so clicking the canvas behind will not close them and throw the answers away. Only Close, Skip and the go button end it. With a dozen questions the list scrolls and the buttons stay pinned to the bottom.
Aspect ratio
MiniMax takes its aspect ratio from the picture and ignores the one the panel asks for. A 16:9 render sent as a 9:16 request does not come back cropped — it comes back squeezed, and a subtly wrong-proportioned building is the kind of mistake that survives all the way to a client.
So the presets follow the picture: a portrait render makes a portrait storyboard and portrait clips. And if a MiniMax video node does not match the render wired into it, it stops before spending, names both shapes, and offers to set the node to the render's. FLUX and LTX genuinely reshape to a ratio you ask for, so they are left alone.
Choosing a model
| Model | Good for | Notes |
|---|---|---|
| Nano Banana 2 | Everyday work, iteration | Fast, all ratios, 0.5K to 4K |
| Nano Banana Pro | The hero shot | 1K to 4K |
| GPT Image 2 | When its look suits the project | Bills by quality and shape — a wide frame costs less than a square one |
| Omni Flash | Video, from any input | Bills by the second, not by pixels. No relay needed |
| MiniMax H3 | Video. 768P to try, 2K to finish | Takes its shape from the picture. Two frames at most |
| LTX 2.5 | Video with sound, up to 4K | Pro is 1080p, Fast reaches 4K. Six to ten seconds |
| FLUX 3 | Long takes and multi-shot films | Five to twenty seconds, up to ten pinned keyframes |
Set the model per node, not per project. A set can mix them — the storyboard planner will do exactly that if it helps.
Each video engine carries its own prompting guide, so the model you pick changes how your sentence is written, not just where it is sent. MiniMax wants named fields and cut timestamps; LTX and FLUX want flowing prose; FLUX past eight seconds wants explicit time phases. You never see any of this — you pick an engine and the writing follows.
Credits and cost
One credit is one US cent. Images are charged by resolution; writing a prompt or asking Nomad anything is 0.5 credits, which is roughly what it costs to serve. The pre-flight check is free and a failed call refunds itself.
Every action that spends shows the number on the button first. Run all totals a whole selection before you commit to it.
Video is charged by the second and by engine, from 70 credits for a four-second 768P clip to 1,270 for twenty seconds at FLUX FHD. Because the range is that wide, the length, the engine and the price all sit on the node before you press play.
The render queue
Press play while something is rendering and it joins the queue rather than being ignored. One at a time, in the order you pressed, with what is waiting named on screen. Click any waiting job to drop it, or press Stop to cancel the line — the one already running finishes rather than being abandoned.
A video render is a long job — a minute or two at 768P, considerably more for 2K or twenty seconds of FHD — so it narrates itself while it works. The node caption ticks through queued, running and the minutes elapsed, then saving the clip. A silent wait reads as a hang, and a hang is what gets a healthy job abandoned.
Open Inspect after any render and it lists the run step by step: what was submitted, how long the polling took, when the clip was ready, whether it was saved and how big it was. If something goes wrong, the step it stopped at is the answer.
The relay
Nomad tries every provider directly first. Some video APIs do not allow requests from a browser at all — that is a rule Chrome enforces, not a choice — and MiniMax's is one of them. When the browser refuses, Nomad falls back to a small relay you run on your own machine.
Run it
Open a terminal and run node nomads-relay.js. Leave the window
open; closing it stops the relay.
Tell Nomad about it
Keys → the relay field → http://localhost:8787. Remembered, so it
is a one-time step.
Watch it work
Every request that passes through is printed, so you can see exactly what left your machine and where it went.
One route — /go?url=… — carrying the whole URL. It never parses the
request and rebuilds it, because a relay that rebuilds a request from its own idea
of the schema silently pins every value it does not know about. It also refuses to
forward anywhere except the providers, since it runs on your machine with your keys
passing through it.
Without a relay, Omni Flash still works, and everything to do with stills works normally.
Keyboard
| Keys | Does |
|---|---|
| Ctrl + Z | Undo — a whole storyboard counts as one step |
| Ctrl + Shift + Z | Redo |
| Ctrl + C / V / D | Copy, paste, duplicate |
| Ctrl + G | Group the selection |
| Ctrl + Shift + G | Ungroup |
| Ctrl + A | Select everything |
| Ctrl + Shift + S | Download the selection, one at a time |
| Ctrl + 0 | Fit the canvas to the screen |
| Delete | Remove the selection |
| Escape | Deselect, or leave presentation mode |
Pencil and touch
The canvas reads a pencil the same way it reads a mouse, so nothing is a cut-down version. There is no tablet mode to switch into.
| Gesture | Does |
|---|---|
| Tap | Select |
| Double-tap | Opens the node picker, as a double-click does |
| Long-press | The context menu — copy, paste, lock, arrange, info |
| Drag | Move a node, or pan the canvas from empty space |
| Pinch | Zoom, around the point between your fingers |
| Two fingers | Move the canvas |
| Pencil drag from a port | Wire two nodes together |
Ports and transform handles roughly double in size and stay visible, because a fingertip is about nine millimetres across and there is no hover to reveal anything. On a pencil or a mouse they stay small and appear as you approach.
A long press cancels whatever drag it interrupted, so the node does not slide away under the menu. Resting your finger still counts as a press — it allows about twelve pixels of wander before it decides you meant to drag.
When something goes wrong
The render changed my building
Tell Nomad what changed, in plain words. It rewrites the prompt against the failed image, naming that specific failure. If it keeps happening, check that your reference is actually wired into the image node — the pre-flight check catches this, but it can be overridden.
Nothing happens when I press play
Something else is probably rendering. Look at the queue bar above the dock — your node will be in the line with its name on it.
The set does not look like one building
Usually the reference reached some shots and not others. Select the reference and check its wires fan out to every image node in the set.
I ran out of credits mid-set
Nothing is lost. The queue stops, everything already made stays, and the prompts are still written. Top up and press play again.
The clip came back the wrong length or the wrong size
Nomad compares what arrived against what it asked for and says so when they differ.
If you are on a relay, check it is the current nomads-relay.js — a relay
that rebuilds requests instead of forwarding them will pin every clip to whatever its
own defaults are.
The video finished but nothing appeared
Open Inspect. The run is listed step by step and the step it stopped at is the answer. The most common cause is a relay answering for itself rather than forwarding, which Nomad now names outright instead of waiting.
The download is a few hundred bytes
That is a refusal page wearing the right filename, and it usually means the link expired. Providers serve finished clips on signed links that last under an hour, so Nomad downloads the bytes into the project as soon as a clip is ready. If it could not — the file was too large, or the host refused — it says so at the time and warns that the link expires. Re-run the clip and it will be saved properly.
My building looks squeezed
MiniMax takes its shape from the picture, so a node set to a different ratio squeezes rather than crops. Nomad stops before spending when it spots this and offers to match the node to the render.
Am I running the file I just saved?
Open Keys — the build is stamped at the top and printed to the browser console. If it is not the one you expect, the browser is serving a cached copy: Ctrl+Shift+R.