ArchvizSI — architectural visualisation super intelligence. Your entire project on one canvas.
Open the appapp.archviz.siSign in ↗Your entire project. One canvas. Images, video, audio, 3D and text.
Scroll to explore
The product, in four figures.
90sec
Message to renderFrom a plain-language message to the first render on your canvas.
6
Output typesImage, edit, video, audio, 3D and text, all from the same chat.
∞
GenerationsNo credits, no caps, no per-image fees. It runs on local GPUs.
1
CanvasThe whole project, from first mood board to client handover, on one board.
Describe it.The agent doesthe rest.
No pipeline to configure, no tool to select. Write what you want in plain language: the agent reads the intent, routes it to the right model, writes the prompt and places the result on your canvas.
In one message- Ask for four options, get four, in parallel
- Attach a sketch and it reads the geometry first
- Change the season and everything else stays
Four facade variations — stone, glass, timber, white render. Dusk light.
Session 000%
Type a scene. The agent writes the render prompt and places a single image or a whole batch of variations, side by side in a labelled frame.O-01 · Image · PNG · Flux.1
Right-click a card and describe the change: material, light, season. Exactly that changes; the composition stays.O-02 · Edit · PNG · Flux.2 Klein
Turn any still into a short camera move or walkthrough, played inline on the canvas.O-03 · Video · MP4 · Wan 2.2
Describe a mood and a length and get a soundtrack for the walkthrough. No audio software needed.O-04 · Audio · MP3 · ACE-Step 1.5
Lift a single image into a 3D mesh you can orbit and inspect right in the browser.O-05 · 3D · GLB · Hunyuan3D 2.1
Concept statements, material specs and client briefs, written from what is already on the canvas.O-06 · Text · Markdown · Qwen 2.5
What one message can produce
Inside the agent. Each layer is as thick as the seconds it takes.
- 01Message
- 02Intent
- 03Route
- 04Prompt
- 05Render
- 06Canvas
Message
Plain language in the chat panel. No prompt engineering, no workflow to pick.
"4 sunset views, stone / glass / timber / white render"Step 01- Text or attached image
- Scoped to the current board
- ~2 sec
Intent
A local language model reads what you want: output type, subject, constraints and count. Attached images get vision analysis first.
intent → image batch · count: 4 · light: duskStep 02- Qwen 2.5 / Qwen 2.5 VL
- Uses what is already on the canvas
- ~3 sec
Route
The intent picks the pipeline: image, edit, video, audio, 3D or text. You never choose a model.
route → Flux.1 · parallel × 4Step 03- Six ComfyUI pipelines
- Ollama for text
- <1 sec
Prompt
The agent writes a technically correct prompt in the syntax that model expects, one per variation.
"stone facade, dusk, deep reveals, warm interior glow…"Step 04- Per-model prompt syntax
- Variations written together
- ~6 sec
Render
Jobs run on local GPUs, side by side. Nothing is sent to a third-party AI service.
job 1–4 dispatched · rendering 3 / 4Step 05- ComfyUI queue
- Local hardware
- ~70 sec
Canvas
Results land as labelled cards in one frame, ready to edit, animate, turn into 3D or present.
placed · frame: Facade study · 4 cardsStep 06- One frame per request
- Every card can be chained
- ~2 sec
The canvas, part by part.
An infinite board. Pan, zoom and arrange freely; every output lands in a labelled frame you can move, resize and group.
Infinite · zoomable · frames · no export neededThe agent lives in the chat panel. Drag canvas images in as references; every conversation is scoped to its board.
Agent · references · per-board historyThe toolbar gives direct access to every pipeline. Write your exact prompt, set the parameters and bypass the agent entirely.
Direct generate · image · video · audio · 3DSelect a card to edit it, animate it, write a brief from it or chain it into the next step.
Edit · animate · brief · chainCreate mode generates immediately. Ask for variations and they all run in parallel.
Create mode · parallel batchesThink mode plans first: the agent proposes the approach and waits for your go before rendering.
Think mode · plan, then renderAny medium.
One canvas.
Open models, chosen per task and run on local GPUs.
Scroll down and the sheet falls into place.
Aperture 000%
The old way.
The new one.
First render
Set up the scene in a 3D tool, light it, render overnight, iterate on feedback.
Describe the scene in one sentence. First render in about 90 seconds.
Variations for a client
Duplicate the scene, adjust materials, re-render every option separately.
Ask for four. They arrive in parallel, labelled, side by side.
Changing one detail
Re-open the 3D file, modify the element, re-render the whole image.
Right-click the card and describe the change. The rest stays untouched.
A walkthrough video
Animation software, camera rig, render farm, video editing, export.
Select a render, add a motion prompt. The video plays on the canvas.
Tools you need
3D modeller · renderer · video editor · audio tool · document app.
One canvas. One agent. Every output type.
Questions
worth asking
Q-01Do I need to know how to write prompts?+
No. Write what you want in plain language, like "four facade variations, stone and glass, late afternoon sun", and the agent turns it into a technically correct prompt for the right model. Prompt engineering is the agent's job.
Q-02Does it run in the cloud? Who owns the results?+
Flux, ComfyUI and Ollama all run on dedicated local hardware. Nothing is sent to a third-party AI service. Your renders, 3D models, audio and briefs are yours.
Q-03What if I know exactly what I want?+
Use the toolbar. It gives direct access to every pipeline: pick the output type, write your exact prompt and generate without any agent interpretation. The agent is optional.
Q-04Can I use my own reference images?+
Yes. Upload from your machine or drag any card already on the canvas into the chat. The agent analyses the geometry, materials and spatial relationships before it responds.
Q-05How do I get access?+
Sign in at app.archviz.si if you have an account. Otherwise, get in touch and request access.
Your canvas
is waiting
ArchvizSI is designed and built by Marcel Žnidarič, architect and BIM specialist. More at znidaric.si.