Architectural rendering agent

Massing screenshot → photorealistic render.

Turn a rough massing screenshot into a photorealistic architectural render — automatically, faithfully, and on your own ChatGPT or Gemini quota. No API key, nothing uploaded to a third party.

Free & open source (MIT) · runs locally on your machine · bring your own ChatGPT login

Faithful — your geometry is locked Quota-saving — incremental edits, not re-rolls Tamper-checked every round 2K / 4K / 8K local upscale

Why it's different

A director that understands the brief, writes a professional prompt, generates, checks the result against your base image, and refines — while you stay in control.

🎯

Faithful by design

Every round is compared to your original. Building form, floor count, windows, straight lines and text are locked — the model improves materials, light and entourage, not your geometry.

💸

Quota-saving iteration

Local flaws are fixed by an incremental edit on the previous render instead of re-rolling a whole image. Defects converge instead of reappearing.

🧭

Confirm gate

The AI first explains what it understood and shows an editable prompt. Approve or adjust — then it starts spending credits.

🖌️

Hands-on region edits

Mark exactly what to change with freehand, line, rectangle, lasso, smear or eraser — the AI edits only inside the mark.

🎨

Two engines, your account

Generate with ChatGPT or Gemini's nano-banana — both drive a browser you're logged into, on your own subscription. No API key.

🧱

Curated ArchViz library

Precise material vocabulary (board-formed concrete, Shou Sugi Ban, Corten, zinc standing seam…), lighting scenarios and camera guidance.

🔍

Offline vision, no account

The Prompt Assistant can read your image with a local Ollama vision model — fully offline, no account, no VPN.

🖼️

Local quality boost

Final deliverables upscale to 2K/4K/8K with a local super-resolution model and de-watermark locally — offline, on the final image only.

🌐

Bilingual UI (中 / EN)

Work in Chinese; the prompt sent to the model is always English (image models are more reliable in English). Your choice is remembered.

How it works

One director chat runs the whole loop; you review at every checkpoint.

  1. Describe & upload. A few sentences plus a base image (kept faithful) and optional reference images for mood, material and light.
  2. Confirm the prompt. The AI paraphrases your intent and shows an editable, professional prompt before any credits are spent.
  3. Auto-iterate. Generate → tamper/quality check against the base → refine the previous render, or redraw from base if the form drifts.
  4. Review & export. Pause every N images to leave feedback, mark a region to fix, or export to your Desktop. Every step is saved.

Download & run — no coding

Grab the package for your OS from the Releases page, unzip, and open the included read-me. First launch sets everything up and asks you to log into ChatGPT once.

Windows*-windows-*.zip · double-click the launcher
macOS*-mac-*.zip · double-click the launcher
Linux*-linux-*.zip · see the read-me

⬇ Get the latest release

Prerequisites: Python 3.10+, Google Chrome, and a ChatGPT account that can generate images.