Case study

Reimagining ArchViz Through AI

From legacy renderings to modern AI-enhanced visualization — a ComfyUI case study with Z-Image Turbo and Flux 2 Klein 9B.

By Roshan Abraham · ComfyUI · Z-Image Turbo + Flux 2 Klein 9B

Reimagining ArchViz Through AI — cover
Reimagining ArchViz Through AI — ComfyUI case study with Z-Image Turbo + Flux 2 Klein 9B.

01 · Overview

Project overview

A visual case study following a single archviz asset through a modern AI pipeline: from a decade-old legacy render, to a ChatGPT-redefined base image, through two ComfyUI 2× upscale branches, and finally to a mobile-friendly presentation output.

Legacy Render

Original image created over a decade ago

ChatGPT Redefined

Modern mood, lighting and material direction

ComfyUI Upscale

Two AI pipelines for 2× enhancement

Presentation Output

Mobile-friendly visual proof

02 · Baseline

Original render showcase

The legacy workflow has strong design intent but limited modern realism — flat lighting, simplified plants, softer edges and dated material response.

Original legacy render with facade, vegetation and pavement detail
Baseline reveals flat lighting, simplified plants, softer edges and dated material response.

03 · Direction

ChatGPT redefined version

Before any upscale stage, ChatGPT reinterprets lighting, material and atmosphere. The facade gains richer material response, the vegetation feels grounded, and the whole scene becomes cinematic.

ChatGPT-redefined base image showing richer material and mood
Lighting reinterpretation · material refinement · atmosphere · stronger realism.

04 · Pipeline A

Z-Image Turbo result analysis

Tile-based 2× upscale with z-image-turbo_fp8_scaled_e5m2_KJ + qwen_3_4b + ae_zimage at 10 steps, denoise 0.22. Sharper details with strong preservation of architecture and mood.

Z-Image Turbo 2× upscale results
Texture sharpness · edge recovery · vegetation realism · material clarity.

05 · Pipeline B

Flux 2 Klein 9B result analysis

Same input, alternate refinement: flux-2-klein-9b-fp8 + qwen_3_8b_fp8mixed + flux2-vae at 20 steps, denoise 0.22. The result trades a little raw sharpness for cinematic coherence and a more unified global lighting feel.

Flux 2 Klein 9B 2× upscale results
Lighting realism · micro texture · cinematic mood · upscale consistency.

06 · Comparison

Direct comparison grid

Same visual region. Four stages. Clear evolution.

Comparison grid: Original, ChatGPT Redefined, Z-Image Turbo 2×, Flux 2 Klein 9B 2×
Original render → ChatGPT redefined → Z-Image Turbo 2× → Flux 2 Klein 9B 2×.

07 · Pipeline

Workflow pipeline

From an old render to a publishable AI-enhanced archviz output.

Workflow pipeline: legacy render → ChatGPT redefinition → Z-Image Turbo → Flux 2 Klein → final output
Designed for repeatable production, not one-off experiments.

08 · Takeaways

Technical insights

The practical value is workflow control.

Modernize legacy assets

Turn older renders into current visual standards.

Preserve architecture

Prompts explicitly protect layout, facade and materials.

Tile-based detail

2× upscale uses 768 px tiles with seam fixing.

Production advantage

More usable outputs from existing render libraries.

Visual storytelling

Atmosphere and material realism become presentation-ready.

Fast iteration

Two ComfyUI branches for comparison and tuning.

Technical insights summary page
Six takeaways from blending creativity, experience, and AI.

The future of architectural visualization

Blending creativity, experience, and AI. Download the full deck or jump into one of the production workflows.

© Roshan Abraham · ArchViz Magic Canvas