Rendering a single interior image at 8K resolution (7680×4320) takes four to eight times longer than rendering at 4K (3840×2160). For a scene that takes 20 minutes at 4K, you are looking at 80–160 minutes at 8K — and that multiplier applies to every camera angle, every revision, every test render. When clients request large-format print deliverables (billboard graphics, exhibition panels, showroom prints at 300 DPI), the render time cost of native high-resolution output can consume an entire production day.
AI upscaling offers a compelling alternative: render at 4K (or even 2K), then upscale to 8K or higher using trained neural networks that reconstruct detail beyond the original pixel resolution. The promise is real — the best upscaling tools produce genuinely impressive results on certain content types. But the technology has critical failure modes that are specifically triggered by architectural visualization content, and using AI upscaling without understanding these failure modes will produce deliverables that look artificial, over-processed, or structurally inaccurate at print inspection distance.
Testing Methodology
We rendered five standard ArchViz scene types at native 8K resolution and at 2K resolution, then upscaled the 2K renders to 8K using three tools. We compared the upscaled results against the native 8K ground truth at 100% zoom across five evaluation criteria.
Tools Tested
- Topaz Gigapixel AI 7.2 — Desktop application, $99 perpetual license, GPU-accelerated
- Real-ESRGAN (4x-ultrasharp model) — Open-source, command-line, free
- V-Ray Frame Buffer built-in denoiser + Lanczos upscale — Traditional non-AI baseline
Evaluation Criteria (scored 1–10)
- Edge Sharpness: Architectural edges, window frames, furniture silhouettes
- Texture Fidelity: Wood grain, stone veining, fabric weave preserved vs hallucinated
- Reflection Accuracy: Mirror, glass, polished metal reflections intact vs smeared
- Text/Signage: Any readable text in the scene (book spines, appliance labels)
- Artifact Level: Halos, ringing, pattern repetition, plastic look (lower = more artifacts)
Benchmark Results
Scene 1: Modern Kitchen Interior (heavy reflections, glass, steel)Criterion | Topaz Gigapixel | Real-ESRGAN | V-Ray Lanczos
-------------------|-----------------|-------------|---------------
Edge Sharpness | 8.5 | 7.5 | 5.0
Texture Fidelity | 7.0 | 6.5 | 4.5
Reflection Accuracy| 5.5 | 5.0 | 6.0
Text/Signage | 6.0 | 4.5 | 3.0
Artifact Level | 7.0 | 6.0 | 9.0
AVERAGE | 6.8 | 5.9 | 5.5Scene 2: Living Room (fabric textures, wood, diffuse surfaces)Criterion | Topaz Gigapixel | Real-ESRGAN | V-Ray Lanczos
-------------------|-----------------|-------------|---------------
Edge Sharpness | 9.0 | 8.5 | 5.0
Texture Fidelity | 9.0 | 8.0 | 4.5
Reflection Accuracy| 7.5 | 7.0 | 6.5
Text/Signage | 7.0 | 5.5 | 3.0
Artifact Level | 8.5 | 7.0 | 9.5
AVERAGE | 8.2 | 7.2 | 5.7Scene 3: Bathroom (tiles, wet surfaces, chrome fixtures)Criterion | Topaz Gigapixel | Real-ESRGAN | V-Ray Lanczos
-------------------|-----------------|-------------|---------------
Edge Sharpness | 8.0 | 7.0 | 5.0
Texture Fidelity | 6.0 | 5.5 | 4.5
Reflection Accuracy| 4.0 | 4.0 | 6.0
Text/Signage | — | — | —
Artifact Level | 5.5 | 5.0 | 9.0
AVERAGE | 5.9 | 5.4 | 6.1Scene 4: Exterior Residential (vegetation, sky, concrete)Criterion | Topaz Gigapixel | Real-ESRGAN | V-Ray Lanczos
-------------------|-----------------|-------------|---------------
Edge Sharpness | 9.0 | 8.5 | 5.0
Texture Fidelity | 8.5 | 7.5 | 4.5
Reflection Accuracy| 7.0 | 6.5 | 6.0
Text/Signage | — | — | —
Artifact Level | 8.0 | 7.0 | 9.5
AVERAGE | 8.1 | 7.4 | 6.3Scene 5: Office Interior (glass partitions, screens, repetitive grid ceiling)Criterion | Topaz Gigapixel | Real-ESRGAN | V-Ray Lanczos
-------------------|-----------------|-------------|---------------
Edge Sharpness | 7.5 | 6.5 | 5.0
Texture Fidelity | 5.0 | 4.5 | 4.5
Reflection Accuracy| 4.5 | 4.0 | 6.0
Text/Signage | 3.0 | 2.5 | 2.5
Artifact Level | 4.0 | 3.5 | 9.0
AVERAGE | 4.8 | 4.2 | 5.4Key Findings
Where AI Upscaling Excels
- Diffuse organic textures: Fabric weaves, wood grain, vegetation foliage, concrete surfaces. AI models are trained heavily on natural texture data and reconstruct these convincingly. Scene 2 (living room) and Scene 4 (exterior) scored highest.
- Architectural edges: Straight lines, window frames, building silhouettes. Both Topaz and Real-ESRGAN sharpen edges better than traditional upscaling, producing crisp architectural boundaries.
- Low-frequency gradients: Smooth walls, sky gradients, soft shadow transitions. No artifacts introduced, clean smooth output.
Where AI Upscaling Fails
- Reflective surfaces: Mirrors, chrome, polished stone, wet tiles. AI models hallucinate detail in reflections, creating phantom textures that do not exist in the reflected scene. Scene 3 (bathroom) scored worst, with Topaz actually performing below the traditional Lanczos baseline for reflections.
- Repetitive geometric patterns: Tile grids, ceiling panel arrays, perforated screens. AI models create slight alignment errors that produce visible moiré patterns at print resolution. Scene 5 (office) ceiling grid was visibly distorted.
- Text and fine detail: Book spines, appliance labels, architectural signage. AI models hallucinate plausible but incorrect characters. Never rely on upscaling to preserve readable text.
- Glass transparency with complex backgrounds: Through-glass views where the AI model must simultaneously upscale the glass surface and the scene visible through it. Depth-ambiguous areas produce inconsistent sharpening.
The Hybrid Workflow: Selective Upscaling
The optimal production workflow is not "always upscale" or "never upscale" — it is selective upscaling that applies AI enhancement only to image regions where it excels and preserves native render quality where it fails.
Python# RenderVault: Selective AI Upscaling with Mask-Based Compositing
# Upscales only non-reflective regions, preserves native render for reflections
import cv2
import numpy as np
from pathlib import Path
def selective_upscale(
native_4k_path: str,
upscaled_8k_path: str,
reflection_mask_path: str,
output_path: str,
feather_radius: int = 15
):
"""
Composite AI-upscaled render with native-resolution render using
a reflection mask to protect reflective surfaces from AI artifacts.
Args:
native_4k_path: Native 4K render (will be Lanczos-upscaled to match 8K)
upscaled_8k_path: AI-upscaled 8K render (from Topaz/Real-ESRGAN)
reflection_mask_path: Render element mask (white = reflective surfaces)
output_path: Output composite path
feather_radius: Mask feather radius for smooth blending
"""
# Load images
native = cv2.imread(native_4k_path)
upscaled = cv2.imread(upscaled_8k_path)
mask = cv2.imread(reflection_mask_path, cv2.IMREAD_GRAYSCALE)
# Upscale native render to match AI output resolution (Lanczos)
target_h, target_w = upscaled.shape[:2]
native_resized = cv2.resize(native, (target_w, target_h),
interpolation=cv2.INTER_LANCZOS4)
# Upscale mask to match
mask_resized = cv2.resize(mask, (target_w, target_h),
interpolation=cv2.INTER_LANCZOS4)
# Feather mask edges for smooth blending
mask_feathered = cv2.GaussianBlur(mask_resized,
(feather_radius*2+1, feather_radius*2+1), 0)
# Normalize mask to 0-1 range
mask_norm = mask_feathered.astype(np.float32) / 255.0
mask_3ch = np.stack([mask_norm] * 3, axis=-1)
# Composite: mask=1 (reflective) → native, mask=0 (diffuse) → AI upscaled
composite = (upscaled.astype(np.float32) * (1 - mask_3ch) +
native_resized.astype(np.float32) * mask_3ch)
cv2.imwrite(output_path, composite.astype(np.uint8))
refl_coverage = np.mean(mask_norm) * 100
print(f"Composite saved: {output_path}")
print(f"Reflection coverage: {refl_coverage:.1f}% of image")
print(f"AI-upscaled: {100 - refl_coverage:.1f}% | Native-preserved: {refl_coverage:.1f}%")
# Usage:
selective_upscale(
"render_4k.png",
"render_8k_topaz.png",
"reflect_filter_4k.png", # From V-Ray/Corona render elements
"final_8k_composite.png"
)The reflection mask comes directly from your render elements — V-Ray's VRayReflectionFilter or Corona's CShading_ReflectFilter. These render elements are already part of a standard ArchViz render setup (see our denoiser article for the element setup), so no additional rendering is required.
Topaz Gigapixel Optimal Settings for ArchViz
Through testing across 40+ architectural renders, these Topaz Gigapixel settings consistently produce the best results for ArchViz content:
- AI Model:
Standard v2(not "High Fidelity" — the HF model over-sharpens architectural edges creating visible halos) - Scale:
2×for 4K→8K,4×for 2K→8K (never exceed 4× — quality degrades significantly beyond this) - Suppress Noise:
Low(your V-Ray/Corona render should already be denoised — additional AI denoising removes legitimate texture detail) - Sharpen:
Medium(Low for scenes with many reflective surfaces, High for exterior vegetation-heavy scenes) - Fix Compression:
Off(your source is an uncompressed render, not a JPEG — this filter adds unnecessary processing) - Output format:
TIFF 16-bit(preserve dynamic range for post-production)
When to Render Native Instead
AI upscaling is not always the right choice. Render at native resolution when:
- The scene is predominantly reflective surfaces (bathrooms, showrooms, glass facades) — upscaling quality drops below acceptable thresholds
- The deliverable contains readable text that must be accurate (signage renders, wayfinding visualization)
- The client specifically requires native resolution (some architecture firms mandate this in their visualization contracts)
- Render time at native resolution is under 30 minutes — the time saved by upscaling is not worth the quality compromise and compositing overhead
Key Takeaways
AI upscaling is a legitimate production tool for ArchViz, saving 60–75% of render time on large-format deliverables — but only when applied selectively. Use it freely on diffuse, organic, and exterior content. Avoid it or composite selectively on reflective, repetitive-pattern, and text-containing content. Topaz Gigapixel produces the best results for architectural content at the Standard v2 model with conservative sharpening. And always deliver the selective composite, not the raw AI upscale — the 15 minutes of compositing time protects you from visible AI artifacts in the client's printed deliverable.
Testing a different upscaling tool for ArchViz? Send us your comparison — we include reader-tested tools in our benchmark updates.