# LookX AI vs Veras: pavilion rendering test

**Status: test kit available; paired renders and scores are not available.** The picture is the original input. It was drawn for 4now from simple geometry, without using a client project, stock photography, or vendor output. It is not a sample produced by either tool. You may use it to run and share this comparison.

## Input and prompt

Use the unmodified 4now pavilion PNG (1024 × 768, SHA-256 a705e26c3aa9390b8ec95d819d0db953cbf07968d1724a0e9c16de9a74e6221c) in both products. The SVG is the vector source. The drawing has three equally sized front windows, a single side door, a flat main roof and a separate lower canopy on two posts.

Paste this same prompt into each product:

> Render this single-storey pavilion in soft, overcast daylight. Use pale timber cladding, clear glass, a charcoal flat roof and a plain concrete plinth. Keep the camera angle, three front windows, side door and two canopy posts exactly as drawn. Do not add people, planting, extra openings or new floors.

## Procedure

1. In each product choose an image-to-image architectural workflow with the PNG as structural input. Use its web version where available. Otherwise record the host application and version and identify the difference rather than implying identical hosts.
2. Use the same uncropped 4:3 input and requested output size, 1024 × 768, if available in both. Use their default structure-preservation settings on the first pass. Record the product model, plan, date, actual output dimensions and all non-default settings. Do not assume the products' strength scales correspond.
3. Keep the first output, without selecting the best image from a batch. If the product generates several images together, use the first in display order and record the batch size. Record elapsed generation time separately from editing time.
4. Check four yes/no conditions: exactly three front windows, one side door, exactly two canopy posts, and the flat main roof with a separate, lower side canopy. Mark a condition as unclear if it cannot be judged. Keep the actual outputs alongside the raw observations.
5. For a second pass, identify the first failed condition using the order above. Request only that correction, up to two attempts or 5 minutes of active editing, whichever comes first. Count active editing time apart from generation waits. Record unresolved defects; never report an unrun or unavailable result as zero.
6. This is a one-scene pilot. It cannot establish which product is generally better. Before publishing output images, check the applicable plan's publication rights and label the vendor, version, settings and date.

## Official references

- LookX pricing: https://www.lookx.ai/price.html
- Veras product: https://www.chaos.com/veras
