One Still-Life, Three Native Generators
The same product brief tested across Grok Imagine, Codex image_gen, and Antigravity generate_image. A direct evaluation of native visual models without external wrapper degradation.
Compare native AI image generators on real product still-life briefs and learn optimal visual routing.
For a quiet product still-life, Grok Imagine, Codex image_gen, and Antigravity generate_image all returned a publishable frame on this brief. Prefer the native path you already have. Do not route FrankX stills through FAL. Trust the tool-result model name.
AI CoE pillar: Technology · media routing
- Product still-life: Any of the three native paths on this page
One Still-Life, Three Native Generators
TL;DR. Same prompt. Three native tools that actually returned pixels: Hermes Grok Imagine, Codex image_gen, Antigravity generate_image. All three passed QA. Earlier interior and droid frames missed the brief and are not shown. FAL was not used. Grok video was not run in this session. The configured Grok quality id still did not appear in the tool result.
Receipt: /research/image-curated-2026-08-16/receipt.json
The brief
A single matte-black precision cylinder on dark honed slate, one hairline gold ring at the equator, warm sidelight from the left, soft falloff, shallow focus, quiet product photograph, 16:9. No logos, no type, no people.
What ran
| Path | Tool result | File | QA |
|---|---|---|---|
| Hermes / xAI | grok-imagine-image (not the configured quality id) | JPEG 1280×720 | Publish |
| Codex CLI | image_gen.imagegen | JPEG 1672×941 | Publish |
Antigravity (agy) | generate_image | JPEG 1376×768 | Publish |
Codex’s -i flag attaches an image to a prompt. Generation used its native image_gen tool. Antigravity used generate_image. Neither is FAL.
The frames
Grok Imagine: one cylinder, gold ring, gold end, slate, no type.
Codex: standing cylinder / jar, gold ring, empty ground, no type.
Antigravity: lying cylinder, gold ring, slate, no type.
More Grok Imagine, same session
Only QA-pass frames. No Arcanea lore names on this site.
Photoreal night street. No people. No readable signs.
Quiet interior. Empty. Gold slit of light.
Original anime study. No logos. No named studio.
Collectible droid study. Soft mascot, not a mechanical kit.
The Grok 4.6 routing facts sit in a code-built panel on the research page so numbers cannot smear. That is not an image model.
What we are not showing
An earlier pass asked Grok Imagine for an empty room and a mechanical droid. The room gained a chair and a lamp that looked on. The droid became a mascot. Those frames are not audience results.
Hermes is configured for grok-imagine-image-quality. The tool still reported grok-imagine-image. There is no quality-vs-fast pair on this page.
Grok Imagine video is a separate backend. This session has no video tool. Not run.
Analysis
On this one still-life, the three native paths are close enough that the routing question is “which CLI is already in the session,” not a fake leaderboard. Codex stood the object up. Grok and Antigravity kept it on its side. All three kept the gold ring and stayed empty of type.
FAQ
Is this Grok 4.6 making pictures?
No. Grok 4.6 is the reasoning model. These stills came from Grok Imagine, Codex image_gen, and Antigravity generate_image.
Why isn’t FAL here?
It is not a native estate path for this work. We did not use it.
Where is Model Arena?
Model Arena is receipt-gated text/agent battles. This is a media sample.
Related: Grok 4.6 brief · Research hub
Read on FrankX.AI — AI Architecture, Music & Creator Intelligence
Stay in the intelligence loop
Field notes on AI systems, production patterns, and builder strategy.
Continue Reading
Evals Are the Experiment
An agent system without a measurement protocol is literature. How to apply the scientific method to agentic loops: hypothesis, instrument, run, ratify, ledger.
Read article
Multi-Agent Model Fabric 2026: Routing, Evals, and Fallbacks
Production architecture for absorbing Grok 4.6, Gemini 3.7 Flash, DeepSeek-V4-Pro, and Qwen3.8-27B. Lane-based routing, structural evals, cost caps, and handoff contracts.
Read article
How to Govern an AI Skill Library
A practical operating model for skill registries, risk tiers, evaluations, and AI CoE ownership.
Read article