The evaluation layer for image & video models

Every model, judged in the same light.

One prompt, run across the leading image and video models — then weighed frame by frame until the strongest answer is obvious.

Ghost Palette compares these models: Nano Banana 2 Lite, FLUX.2 [pro], SD 3.5 Large, Recraft V3, Seedream 4, Qwen Image, Ideogram V3, Kling 2.6 Pro, Kling 2.5 Turbo, Luma Dream Machine, MiniMax Hailuo 02.

Open Studio

How it works

Three moves from prompt to pick.

Step 01
Prompt once
Write a single brief and choose the image and video models worth testing.
Step 02
Generate together
Every model renders the same prompt in parallel — no re-typing, no switching tabs.
Step 03
Judge in the same light
Weigh detail, motion, and fidelity side by side, then keep the output that wins.

Selected frames

08 stills · FLUX.2 · scroll the strip

01 / 08
02 / 08
03 / 08
04 / 08
05 / 08
06 / 08
07 / 08
08 / 08

Why Ghost Palette

Built to compare, not just generate.

01
Every model, one run
Black Forest Labs, Stability, ByteDance, Alibaba, Recraft, Ideogram, Kling and more — image and video, no tab-switching.
02
Evaluation built in
Benchmark signals for speed, cost, and fidelity sit next to the outputs, so the pick is evidence, not a hunch.
03
Refine from a reference
Upload a direction, generate a refinement, and keep only the frames that move the image closer to finished.

Models compared

7 image · 4 video