+ We ran Scenario's catalogue through PixelKiln: ten text-to-image
+ models priced from 1 to 11 compute units, six single-image tool
+ models, reference-image edits, a three-model chain, and a real
+ isometric pixel-art game's structures, covers, backdrops, and sound
+ effects. Every paid request is preceded by a free quote and held
+ under a ceiling you set. These are our own benchmark images, one
+ sample each.
+
+
+
+
+
+
Price does not predict fit
+
+
+
+ 1 CU FLUX.2 Klein 9b
+
+
+
+ 2 CU Krea 2 Turbo
+
+
+
+ 3 CU Ernie Image Turbo
+
+
+
+ 11 CU GPT Image 2
+
+
+
+ The same brief, seed, and 512 px request to four models. The 1 and
+ 2 CU models gave a usable silhouette and the most authentic pixel
+ texture; GPT Image 2 followed the brief best. None returned a
+ native-grid, closed-palette asset, and four ignored the requested
+ canvas size.
+
+
+
+
+
Render, pixelate, cut out, refine
+
+
+
+ 1 CU render
+
+
+
+ 5 CU Pixelate
+
+
+
+ 2 CU Birefnet cutout
+
+
+
+ free refine, 64×64
+
+
+
+ One manifest, one pixelkiln gen, 8 CU: each step names its
+ parent's style, so the chain runs in order and a 64×64 transparent
+ sprite lands at the end, held at the same review gate as any other
+ asset. On a real game's 240 px landmarks the same chain returned
+ the sprite's exact size at high confidence with far more detail.
+
+ revision.fromStyle lets a style that uses one model
+ revise another style's output, so a chain across models is declared,
+ priced, and ordered instead of stitched by hand.
+
+
+
+
Where it fits
Large landmarks, covers and key art with legible text, flat backdrops, image-to-image edits at 1 CU, and cheap sound effects
+
Where it does not
32 px props, characters, terrain tiles and cliff blocks, which stay on a tool built for them
+
Cost control
A free quote before every paid call, a per-style ceiling, and a command budget
+
+
+
+
+
+ Set up Scenario
+
+
+ Chain models across styles
+
+
+ Read the game pilot ↗
+
+
+ Inspect the model bake-off ↗
+
+