ScenarioExplorations Blender 2D Animation

Three looks. One Blender script. 30 Scenario generations

Blender 2D
Animation.

Generate the art in Scenario.
Animate it once in Blender.

Every picture and the hero's run come from Scenario. Blender lays them out as flat , moves the camera and runs the physics.
Switch the look, or pull the layers apart.

0 / 9 0.0 s

Made once, dressed three times.

The animation lives in Blender; the look lives in the pictures. Scenario makes the pictures, so a new theme is a new set of generations, not a new animation.

Flat cards, real depth

Sky, far layer, mid layer, ground and foreground are separate images with a , each on its own plane. An orthographic camera follows the hero; each layer's horizontal position is driven by the camera.

Animated once

Jumps with squash and stretch and the coins are plain keyframes. The pots, idol and coconuts are whose collision shapes are traced from each sprite's outline.

Swap the art

The Halloween and snow looks run through the same script: build_scene.py --set assets_snow. Only the images change, plus a few numbers measured from them (horizon, ground line, hero size).

The ingredients, jungle run.

Ten Scenario generations per look, through the Scenario MCP. The concept frame is the reference for every other image; the hero is drawn on , and its left pose is the of the run clip. Open a card for its prompt.

Ingredient

Ingredient

Prompt

How it was made

One recipe, three looks, 30 generations.

Made with Claude Code, the Scenario MCP and Blender 5.2. The brief: a side-scrolling platform shot with "two or three backgrounds with the parallax effect and… one or two foregrounds", a main character with a run and a jump, objects falling, and "all should be built in Blender".

  1. 1

    Lock the look with one frame

    An in-game screenshot of the whole idea: sky, sea, jungle ruins, ground, hero, coins. Two takes, one kept. It is the reference image for every later asset, so the set stays consistent.

    With Scenario MCP · GPT Image 2.5 Sunburst · 1536×864 · 23 CU

  2. 2

    Generate each layer on its own

    Sky (opaque), far sea, mid jungle, a ground strip with a flat walkable top, and a foreground sheet, each on a transparent background. A script makes the tiled layers repeat without a seam.

    With GPT Image 2.5 Sunburst · concept as reference · 13 to 15 CU each

  3. 3

    Props on one sheet

    Eight props with wide gaps, cut apart by connected components. The palm is drawn without coconuts, which become separate physics bodies.

    With GPT Image 2.5 Sunburst · 14 CU (props) · 13 CU (palm)

  4. 4

    The hero and its run

    Two poses on flat magenta. The left pose is the first frame of a 5-second clip that runs "like a treadmill"; a script keys the magenta and keeps the frames that loop most cleanly (9 for the panda, 12 for the cat and the snow leopard).

    With GPT Image 2.5 Sunburst · 24 CU Then Minimax H3 · image to video · 768p · 80 CU

  5. 5

    Lay it out in Blender

    The Scenario MCP makes every asset, and one Python script, run by Claude in Blender, builds the whole scene: it puts every image on a flat plane (emission with alpha), stacks the planes in depth, adds an orthographic camera that follows the hero, and drives each layer's position from the camera.

    With Blender 5.2 · build_scene.py

  6. 6

    Animate and simulate

    The run plays as an image sequence; the jump swaps to the jump pose with squash and stretch. Coins flip up and out of frame. Pots, idol, amphora and coconuts are rigid bodies with convex hulls traced from their outlines, kept in one plane by invisible walls.

    With Blender keyframes · rigid bodies · baked physics cache

  7. 7

    Swap the theme

    A Halloween street with a black cat, then a white winter valley with a snow leopard, made with the same recipe: a new concept frame, then the same nine asset roles with it as reference. The same script builds each one with --set assets_halloween or --set assets_snow and a small config of measured numbers.

    With 10 generations · 215 CU per look

  8. 8

    Render

    EEVEE, 1920×1080, 24 fps, 184 frames for each look. For the film, the scene is also rendered from a moving camera to show the planes in depth.

    With Blender 5.2 · EEVEE

What did not work

  • The tiling blend faded the big jungle arch. The seam now goes where the silhouette is lowest.
  • The palm floated: its image has empty rows under the trunk. It is now offset by its measured bounds and set just behind the grass lip.
  • The first physics runs flung pieces off the pedestal. Outline-traced hulls, lane walls and a hero body that slides in sideways fixed it.
  • The cat's run plane came out with no size: Blender reports 0×0 for an image sequence until it plays. The script now measures the first frame.
  • The snow leopard's run clip drifted from magenta to green. Each frame is keyed on its own background colour, and the loop is taken from the green part, away from the red scarf.

Tip

Ask for each layer alone, on a transparent background, with one concept frame as the reference. Keep the animation in Blender, and a new look is only new pictures.

Credits

656 CU

30 finished Scenario generations for the three looks, 2 models (GPT Image 2.5 Sunburst, Minimax H3), from the job records. The film's narration and music are not counted.

Agent skill · ZIP

scenario-blender-2d-animation

Download the skill

PDF · Illustrated

The making-of

Read the making-of

Scenario

Make yours

Start creating