Turning 12 Chibi Monsters with AniSora

Last time we made a roster of 12 chibi monsters in Scenario. Before any of them can walk, Sprute needs each one from eight directions. For humanoid characters, Sprute gets those views from a turntable: AniSora (a Wan 2.2 image-to-video finetune) spins the character once, and we keep every tenth frame.

Would the same turntable work on monsters that have wings, horns, shells and no clothes at all?

Setup

• Command: sprute turntable character. It runs AniSora V3.2 (Wan 2.2 I2V, fp8) and makes 81 frames.
• Output: Frames 0, 10, …, 70 become the S, SE, E, NE, N, NW, W, SW cells.
• Speed: About 30 seconds per monster on one RTX PRO 6000.
• Inputs: 12 monsters in two versions each:
  • The three-quarter (SE-facing) chibis from the last post.
  • Front-facing versions from a new prompt we call "sprute". It asks for a small, muted, front-facing sprite, so the art matches the pixel density of the Sprute characters.
• Prompts: Two, each run on all 24 inputs.
  • base: The prompt Sprute already uses for humanoids. It asks the model to "preserve the character's reference proportions, clothes and hairstyle".
  • creature: The same structure, rewritten for monsters. It asks the model to "preserve the creature's proportions, colors, horns, wings, tail and markings".

That makes 48 turntables, about 25 minutes of GPU time.

Results

All 12 monsters turning, front-facing input, base prompt:

The eight directions Sprute keeps from each turntable:

Most of these are usable as they are. The back views look like the backs of the same monster. The frog, mushroom, slime, imp, wolf, bat, skeleton and goblin hold up all the way around.

Three are weaker:

• Fire spirit: Its back washes out into a pale cream lump.
• Beetle: From the side (E and W), its two horns merge into one.
• Dragon: Its back views (NW and W) turn into a mass of fur.

Naming body parts makes them appear

We expected the creature prompt to help. It made things worse:

When the prompt says "horns, wings, tail", the model draws horns, wings and a tail on the monster, whether or not the reference has them. The frog grew a crocodile tail and a back plate. The golem grew horns and a little yellow tail. The slime grew bat wings. This happened on both input sets, to the beetle, goblin, skeleton and fire spirit as well. The imp did something different: with the creature prompt it never turned at all.

The humanoid prompt names no body parts, so it adds none. We're keeping the base prompt for monsters.

Start from the front

The turntable assumes the first frame faces the camera, because that frame becomes the S cell. The three-quarter inputs still turn cleanly, but every direction then sits about half a step off. The S cell shows a three-quarter view, and the views after it are rotated the same amount:

With front-facing inputs the S cell is a true front view, and the other seven directions line up with the ones Sprute's humanoid characters use.

Next steps

• Sort the monsters by body type and animate them:
  • Bipeds (imp, goblin, skeleton, mushroom) driven by Template-kun.
  • Four-legged monsters driven by a quadruped rig.
  • Blobs with procedural squash and stretch.
  • Flying monsters later.
• Redo the fire spirit and the dragon with new seeds.

‒ Sprited Dev 🐛