Sloth Studio Sprite3D: 8-direction sprites from one image (Claude Code skill)

Give Claude Code one picture of your character. Sloth Studio Sprite3D turns it into a textured, rigged 3D model and renders game-ready sprites from it:

• 3 moves: walk (8 frames), run (8 frames) and punch (a one-two: jab then cross, 12 frames)
• 8 directions: S, SE, E, NE, N, NW, W, SW
• 256 px transparent PNG strips, one per move and direction (24 strips)
• 96 px sprite sheets, one per move (one row per direction), plus preview GIFs
• The rigged, textured GLB and a browser viewer that plays the moves and re-exports sprites

Because every frame is rendered from the same 3D model, the character keeps the same face, clothes and proportions in all directions and frames. The feet stay planted on the floor in walk and punch (leg IK).

How it works: your character is redrawn as a 4-view T-pose sheet (1 image call), Hunyuan3D-2 turns it into a 3D mesh on a free Hugging Face Space, the drawings are projected onto the mesh and the hidden parts are filled by about 10 more image calls, a 17-bone skeleton is fitted and skinned in Blender, and the moves are rendered headless in Chrome. Every step prints pass / warn / FAIL checks (feet on the floor, texture coverage, silhouette match, fist clearance), so Claude can tell when to redo a step.

Made with the same tech behind Sloth Studio: https://slothstudio.supercapybara.com/

What you need

• Claude Code, recommended model Opus 5.5 or above (the scripts also run by hand)
• uv, Blender 4.2 or newer, Node.js 22 or newer, Chrome or Chromium; ffmpeg optional
• An image provider: the codex CLI logged in with a ChatGPT plan that includes image generation, or an OpenAI or OpenRouter API key
• A free Hugging Face account token (recommended)
• If a step fails on your machine, the skill's built-in How to fix me guide tells Claude how to adapt it to your OS and tools and re-run only that step.

Cost and time per character (measured)

On a new test character, from a cold start following the skill's instructions: about 16 minutes, 12 image calls (14 attempts including automatic retries), 1 free Hugging Face shape call.

• With the codex CLI and a ChatGPT plan: no per-image charge; the calls count against your plan's limits.
• With the OpenAI or OpenRouter API: about 11 to 14 calls, roughly $0.30 to $0.45 per character at about $0.03 per call. Check your provider's current price.
• The free Hugging Face GPU quota runs out after a few characters and refills within about 24 hours.

Works best with

Humanoid characters with visible arms and legs, fitted clothes, short or tied hair. Not suited: capes, cloaks, long robes, wide skirts, wings, tails, big weapons or props, very bulky armour, quadrupeds. The move set is fixed: walk, run, punch.

Licence and territory: read before buying

• EU, UK and South Korea: the shape step uses Tencent Hunyuan3D-2, whose licence does not apply in the European Union, the United Kingdom or South Korea, and forbids using its output there. If you are there, you need your own T-pose mesh from another tool (shape --mesh your.glb), or this skill is not for you.
• Your sprites are yours to use, including commercially, within the Hunyuan3D terms (territory; no using outputs to train other AI models), your image provider's terms, and your rights to the input character. Tencent claims no rights in the outputs.
• Hunyuan3D large-company clause: companies whose products had more than 1 million monthly active users must request a licence from Tencent.
• The skill itself is GPL-3.0-or-later (its Blender scripts must be GPL-compliant). You may use, change and share it under the GPL. three.js is bundled under MIT. Full details in LICENSES.md in the download.
• The free Hugging Face Space is run by Tencent and can change or go offline; the skill then needs another shape source (--mesh).