Practical Pipeline for Converting 3D Video to Line-Art Animation via MiniMax H3 and DepthAnything
A step-by-step workflow converting 3D steampunk video into line-art animation using GPT-6 Astra, BlenderMCP, ComfyUI DepthAnything, GPT Image 2.5, and MiniMax H
Converting three-dimensional computer-generated animation into traditional two-dimensional hand-drawn or line-art aesthetic has long challenged digital artists, often resulting in lingering synthetic textures and rigid polygonal geometry. AI content developer Kiyoshi Shin (@kiyoshi_shin) has published a practical multi-stage pipeline that starts from a steampunk 3D city scene generated with GPT-6 Astra and BlenderMCP, successfully transforming it into fluid, expressive line-art animation by combining ComfyUI DepthAnything, GPT Image 2.5, and MiniMax H3.
Image source: 新清士@AIコンテンツ開発者 (@kiyoshi_shin) on X
Overcoming 3D Artifacts: The Mandatory Depth Map Preprocessing Stage
Directly submitting raw 3D rendered footage alongside style reference images into generative video engines like MiniMax H3 (Hailuo) fails to achieve convincing artistic stylization. Initial tests revealed that when fed standard 3D video, the diffusion model prioritizes preserving the original volumetric shading, specular highlights, and surface textures, producing an awkward hybrid that retains heavy 3D CG characteristics rather than adopting a genuine 2D illustrative style.
To resolve this limitation, the developer processed the entire 15-second 3D source clip through ComfyUI using the DepthAnything model to generate a clean depth video. Stripping away all surface shaders, lighting information, and baked textures yields a pure grayscale geometric representation of the environment. Using this depth map as the structural motion conditioning signal allows the video generation model to remain free from 3D material constraints while maintaining full spatial and camera motion integrity.
Reference Image Calibration and Mitigating Temporal Frame Drift
Curating the style reference images required careful experimentation to balance visual fidelity with temporal coherence across the sequence. To anchor the artistic transformation, three key transition frames were extracted from the source 3D video and stylized into distinct concept line-art illustrations using GPT Image 2.5.
Subsequent experiments tested expanding the reference pool to six stylized images, but this introduced severe visual instability, including flickering, architectural distortion, and noticeable frame-to-frame drift as competing style prompts conflicted during generative denoising. Restricting the reference set to approximately three keyframes and granting the video model autonomy to infer intermediate transitions provided superior temporal consistency, preserving sharp linework without disruptive visual hallucinations.
Orchestration and Prompt Engineering Guided by Claude Astra
The entire end-to-end creative loop was coordinated within the Claude Astra agentic environment. From initial scene blocking in Blender via BlenderMCP to configuring data formatting between pipeline stages, Claude Astra authored the precise directorial prompts and camera motion parameters for Hailuo while systematically documenting the technical post-mortem.
By establishing spatial geometry in Blender, abstracting motion via DepthAnything, setting artistic direction through GPT Image 2.5, and delegating temporal synthesis to MiniMax H3, this modular pipeline establishes a reproducible production standard for artists seeking to bridge 3D animation and 2D illustrative craft.