> Paper: Pose-Aware Diffusion for 3D Generation > Authors: Zihan Zhou, Luxi Chen, Jingzhi Zhou, Yuhao Wan, Min Zhao, Baoyu Fan, Chongxuan Li > arXiv: 2605.00345 | 2026-04-29
1. The Awkwardness of "Generate First, Pose Later"
Imagine designing a game character. The traditional 3D generation workflow looks like this:
1. Generate a 3D model in a canonical pose 2. Manually rotate it into the target pose 3. Problems:
- Rotation may deform the geometry
- Spatial mismatch
- Pose ambiguity
- Post-processing required
- Spatial mismatch: canonical and target poses have different geometric structures; rotation causes deformation.
- Pose ambiguity: correspondence between different poses of the same object is uncertain, making learning difficult.
- Complex post-processing: manual adjustment, low automation, poor efficiency.
- Spatial consistency: generation happens directly in the target space — no rotation-induced deformation.
- No ambiguity: pose is an explicit condition, so the generation target is clear and learning is simpler.
- End-to-end: one-step generation, no post-processing, higher efficiency.
- Solve the root cause > patch the symptoms
- The right problem > an elegant solution
- End-to-end > pipeline
It's like generating a "standing person" and then forcibly bending them into a "running pose" — joints may twist, clothes may clip, and the result feels unnatural. What's needed is direct generation of the target pose: no intermediate steps, end-to-end, spatially consistent.
2. PAD: Generating Directly in Observation Space
The paper proposes Pose-Aware Diffusion (PAD):
> Core idea: Abandon the "canonicalize first, rotate later" paradigm. Generate 3D geometry directly in observation space, using monocular depth as a 3D geometric anchor to eliminate pose ambiguity at the root.
Technical approach:
1. Observation-space generation — no canonical pose; generate directly at the target pose, in one step, without ambiguity. 2. Monocular depth back-projection — back-project a 2D depth map into a partial point cloud that serves as a 3D geometric anchor, providing spatial constraints. 3. Explicit pose injection — pose information is injected explicitly rather than learned implicitly, enabling direct and precise control. 4. End-to-end diffusion — from noise to 3D with the target pose as a condition; no post-processing needed.
Analogy: the traditional method is like sculpting a clay figure and then bending it into a pose; PAD is like sculpting the clay directly into the target pose — one step, more natural, more accurate.
3. Why Direct Generation Beats "Canonicalize + Rotate"
Problems with traditional methods:
PAD's advantages:
4. A Feynman-Style Judgment: Doing the Right Thing Matters More Than Doing Things Right
Quoting Feynman's spirit — "knowing the name of something" differs entirely from "understanding something":
> "Generate in canonical pose, then rotate" is *correctly doing the wrong thing*. The root problem: a 3D object's pose cannot be fixed by post-processing — pose is part of the generation process. PAD's insight is to generate directly in the target pose space, fundamentally eliminating ambiguity.
This reflects good design thinking:
5. Takeaways
If you work on generative models or 3D vision, ask yourself:
1. Does my generation pipeline have unnecessary intermediate steps? 2. Is pose part of generation, rather than a post-processing step? 3. Can depth information serve as a 3D anchor? 4. Does end-to-end generation outperform a pipeline?
PAD reminds us: the best 3D generation isn't "generate + transform" — it's "generate directly."
When diffusion models learn to "sculpt" 3D objects directly in the target pose within observation space, they evolve from "canonical model generators" into "pose-aware sculptors." In 3D generation, the best models aren't the most flexible, but the most direct and precise. In the 3D world, pose isn't clothing — it's the skeleton.