[CVPR 2025] RelaxFlow: Solving Amodal 3D Ambiguity with Low-Pass Semantic Corridors
RelaxFlow: Text-Driven Amodal 3D Generation
RelaxFlow is a training-free dual-branch framework designed for text-driven amodal 3D generation, enabling users to specify the semantic completion of occluded or ambiguous regions in an image. By decoupling observation fidelity from semantic intent, it achieves state-of-the-art results in generating 3D assets that are both faithful to visible pixels and consistent with textual prompts.
TL;DR
Generative AI models often struggle with "hallucination" when faced with occluded images. If you only see the back of a wooden board, is it a sofa or a bed? RelaxFlow introduces a training-free dual-branch framework that uses text prompts to steer 3D completion. By applying a low-pass relaxation to the semantic branch, it allows high-level text intent to guide the global shape while letting the original image pixels strictly dictate the visible details.
The Tension Between Pixels and Pixels
When we look at a partially occluded object, our brains perform "amodal perception"—we mentally fill in the blanks. Current image-to-3D models try to do this but face a fundamental conflict:
- Observation Fidelity: The visible parts must match the input image perfectly (Rigid Control).
- Prompt Following: The hidden parts must follow the user's text prompt (Relaxed Control).
Existing models treat these two forces with the same level of "gravity." This leads to a tug-of-war where either the 3D shape looks nothing like the prompt, or the visible pixels get blurred and distorted to accommodate the text.
Methodology: Decoupling via Relaxation
The core insight of RelaxFlow is that semantic intent is low-frequency, while visual evidence is high-frequency. If we want a prompt to guide a "sofa" shape, we don't need it to specify every textile fiber; we just need it to define the "corridor" of valid sofa-like geometries.
1. Dual-Branch ODE Flow
Instead of a single generation path, RelaxFlow runs two in parallel:
- Observation Branch: Uses the original image tokens to anchor high-frequency details.
- Semantic-Prior Branch: Uses visual proxies generated from your text prompt to "steer" the hidden geometry.
2. Multi-Prior Consensus
To avoid the quirks of a single reference image, RelaxFlow uses multiple visual priors for the same text (e.g., three different "sofas"). By concatenating these in the cross-attention layer, the model naturally pays attention to what they have in common (the "sofa-ness") and ignores the inconsistent textures.
3. Low-Pass Relaxation (The "Secret Sauce")
The authors implement a Gaussian blur on the cross-attention logits of the semantic branch.
Figure 3: The RelaxFlow pipeline showing the separation of rigid observation and relaxed semantic guidance.
Theoretically, they prove this is equivalent to applying a low-pass filter to the generative vector field. This "thickens" the semantic corridor, making it easier for the model to find a path that is both semantically correct and compatible with the hard pixel constraints of the observation.
Experimental Results: Disambiguating the Unseen
The authors introduced two new benchmarks: ExtremeOcc-3D (for objects 80% hidden) and AmbiSem-3D (for objects that could be two different things).
Figure 4: Note how RelaxFlow (bottom) correctly transforms an ambiguous wooden board into a bed based on the prompt, whereas others collapse into single, often incorrect modes.
| Method | CLIP (Text) ↑ | Point-FID ↓ | User Preference ↑ |
|---|---|---|---|
| SAM3D (Baseline) | 24.08 | 100.38 | 9.22% |
| RelaxFlow (Ours) | 27.26 | 81.11 | 68.52% |
The quantitative gains are massive. On the SAM3D backbone, RelaxFlow dropped the Point-FID (a measure of 3D semantic error) by nearly 20 points without requiring any additional training.
Deep Insight: Why This Matters
RelaxFlow proves that you don't always need more data or bigger models to solve complex 3D ambiguity. Sometimes, it's about gradient surgery. By understanding the frequency characteristics of "intent" vs. "evidence," RelaxFlow allows us to use existing powerful feedforward models in a modular, controllable way.
Limitations
While powerful, RelaxFlow can struggle if the text prompt and the image are fundamentally contradictory (e.g., trying to turn a clear image of a car into a bird). In these cases, it results in a "soft concept fusion" or morphing effect.
Conclusion
RelaxFlow bridges the gap between human-like amodal reasoning and machine-driven 3D generation. It’s a "plug-and-play" module that makes 3D generation interactive and context-aware—critical for the future of AR/VR and robotics where we rarely see the "whole" object at once.