Research map / 3D vision / 3D generation
CoSER: Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoSER, a hybrid attention + selective state space model (Mamba) multiview text-to-image generator that fine-tunes Stable Diffusion LDM on G-Objaverse to produce dense, consistent 12-view images for text-to-3D.
Problem
Generating dense, consistent multiview images from text prompts for 3D asset creation; existing dense cross-view attention is expensive and sparse/sparse-view methods introduce inconsistencies.
Approach
Combines neighbor-view attention (appearance awareness and rotation-based detail refinement over a 3x3 window in adjacent views) with a spiral bidirectional Mamba scan over all views and text-score-based weighted pooling for sparse global attention correction.
Key results
- CoSER achieves quality score 33.07 vs 32.19 for VideoMV and 31.26 for Hash3D; alignment 37.7 vs 34.4 for VideoMV.
- In user study, CoSER is preferred at 0.31 semantic, 0.29 consistency, 0.32 texture, the highest of all compared methods.
- Adds about 103M parameters and about 460 GMACs over the AA-only baseline (969.95M / 1487G vs 867.45M / 1036G).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoSER instead of VideoMV | Text-to-3D generation on T3Bench prompts · T3Bench prompts, generated 3D assets | Quality: 32.19 -> 33.07 | Helped |
| Use CoSER instead of VideoMV | Text-to-3D generation on T3Bench prompts · T3Bench prompts, generated 3D assets | Alignment: 34.4 -> 37.7 | Helped |
| Use CoSER instead of VideoMV | Text-to-3D generation on T3Bench prompts · T3Bench prompts, generated 3D assets | Consistency preference: 0.25 -> 0.29 | Helped |
| Use CoSER instead of VideoMV | Text-to-3D generation on T3Bench prompts · T3Bench prompts, generated 3D assets | Texture preference: 0.22 -> 0.32 | Helped |
| Add Detail Refinement to Appearance Awareness (ablation) | CoSER · Multiview image consistency | Cross-view consistency | Helped |
| Add Rapid Glance after Appearance Awareness and Detail Refinement (ablation) | CoSER · Multiview image generation | Ambiguity reduction | Helped |
| Add Accumulated Inconsistency Rectification after Rapid Glance (ablation) | CoSER · Multiview image consistency | Cross-view consistency | Helped |
| Replace Rapid Glance's spiral bidirectional scan with spatial-first bidirectional scan (ablation) | CoSER · Multiview image generation | Consistency | Hurt |
| Add spiral bidirectional scan to Rapid Glance (ablation) | CoSER · Multiview image generation | Semantic and view consistency | Helped |
| Remove the score map from Accumulated Inconsistency Rectification (ablation) | CoSER · Multiview image generation | Retention of critical information | Hurt |
| Remove Rapid Glance while retaining the score map (ablation) | CoSER · Multiview image generation | Score map accuracy | Hurt |
| Replace Rapid Glance's SSM with temporal attention (ablation) | CoSER · Multiview image generation | Performance | Hurt |
Surprising findings
- Temporal attention for global consistency impaired performance compared with the proposed SSM-based rapid glance.
Limitations
- Global consistency relies on sparse pooled attention; dependence on learned text-relevance scores not explicitly supervised.
- Evaluation relies on CLIP score and small-scale user study (100 answers on 30 text-3D pairs).
Experimental setup
Datasets: G-Objaverse LAION-5B
Benchmarks: T3Bench prompts (100 prompts)
Metrics: CLIP score text-caption recall (alignment) user study preference (semantic, consistency, texture)
Baselines: GaussianDreamer Hash3D VideoMV
Builds on: Latent Diffusion Models Mamba MVDream VideoMamba NeuS
Closest papers by meaning
- ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models (CVPR 2024 · similarity 0.88)
- IT3D: Improved Text-to-3D Generation with Explicit View Synthesis (AAAI 2024 · similarity 0.87)
- IM-3D: Iterative Multiview Diffusion and Reconstruction for High-Quality 3D Generation (ICML 2024 · similarity 0.86)
- Flex3D: Feed-Forward 3D Generation with Flexible Reconstruction Model and Input View Curation (ICML 2025 · similarity 0.86)
- EfficientDreamer: High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors (CVPR 2024 · similarity 0.86)
- 3D-SceneDreamer: Text-Driven 3D-Consistent Scene Generation (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.