DynaMesh: Dynamic 3D Texture Generation

Raj Hansini; Guan Chen; Rana Hanocka; Itai Lang

cs.CVOct 4, 2026

DynaMesh: Dynamic 3D Texture Generation

Authors: Raj Hansini, Guan Chen, Rana Hanocka, Itai Lang

Organizations: University of Chicago

Abstract

We present DynaMesh, a dynamic texture generation method for 3D meshes. Given a textureless shape and a text prompt describing an effect, our method produces an appearance that evolves while the object's geometry remains unchanged. Previous works on dynamic 3D content generation have focused on motion, where an object's geometry and position change while keeping its appearance the same. Methods on texture generation sit on the other side of the problem, painting appearance onto a shape as a fixed surface property and not as an evolving process. Neither addresses a visual effect that propagates on a 3D object. A natural route consists of two generators: a video model that shows the effect from a single view, and an image-to-3D generator that lifts each frame to 3D. However, the latter has no notion of time, so running it per video frame produces a sequence that flickers, loses effect details, and yields a different mesh at every video frame. Our method addresses these failures by conditioning a video model on a render of the mesh and the prompt to obtain a reference video, then running a frozen image-to-3D generator on the video with two changes. The conditioning of each frame is blended over a temporal window, and low-rank adapters are fit per shape to restore the lost details. The mesh is encoded once for the whole sequence, so geometry is constant by construction, and the output is a single mesh with a texture per frame. Applied to various objects and effects, DynaMesh substantially improves over recent video-to-4D and texturing methods, and can generalize its temporal effect to different shapes never seen during training. Our project page is at https://threedle.github.io/dynamesh/.

Figures & tables

Figure 2 : Gallery of results. Four objects driven by four different effects. Each panel shows the driving video as a film strip and, below it, our output rendered from two novel viewpoints. The texture remains coherent even at viewpoints unseen in training.

Figure 3 : System overview. A render of the input mesh, together with a text prompt describing the effect, is passed to a video generation model, which returns the reference video for our pipeline. The 3D generator then produces the textured object at every frame. Our temporal cross-attention (the new module) conditions each frame’s generation on a window of neighboring video frames. Low-rank adapters on the attention modules are fit per scene to recover detail. Geometry is shared across frames, so the output is one mesh whose texture evolves with the video.

Figure 4 : Method overview. The token maps of an eleven-frame window are blended by a temporal attention layer. The blend conditions the generator’s spatial cross-attention, which writes the frame’s appearance onto the voxel tokens of the input mesh, and the 3D self-attention then propagates that appearance across the surface. Every weight except the rank-4 adapters on those two attentions is frozen.

Figure 5 : Generalization. Per-scene adapters fit on a reference video are applied, without any retraining, to meshes never seen during fitting. Both effects clearly propagate on the new shapes according to the video. The adapters change only how appearance is generated, and each object keeps its own geometry, so the same spreading blots or lava cracks work even on different shapes.

Figure 6 : Temporal coherence. A chair with moss spreading over its surface is shown at four time frames. The insets magnify the same patch of the backrest. Frozen TRELLIS.2 changes the texture abruptly from one moment to the next (flickers), while our method keeps the moss growing smoothly, as in the video.

Figure 7 : Making an existing static texture dynamic. Our method can operate on objects with a given texture and make it change over time, as shown for the moving spots over the duck.

	Unseen views		Fidelity vs . GT
Method	Flicker $\downarrow$	Accel. $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
SV4D 2.0 † [ 53 ]	0.023	0.035	34.8	0.963
DG4D † [ 35 ]	0.074	0.142	13.3	0.393
L4GM [ 36 ]	0.007	0.011	21.6	0.728
MeshNCA [ 31 ]	0.018	0.033	7.7	0.105
TRELLIS.2 [ 47 ]	0.018	0.027	12.1	0.303

Table 1: Quantitative comparison over 42 objects and effects and all 150 frames. Flicker and acceleration are per-step temporal errors, and PSNR/SSIM, fidelity metrics to the ground truth video, are computed in the supervised view against the video. We achieve better results than state-of-the-art methods. † SV4D 2.0 and DG4D are grayed out as their outputs have fewer frames and cannot be measured on the same frame grid, discussed more below.

Figure 8 : Qualitative comparison. We compare our method against the baselines on one reference video (top strip), with each method shown at four timesteps from a single unseen view. The baselines either flicker or degrade the geometry at that view. Only our method tracks the effect consistently and preserves the shape’s geometry.

Figure 9 : Failure case. DynaMesh relies on effects where the property of the pattern is local. Here, it replicates the local patterns (flowers, dots, etc.), but it does not recover them from unseen views, since the flowers and lines follow a global arrangement that a single view does not capture.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 10 : Additional gallery. DynaMesh generates a different temporal effect for each object, and the effect holds at an unseen viewpoint despite being fit to a single camera.

Figure 11 : Qualitative comparison. Given the same reference video, the baselines either lose the effect or distort the geometry at the unseen viewpoint, whereas DynaMesh preserves both.

Figure 12 : Comparison at the second novel view. All methods are shown from a camera behind the object, which none of them was supervised on, at the same four frames as Fig. 11 . The baselines that construct their own geometry or operate in video space lose the effect. In contrast, our method presents the same texture it shows at the supervised view, since the effect is written onto the input mesh rather than onto a camera.

Figure 13 : Flicker over the clip. We plot flicker at each time step for frozen TRELLIS.2 and for DynaMesh, over the 42 objects of Tab. 1 . The gap holds across the whole clip and is largest at the start, before the frozen texture settles.

Attention			Unseen views		Fidelity vs . GT
Cross	Self	Temp.	Flicker $\downarrow$	Accel. $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
✗	✗	✗	0.01818	0.02720	12.14	0.3035
✓	✗	✗	0.00727	0.01039	23.80	0.7542
✓	✓	✗	0.00702	0.00998	24.86	0.7970
✓	✓	✓	0.00499	0.00531	24.89	0.7980

Appendix

Table 2: Ablation study. We turn on the components of our method one at a time and evaluate on the same 42 objects as Tab. 1 , with the same metrics. The first row is the frozen generator and the last row is our full method, which is best on every metric.

Figure 14 : Continuing an effect across clips. A second reference video, generated from the last frame of the first, continues the same effect, so a progression that does not fit in one clip can be carried across several.

Figure 15 : MeshNCA conditioning. Conditioned on a single frame, MeshNCA reproduces it and does not advance through time, so its result reflects the conditioning image rather than the progression of the effect.

	Q1 effect fidelity (%) $\uparrow$		Q2 temporal coherence $\uparrow$		Q3 geometry $\uparrow$
Method	Supervised	Unseen	Supervised	Unseen	Supervised	Unseen
Frozen TRELLIS.2	85.9	85.6	3.66	3.71	4.90	4.86
MeshNCA	31.4	32.9	2.62	2.71	4.70	4.75
SV4D 2.0	98.4	72.3	4.55	4.04	4.98	3.62
DreamGaussian4D	45.4	27.4	2.66	2.42	4.69	3.65
L4GM	98.9	60.3	4.64	3.75	4.90	3.47

Appendix

Table 3: LLM as a judge. We ask a vision-language model three questions about each rendered clip, at the supervised view and at the three unseen views, over all 42 objects with three repeats per clip. Effect fidelity is the percentage of five yes/no questions passed, while temporal coherence and geometry are rated from 1 (worst) to 5 (best). Our method scores best at the unseen views on all three questions, and the gap to the video-based baselines widens as the camera moves away from the supervised one.

Effect	Shapes	Flick. $\downarrow$	Accel. $\downarrow$	Q1 (%) $\uparrow$	Q2 $\uparrow$
Rorschach	source	0.004	0.005	100.0	5.00
Rorschach	unseen	0.006	0.009	100.0	4.73
Lava	source	0.007	0.005	86.7	5.00
Lava	unseen	0.007	0.006	85.5	4.64

Appendix

Table 4: LLM as a Judge for Generalization. Flicker, acceleration, and judge scores (Q1 effect fidelity, Q2 temporal coherence) for each adapter on the shape it was fit on and averaged over eleven unseen shapes. Effect fidelity is the percentage of five shape-neutral questions passed per effect, generated from the reference video, so the same questions score every shape. The transferred effects keep the temporal quality and the fidelity of the source result.

Question	Reference	Score
Q1 effect fidelity	reference video	% of 5 questions passed
Q2 temporal coherence	none	1 to 5
Q3 geometry correctness	gray render of the input mesh	1 to 5

Appendix

Table 5: Judge protocol. We ask three questions about every rendered clip, each scored against the reference its claim needs.

Figure 16 : Temporal window width. Widening the conditioning window reduces flicker, but fidelity peaks at a width of five and declines beyond it, so we select eleven to balance the two.

	Temporal		Reconstruction
Temporal operator	Flicker $\downarrow$	Accel. $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$
Temporal attention	0.00680	0.00921	24.89	0.7171
Spatio-temporal attn.	0.00683	0.00924	24.86	0.7165
Parameter-free (ours)	0.00506	0.00474	24.82	0.7149

Appendix

Table 6: Temporal operator ablation. We compare two learned temporal operators against the parameter-free blend, at the same eleven-frame window with everything else held fixed. Metrics in the table are computed on the texture rather than on the renders, so they differ from Tab. 2 . The blend has lower flicker and acceleration on all 42 objects, at a cost of

0.07

dB in PSNR and

0.002

in SSIM.

Stage	Setting	Value
Reference video	model	Kling 3.0 [ 24 ]
	clip length	121 to 150 frames
	training res.	$960^{2}$
Generator	backbone	TRELLIS.2 [ 47 ]
	image encoder	DINOv3 ViT-L/16 [ 39 ]
	flow blocks	30

Appendix

Table 7: Implementation details. We list the generator, the blending window, the adapter, the loss, and the hardware used to produce every result in the paper. All objects are fit with these same values.

Explore similar work

Jul 1, 2026cs.CV

Ink3D: Sculpting 3D Assets with Extremely Complex Textures via Video Generative Models

Recent 3D generative models can synthesize high-quality geometry but often struggle to reproduce intricate textures from reference images, largely due to the scarcity of large-scale 3D training data with rich surface appearance. In contrast, visual generative models are trained on datasets several orders of magnitude larger and excel at modeling complex visual patterns. Motivated by this gap, we introduce Ink3D, a framework that bridges 3D generation with large-scale video generative models to synthesize extremely complex textures. Ink3D first reconstructs a white-mesh geometry using an off-the-shelf 3D generation model. It then employs OrbitPainter, a conditional video generative model, to produce dense orbit-scan videos capturing object appearance across viewpoints. To convert these views into coherent textures, we introduce TextureOptimizer, a neural baking module that integrates dense multi-view observations while mitigating geometry inconsistencies arising from video generation. By decoupling geometry and texture synthesis and leveraging large-scale pretrained video priors, Ink3D enables significantly richer and more faithful texture generation than prior approaches.

Yue Han, Chong Li, Zhening Liu +5

ZGCA & ZGCI · Zhejiang University · Microsoft Research +1

3D Generative Models 3D Assets Texture Generative Video Models 3D Generation Video Generation+1

Ask PDF

Apr 29, 2026cs.CV

AnimateAnyMesh++: A Flexible 4D Foundation Model for High-Fidelity Text-Driven Mesh Animation

Recent advances in 4D content generation have attracted increasing attention, yet creating high-quality animated 3D models remains challenging due to the complexity of modeling spatio-temporal distributions and the scarcity of 4D training data. We present AnimateAnyMesh++, a feed-forward framework for text-driven animation of arbitrary 3D meshes with substantial upgrades in data, architecture, and generative capability. First, we expand the DyMesh-XL dataset by mining dynamic content from Objaverse-XL, increasing the number of unique identities from 60K to 300K and substantially broadening category and motion diversity. Second, we redesign DyMeshVAE-Flex with power-law topology-aware attention and vertex-normal enhanced features, which significantly improves trajectory reconstruction, local geometry preservation, and mitigates trajectory-sticking artifacts. Third, we introduce architectural changes to both DyMeshVAE-Flex and the rectified-flow (RF) generator to support variable-length sequence training and generation, enabling longer animations while preserving reconstruction fidelity. Extensive experiments demonstrate that AnimateAnyMesh++ generates semantically accurate and temporally coherent mesh animations within seconds, surpassing prior approaches in quality and efficiency. The enlarged DyMesh-XL, the upgraded DyMeshVAE-Flex, and variable-length RF together deliver consistent gains across benchmarks and in-the-wild meshes. We will release code, models, and the expanded DyMesh-XL upon acceptance of this manuscript to facilitate research in 4D content creation.

Zijie Wu, Chaohui Yu, Fan Wang +1

School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China · DAMO Academy, Alibaba Group, Hangzhou, China · Hupan Lab, Hangzhou, China

4D Generation High-Fidelity Wireless Foundation Models

Ask PDF

Jun 1, 2026cs.CV

MORPHOS: Autoregressive 4D Generation with Temporal Structured Latents

We present MORPHOS, a novel autoregressive framework that generates dynamic 3D assets from videos across diverse representations, including meshes, 3D Gaussians, and radiance fields. Existing methods are typically limited to a single representation, struggle to model topological changes, or fail to maintain temporal consistency over long videos. To address these limitations, we introduce the Temporal Structured Latents (T-SLAT), a unified 4D representation that jointly encodes geometry and appearance along the temporal dimension. Leveraging T-SLAT, MORPHOS autoregressively generates dynamic 3D assets via causal attention, conditioning each frame on its preceding history to ensure temporal consistency while handling evolving topologies. We also propose a temporal-structural augmentation to mitigate error accumulation in autoregressive generation. MORPHOS achieves state-of-the-art performance in appearance and competitive results in geometry across multiple benchmarks, demonstrating superior generalization across various representations and robustness in long-horizon generation.

Minkyung Kwon, Jinhyeok Choi, Youngjin Shin +3

KAIST AI

Autoregressive Video Generation 3D Assets Autoregressive Generation 3D Gaussian Stemming

Ask PDF