GIFSplat: Generative Prior-Guided Iterative Feed-Forward 3D Gaussian Splatting from Sparse Views

Tianyu Chen; Wei Xiang; Kang Han; Yu Lu; Di Wu; Gaowen Liu; Ramana Rao Kompella

cs.CVFeb 26, 2026

GIFSplat: Generative Prior-Guided Iterative Feed-Forward 3D Gaussian Splatting from Sparse Views

Authors: Tianyu Chen, Wei Xiang, Kang Han, Yu Lu, Di Wu, Gaowen Liu, Ramana Rao Kompella

Organizations: La Trobe University, Melbourne, VIC 3086, Australia · Cisco Research, San Jose, CA, USA

Abstract

Feed-forward 3D reconstruction offers substantial runtime advantages over per-scene optimization, which remains slow at inference and often fragile under sparse views. However, existing feed-forward methods still have potential for further performance gains, especially for out-of-domain data, and struggle to retain second-level inference time once a generative prior is introduced. These limitations stem from the one-shot prediction paradigm in existing feed-forward pipeline: models are strictly bounded by capacity, lack inference-time refinement, and are ill-suited for continuously injecting generative priors. We introduce GIFSplat, a purely feed-forward iterative refinement framework for 3D Gaussian Splatting from sparse unposed views. A small number of forward-only residual updates progressively refine current 3D scene using rendering evidence, achieve favorable balance between efficiency and quality. Furthermore, we distill a frozen diffusion prior into Gaussian-level cues from enhanced novel renderings without gradient backpropagation or ever-increasing view-set expansion, thereby enabling per-scene adaptation with generative prior while preserving feed-forward efficiency. Across DL3DV, RealEstate10K, and DTU, GIFSplat consistently outperforms state-of-the-art feed-forward baselines, improving PSNR by up to +2.1 dB, and it maintains second-scale inference time without requiring camera poses or any test-time gradient optimization.

Figures & tables

Figure 1: Conceptual comparison of reconstruction paradigms. Gradient optimization performs thousands updates, incurring heavy test-time cost, often achieving high quality in dense-view scenarios but struggling in sparse-view scenarios; One-shot feed-forward [ 7 ] is efficient but leaves noticeable artifacts; Our iterative residual feed-forward scheme keeps feed-forward efficiency and achieves higher reconstruction quality without test-time gradient backpropagation.

Figure 2: Overview of GIFSplat. Our framework consists of a Gaussian initializer, an iterative Gaussian head, and a generative prior fusion module. The initializer takes sparse input views and predicts camera parameters and initial 3DGS $g_{0}$ . The iterative head then refines the Gaussians over several forward-only steps by updating per-Gaussian parameters $g_{i}$ using residual corrections $\Delta g$ predicted from the concatenated state and cues $\{g_{i},\mathbf{o}_{i},\mathbf{p}_{i}\}$ . At each step, we render reference and novel views, compute observation evidence $\mathbf{o}_{i}$ from feature differences between input and rendered views, and derive generative prior cues $\mathbf{p}_{i}$ by enhancing the renderings with a frozen diffusion model DIFIX and taking feature-space residuals. These Gaussian-level signals are fed back into the iterative head to progressively improve the 3DGS, particularly in under-constrained regions.

Figure 3: Visualizing iterative residual refinement. Starting from the initial 3DGS prediction (left column), our iterative Gaussian head progressively refines geometry and appearance over three forward-only steps (S1–S3). The zoomed regions highlighted by red dashed boxes show reduced blur, sharper edges, and fewer artifacts as the iteration proceeds, illustrating how the proposed updates gradually improve the scene representation without test-time gradient backpropagation.

Figure 4: Generative prior fusion. Given sparse input views (left), our feed-forward 3DGS first produces a rendered view (middle). A frozen diffusion-based enhancer then refines this rendering into an enhanced image (right) with sharper textures and richer details, which is converted into Gaussian-level prior cues for subsequent residual updates.

	Method	Small			Medium			Large			Average
	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Pose-Required	pixelNeRF [ 28 ]	18.417	0.601	0.526	19.930	0.632	0.480	20.869	0.639	0.458	19.824	0.626	0.485
	AttnRend [ 3 ]	19.151	0.663	0.368	22.532	0.763	0.269	25.897	0.845	0.186	22.664	0.762	0.269
	pixelSplat [ 1 ]	20.263	0.717	0.266	23.711	0.809	0.181	27.151	0.879	0.122	23.848	0.806	0.185
	MVSplat [ 2 ]	20.353	0.724	0.250	23.778	0.812	0.173	27.408	0.884	0.116	23.977	0.811	0.176
Pose-Free	DUST3R [ 21 ]	14.101	0.432	0.468	15.419	0.451	0.432	16.427	0.453	0.402	15.382	0.447	0.432

Table 1: Novel view synthesis performance comparison on the RealEstate10K with 2 views as input . Our method largely outperforms pose-free and pose-required methods across all overlap settings, especially for small overlap setting. The best and second-best results are highlighted.

Figure 5: Qualitative comparisons on representative indoor scenes from RealEstate10K. Columns: sparse input views, FLARE, AnySplat, our GIFSplat with generative prior, and ground truth (GT). Red dashed boxes highlight that GIFSplat recovers sharper boundaries (e.g., door frames, wall corners), more faithful textures, and fewer artifacts such as texture sticking.

Method	Pose?	Metrics @ 8 views
		PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
PixelSplat [ 1 ]	✓	22.55	0.727	0.192
MVSplat [ 2 ]	✓	22.08	0.717	0.189
FLARE [ 32 ]	✗	23.33	0.746	0.237
AnySplat [ 7 ]	✗	23.76	0.762	0.187
IFSplat(Ours)	✗	24.69	0.809	0.171

Table 2: Quantitative comparison on DL3DV with 8 input views . The pose-needed column indicates whether a method requires camera parameters. Metrics are PSNR

\uparrow

, SSIM

\uparrow

, and LPIPS

\downarrow

. Among feed-forward approaches, Ours requires neither intrinsics nor extrinsics and achieves the highest PSNR/SSIM and the lowest LPIPS. The best and second-best results are highlighted.

Figure 6: Qualitative comparisons on DL3DV . Columns: FLARE, AnySplat, GIFSplat (ours), and GT. Across representative scenes, GIFSplat preserves sharper edges and textures while suppressing blur and texture sticking compared with feed-forward baselines.

Methods	Pose?	DTU @ 2 views
Methods	Pose?	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
pixelSplat [ 1 ]	✓	11.551	0.321	0.633
MVSplat [ 2 ]	✓	13.929	0.474	0.385
NopoSplat [ 27 ]	✗	17.899	0.629	0.279
FLARE [ 32 ]	✗	17.528	0.596	0.283
AnySplat [ 7 ]	✗	18.122	0.632	0.276

Table 3: Out-of-distribution performance comparison . Our method shows superior performance when cross-dataset evaluation on DTU using the model solely trained on RealEstate10K. The best and second-best results are highlighted.

Components	Metrics
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
w/o R efinement	24.781	0.826	0.169
w/o w indow att.	25.327	0.837	0.152
w/o G en. Prior	26.291	0.854	0.145
Full	26.559	0.867	0.138

Table 4: Ablation study. We ablate the iterative refinement module ( w/o R efinement), the window attention ( w/o w indow att.), and the generative prior ( w/o G en. prior) from our full model, respectively. The full configuration achieves the best performance across all metrics. The best and second-best results are highlighted.

Steps	Metrics@IFSplat			Metrics@GIFSplat
Steps	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
initial	24.901	0.831	0.164	24.901	0.831	0.164
1	25.512	0.843	0.155	25.774	0.845	0.151
2	25.873	0.847	0.150	26.107	0.849	0.147
3	26.291	0.854	0.145	26.559	0.867	0.138
4	26.311	0.855	0.146	26.561	0.865	0.137

Table 5: Analysis of iterative steps. We report metrics as the number of refinement steps

T

increases from the initial prediction, for both IFSplat and our GIFSplat. Performance improves steadily and saturates around 3 iterations.

Figure 7: Time complexity analysis. Inference time of our methods IFSplat and GIFSplat on two datasets as a function of the number of refinement steps. Both methods scale approximately linearly with

T

Explore similar work

ForeSplat: Optimization-Aware Foresight for Feed-Forward 3D Gaussian Splatting
May 21, 2026Yuke Li, Weihang Liu, Cheng Zhang +83D Gaussian Splatting Meta-Learning
ShanghaiTech University · GGU Technology Co., Ltd · Stereye
ViewSplat: View-Adaptive 3D Gaussian Splatting for Feed-Forward Synthesis
Mar 26, 2026Moonyeon Jeong, Seunggi Min, Suhyeon Lee +13D Gaussian Splatting Novel View Synthesis
University of Seoul · Korea Electronics Technology Institute
GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens
Apr 16, 2026Roni Itkin, Noam Issachar, Yehonatan Keypur +33D Gaussian Splatting 3DGS Acceleration
The Hebrew University of Jerusalem · Westlake University

May 21, 2026J/K move · Enter open · S save

ForeSplat: Optimization-Aware Foresight for Feed-Forward 3D Gaussian Splatting

Yuke Li, Weihang Liu, Cheng Zhang, Yuefeng Zhang+7

ShanghaiTech University · GGU Technology Co., Ltd · Stereye

3D Gaussian Splatting Meta-Learning 3D Reconstruction Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

Feed-forward 3D Gaussian Splatting models offer fast single-pass reconstruction,but scaling them to match per-scene optimization quality is fundamentally hindered by the scarcity of large-scale 3D annotations. A practical compromise is predict-then-refine,where post-prediction optimization compensates for the limited capacity of the feed-forward network. However,standard feed-forward 3DGS is trained solely for zero-step rendering error,ignoring whether its output constitutes a good initialization for the downstream optimizer. We present ForeSplat,an optimization-aware training framework that equips feed-forward 3DGS models to produce initializations explicitly designed for rapid,effective refinement. By offloading part of the scene-modeling burden to the optimizer,ForeSplat substantially reduces the capacity pressure on the feed-forward model,making high-quality reconstruction feasible even with compact networks. At its core is MetaGrad,a lightweight multi-anchor meta-gradient training rule that bypasses costly higher-order differentiation through the 3DGS optimizer. MetaGrad unrolls a short inner-loop refinement trajectory,samples anchor states,and back-propagates aggregated first-order gradients to the prediction head as a surrogate optimization-aware signal. This fine-tuning adds no inference cost and enables high-quality reconstruction within seconds after a few refinement steps. We instantiate ForeSplat on diverse backbones,including AnySplat,Pi3X,and a distilled variant tailored for edge deployment. Across all tested architectures,a ForeSplat-trained initialization converges in fewer refinement steps and reaches a higher peak reconstruction quality than its vanilla counterpart,even fully converged. The framework consistently bridges the gap between amortized prediction and per-scene optimization,establishing a practical path toward lightweight,high-fidelity 3D reconstruction.

Abstract

Explore similar work

ForeSplat: Optimization-Aware Foresight for Feed-Forward 3D Gaussian Splatting

ViewSplat: View-Adaptive 3D Gaussian Splatting for Feed-Forward Synthesis

GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens