Existing feed-forward Gaussian splatting models focus on visual quality, while the few approaches targeting geometric quality are limited to indoor scenes and image pairs. This thesis focuses on enabling feed-forward models to achieve geometrically accurate 3D representations of unbounded scenes from 8 or 16 sparse views at 256 by 448 resolution, evaluated via extracted meshes and rendered depth maps. Leveraging the recent ReSplat model, we make three major contributions towards this goal: First, we stabilize training by applying a 3D filter to the predicted Gaussians, preventing them from exceeding the Nyquist frequency. Second, we explore the behavior of common self-supervised losses for geometric accuracy in feed-forward models. We find that edge-aware depth smoothness and image warp losses lead to unstable training. Instead, we use an edge-aware normal consistency loss. Third, we successfully fine-tune a ReSplat encoder, trained with gsplat as its decoder, using the Gaussian Opacity Fields and the Geometry-Grounded Gaussian Splatting (GGGS) decoders alongside the edge-aware normal consistency loss. These fine-tuned encoders achieve similar geometric accuracy and higher visual quality compared to training from scratch. We evaluate our GGGS fine-tuned encoder and observe strong performance across three key metrics. For mesh reconstruction on the Tanks and Temples dataset, our method improves the F1-score by 30% over the photometric-only encoder, performs on par with per-scene optimization, and surpasses Depth Anything 3 by a factor of six. For depth estimation across MipNeRF360, Tanks and Temples, and DL3DV, our method reduces absolute relative distance by 7% against the photometric-only baseline, 31% against Depth Anything 3, and performs on par with per-scene optimization. These improvements come at a minimal cost to visual fidelity, maintaining a PSNR within 0.12dB of the photometric-only encoder.