Meta tags:
description=Generalizable Sparse-View 3D Reconstruction from Unconstrained Images paper. Official web with qualitative comparisons, links to the source code, and additional materials.;
keywords=genwildsplat,3dgs,nerf,official,code;

Headings (most frequently used words):

of, appearance, view, reconstruction, from, input, scene, on, comparison, genwildsplat, sparse, video, results, count, scenes, views, methods, the, side, in, generalizable, 3d, unconstrained, images, across, varying, curated, megascenes, dataset, analysis, failure, cases, constant, renderings, flexible, viewpoint, and, control, cross, illumination, transfer, effect, interpolation, depth, prediction, our, framework, upper, bound, with, prior, synthetic, data, generation, pipeline, for, curriculum, training, gallery, 100, against, state, art, by, modeling, large, motion, changes, small, objects, related, work, bibtex, tl, dr, feed, forward, wild, 3s, a6000, gpu,

Text of the page (most frequently used words):
the (45), and (27), views (26), view (22), input (20), #appearance (18), from (17), our (16), genwildsplat (15), scene (13), lighting (13), with (12), model (12), reconstruction (10), video (9), #sparse (8), nexussplats (8), wildgaussians (8), novel (8), scenes (8), images (7), rendering (7), methods (7), method (7), results (7), geometry (6), transfer (6), depth (6), that (6), while (6), wild (5), reconstructions (5), analysis (5), ours (5), more (5), across (5), are (5), different (5), unconstrained (4), megascenes (4), forward (4), gaussian (4), for (4), transient (4), objects (4), next (4), stage (4), only (4), illumination (4), generalizable (3), outdoor (3), feed (3), splatting (3), preserving (3), artifacts (3), segmentation (3), prev (3), under (3), comparison (3), inputs (3), training (3), using (3), dataset (3), synthetic (3), this (3), approach (3), appearances (3), prior (3), quality (3), varying (3), geometric (3), prediction (3), enables (3), same (3), count (3), gupta (2), vinayak (2), lin (2), chih (2), hao (2), wang (2), shenlong (2), bhattad (2), anand (2), huang (2), jia (2), bin (2), cvpr (2), 2026 (2), synthesis (2), scale (2), large (2), visual (2), loss (2), diffusionrenderer (2), cases (2), struggles (2), accurately (2), leading (2), rendered (2), output (2), building (2), statues (2), train (2), occlusions (2), target (2), struggle (2), holes (2), these (2), framework (2), use (2), controls (2), below (2), through (2), failure (2), changes (2), each (2), baseline (2), conditions (2), side (2), varied (2), yields (2), consistent (2), against (2), browse (2), test (2), seen (2), during (2), performance (2), increasing (2), diverse (2), generating (2), data (2), allowing (2), 216 (2), significantly (2), curated (2), viewpoints (2), both (2), which (2), representation (2), light (2), encoder (2), captures (2), gaussians (2), rgb (2), show (2), between (2), demonstrate (2), robust (2), free (2), can (2), not (2), flexible (2), viewpoint (2), ability (2), consistency (2), constant (2), code (2), post (2), article, gupta2026genwildsplat, title, author, journal, year, bibtex, level, datasets, anysplat, vggt, grounded, transformer, generalisable, ccpl, contrastive, coherence, versatile, style, neural, inverse, diffusion, models, efficient, related, work, certain, predict, double, case, internal, disentangle, hence, gets, baked, existing, wrongly, classifies, most, effective, solution, new, identifies, tourists, ignoring, when, far, away, may, generate, coherent, examples, illustrate, current, limitations, navigate, four, discussed, scenarios, motion, small, left, right, pair, compares, identical, modeling, previous, row, shows, result, state, art, hundreds, never, reconstructed, showcasing, its, wide, range, settings, gallery, 100, three, fashion, progressively, complexity, single, dl3dv, employ, relight, random, help, resolve, ambiguities, extends, multiple, generalize, iii, introduce, task, remove, them, representations, learned, earlier, stages, generation, pipeline, curriculum, require, accurate, nexussplat, reaches, around, whereas, achieves, high, just, upper, bound, showcase, few, set, used, evaluations, include, significant, variations, occluders, providing, challenging, generalization, capability, given, unposed, first, extracts, multi, features, capturing, semantic, information, dedicated, heads, estimate, camera, parameters, attributes, then, mapped, into, canonical, per, image, modulate, colors, consistently, adapter, pre, trained, network, masked, out, focuses, static, content, photorealistic, reference, rasterizing, centers, slider, interactive, blending, moving, blue, dot, along, axis, resulting, smooth, transitions, learns, semantically, meaningful, codes, effectively, interpolation, directly, correlates, context, enhances, low, leads, higher, number, hole, effect, example, numbers, illustrates, improves, remains, even, two, experiments, limited, six, due, computational, constraints, though, handle, practice, one, another, enabling, controlled, such, feasible, like, couple, optimization, cross, supports, change, preserve, combinations, control, full, relighting, typically, maintain, renderings, note, all, presented, real, world, were, showing, processed, syncfix, whenever, compare, baselines, don, process, ensure, differences, remain, clearly, visible, a6000, gpu, paper,

Text of the page (random words):
generalizable sparse view 3d reconstruction from unconstrained images genwildsplat generalizable sparse view 3d reconstruction from unconstrained images vinayak gupta chih hao lin shenlong wang anand bhattad jia bin huang cvpr 2026 paper code tl dr feed forward in the wild scene reconstruction from sparse views in 3s on a a6000 gpu note all video results are presented on real world scenes that were not seen during training video results showing only our method are post processed with syncfix whenever we compare against baselines we don t post process both baseline and our method to ensure differences between methods remain clearly visible constant appearance renderings our method enables rendering a scene under a constant appearance code while preserving full 3d consistency across views an ability that 2d appearance transfer or 2d relighting methods typically struggle to maintain input appearance 1 appearance 2 appearance 3 flexible viewpoint and appearance control our method supports flexible appearance change and free viewpoint rendering from a 3d scene representation the results demonstrate the model s ability to preserve geometric consistency while generating novel combinations of viewpoints and illumination conditions input view same view different lighting different view same lighting different view different lighting cross scene illumination transfer genwildsplat can transfer lighting or appearance from one scene to another enabling controlled appearance changes while preserving geometry such appearance transfer is not feasible with prior methods like wildgaussians and nexussplats which couple appearance and geometry optimization video results across varying input view count 2 input views example reconstructions from varying numbers of input views use the controls below to browse results for 2 6 input images this illustrates that reconstruction quality improves with more views while genwildsplat remains robust even with only two inputs experiments are limited to six views due to computational constraints though the method can handle more in practice prev next effect of input view count on scene reconstruction the results show that 3d reconstruction quality directly correlates with input view count increasing context views 2 to 6 significantly enhances novel view synthesis low view input 1 3 views leads to geometric holes and artifacts while a higher number of views 5 6 views yields a more robust hole free reconstruction 1 view 2 views 3 views 4 views 5 views 6 views appearance interpolation a slider enables interactive blending between different appearances of the same scene by moving the blue dot along the axis the resulting smooth transitions demonstrate that our light encoder learns semantically meaningful lighting codes and effectively captures diverse scene appearances depth prediction for reference we show the depth prediction rendered by rasterizing the gaussians centers rgb depth rgb depth our genwildsplat framework given sparse unposed images with varying appearance and transient objects our approach first extracts multi view features capturing both semantic and geometric information dedicated prediction heads estimate depth camera parameters and 3d gaussian attributes which are then mapped into a canonical 3d representation a light encoder captures per image illumination allowing the model to modulate the gaussians colors consistently across views using the appearance adapter using a pre trained segmentation network transient objects are masked out and the reconstruction loss focuses on static scene content this enables photorealistic view consistent reconstructions from sparse in the wild images curated scenes from megascenes dataset we showcase a few scenes from our curated set from the megascenes dataset used in our evaluations these outdoor scenes include sparse viewpoints significant illumination variations and transient occluders providing a challenging test of genwildsplat s generalization capability upper bound comparison with prior methods prior methods require significantly more views to approach accurate geometry nexussplat reaches our performance only at around 216 inputs whereas genwildsplat achieves high quality reconstructions with just 6 views input 6 views 36 views 216 views ours 6 views synthetic data generation pipeline for curriculum training we train our model in a three stage fashion with progressively increasing complexity in stage i using only a single scene from the dl3dv dataset we employ diffusionrenderer to relight images with diverse random lighting generating synthetic data to help the model resolve geometry and appearance ambiguities stage ii extends this approach to multiple scenes allowing the model to generalize across scenes and varied appearances in stage iii we introduce synthetic occlusions and task the model to remove them building on the representations learned in the earlier stages gallery of results 100 scenes browse through hundreds of megascenes test scenes never seen by our model during training reconstructed by genwildsplat showcasing its performance across a wide range of outdoor settings prev next video comparison against state of the art methods each row shows the input view reconstructions from wildgaussians and nexussplats and our genwildsplat result under sparse inputs and varied lighting our method yields more consistent novel views input wildgaussians nexussplats genwildsplat ours input wildgaussians nexussplats genwildsplat ours input wildgaussians nexussplats genwildsplat ours input wildgaussians nexussplats genwildsplat ours previous 1 6 next side by side comparison of appearance modeling each left right pair compares a baseline method with genwildsplat under identical target views and lighting conditions wildgaussians genwildsplat nexussplats genwildsplat analysis of failure cases 1 large motion changes on small objects these examples illustrate the current limitations of our framework use the controls below to navigate through the four discussed failure scenarios prev next input video novel view rendering analysis when the target novel view is far away from the input views the model may struggle to generate a coherent output leading to visual artifacts and holes input video novel view rendering analysis the existing segmentation model wrongly classifies statues as transient objects the most effective solution is to train a new segmentation model that identifies tourists as occlusions while ignoring statues input video novel view rendering analysis in the case of internal lighting in the scene lighting from the building the model struggles to accurately disentangle the lighting from the scene and hence the lighting gets baked in input video novel view rendering analysis in certain cases our model struggles to accurately predict depth leading to double geometry artifacts in the rendered output related work in the wild reconstructions wildgaussians 3d gaussian splatting in the wild nexussplats efficient 3d gaussian splatting in the wild appearance lighting transfer methods diffusionrenderer neural inverse and forward rendering with video diffusion models ccpl contrastive coherence preserving loss for versatile style transfer generalisable feed forward reconstruction vggt visual geometry grounded transformer anysplat feed forward 3d gaussian splatting from unconstrained views large scale outdoor datasets megascenes scene level view synthesis at scale bibtex article gupta2026genwildsplat title generalizable sparse view 3d reconstruction from unconstrained images author gupta vinayak and lin chih hao and wang shenlong and bhattad anand and huang jia bin journal cvpr year 2026