Edit3r: Instant 3D Scene Editing from Sparse Unposed Images¶

Conference: ECCV 2026
Paper: ECCV page | Project page
Area: 3D Vision
Keywords: 3D Scene Editing / Feed-forward Reconstruction / Gaussian Splatting / Multi-view Consistency / Pose-free Reconstruction

TL;DR¶

Edit3r turns 3D scene editing into a single forward pass of sparse unposed Gaussian reconstruction: it trains on a SAM2-based recoloring pretext that is view-consistent by construction, and deliberately pairs one recolored view with one raw view so the feed-forward backbone learns to fuse conflicting per-view 2D edit proposals into a geometrically coherent edited 3D Gaussian scene, producing results in about 0.5 seconds at inference.

Background & Motivation¶

Text-driven editing of a real 3D scene is currently dominated by a reconstruct–edit–refit pipeline: reconstruct a neural representation (NeRF or 3D Gaussian Splatting) from posed images, apply 2D generative editing (InstructPix2Pix and friends) to the rendered views, then re-optimize the 3D representation to fit those edited images. Instruct-NeRF2NeRF, GaussCtrl, GaussianEditor, and EditSplat all follow this recipe and achieve impressive visual quality and instruction alignment. But the cost is twofold. Every scene must be refit from scratch, which takes minutes to tens of minutes and rules out interactive editing; and the "edit each frame, then refit" loop actively amplifies inconsistency — a 2D editor inevitably produces different results for the same object across viewpoints, and re-optimization does not resolve that disagreement so much as bake it into the 3D representation, where it surfaces as blur, floaters, and ghosting in novel views.

A second line of progress is the Large Reconstruction Model family (PixelSplat, MVSplat, GS-LRM, NoPoSplat): by amortizing computation into large-scale pretraining, these models emit Gaussians from sparse views in one forward pass with no per-scene optimization, and NoPoSplat goes further by dropping camera poses entirely and predicting Gaussians in a canonical frame directly from unposed images. This looks like exactly the substrate 3D editing needs, but using it naively runs into two walls. The first is supervision: an editing model must have seen "edited and cross-view consistent" 3D ground truth, yet no such data exists because 2D editing is stochastic and applied per image; falling back on multi-view images produced by a 2D editor amounts to injecting label noise, and training oscillates and grows blurry artifacts. The second is input consistency: a feed-forward reconstruction backbone assumes its multi-view inputs are consistent observations of one static scene, and its attention mechanisms are biased toward averaging across views — but edited inputs are precisely inconsistent, and averaging yields a blurry compromise.

This paper's angle is that when consistent ground truth is unavailable, one should train the fusion capability on a surrogate task that is consistent by construction. Core idea: reduce editing to a controllable recoloring pretext — sample one color transform per SAM2 object and reuse it across every frame, yielding supervision that is inherently multi-view consistent; then deliberately pair a recolored view with a raw view as an asymmetric input, forcing the network to propagate the edit semantics from the reference to the auxiliary view while the raw view keeps the geometry honest; at inference the same backbone is simply plugged behind any 2D editor and produces edited 3D Gaussians in a single forward pass.

Method¶

Overall Architecture¶

Given unposed images with intrinsics \(\{(I_v, k_v)\}_{v=0}^{V-1}\) and an editing prompt \(T\), the goal is to reconstruct a geometrically consistent 3D scene \(S_T\) semantically aligned with \(T\). The pipeline has exactly two stages: first a 2D image editor edits the input views into \(\{I_v^\star\}\), then Edit3r performs a single forward pass that predicts anisotropic Gaussians \(\mathcal{G}=\{(\mu_j,\Sigma_j,c_j,\alpha_j)\}\) in a fixed world frame, where \(\mu\) is the center, \(\Sigma\) the covariance, \(c\) spherical-harmonic color coefficients, and \(\alpha\) opacity; standard 3D Gaussian Splatting rasterizes them into novel views. No pose estimation, no mesh, and no test-time optimization anywhere — that is what "instant" means in the title.

The backbone follows NoPoSplat's pose-free reconstruction framework, but its role changes: it is no longer merely a reconstructor, it is the fuser of cross-view edit proposals. Concretely, each view's intrinsics are embedded by a small MLP \(\phi\) and concatenated with the image tokens before a weight-shared ViT encoder encodes every view independently into a unified feature space; a ViT decoder then fuses those features with self-attention and cross-view attention, resolving occlusions and appearance differences through attention. The fused features feed two lightweight Gaussian heads, each containing two DPT-based predictors — one regresses Gaussian centers from transformer features alone (keeping geometry stable and immune to image texture), the other additionally takes RGB image shortcuts to predict opacity, covariance, and low-order spherical harmonics (keeping appearance detail). Per-view dense Gaussians are concatenated in the canonical frame to form the whole scene, and training uses only photometric losses, so neither pose inputs nor pose-based warping are needed.

Training and inference construct their inputs differently, and this asymmetry is the crux of the paper: training feeds an asymmetric pair (one recolored view plus one raw view), whereas inference edits all frames to provide richer edit evidence. To keep evaluation honest, the authors also build DL3DV-Edit-Bench: 20 indoor/outdoor real scenes from the DL3DV test split, where Grounding-DINO first extracts object-level labels and region proposals, an LLM synthesizes candidate prompts covering four edit types (Add / Remove / Modify / Global), and manual vetting keeps 5 valid prompts per scene for 100 edit instances in total; evaluation fixes the 2D editor, shares one random seed and one text prompt across all views of a scene, and confines local edits with masks derived from the proposals.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Sparse unposed views<br/>+ text prompt"] --> B["Per-view 2D editing"]
    B --> C["Pose-free feed-forward<br/>Gaussian backbone ≈0.5 s"]
    E["SAM2-based recoloring"] --> C
    F["Asymmetric input"] --> C
    H["Random drop of reference Gaussians"] --> C
    C --> G["Gaussian-center 3D regularization"]
    C --> D["Edited 3D Gaussians<br/>consistent novel views"]

Key Designs¶

1. SAM2-based recoloring: a view-consistent surrogate for ground truth that does not exist

Supervision is what a feed-forward editing model lacks most. A 2D editor produces different results for the same object across viewpoints, so using those results as targets teaches the network to average contradictory labels, yielding unstable convergence and blurry output — in the ablation, replacing recoloring supervision with 2D edited images degrades C-FID from 171.3 to 215.0 and C-KID from 0.116 to 0.141. The authors instead approximate editing with a controllable recoloring task. SAM2's automatic mask generator proposes all candidate objects on the first frame without any manual prompt, each proposal carrying a soft mask, area, stability score, predicted IoU, and a bounding box; a deterministic multi-criterion filter then keeps a proposal only if its area, stability, and predicted IoU clear thresholds, its aspect ratio falls inside a range, and its box retains enough pixel margin from the image border. Retained instances receive persistent IDs and prompt SAM2's video segmentation predictor, which propagates their masks along the sequence. Identity drift is suppressed by processing a frame only when its set of active object IDs overlaps the most recently processed frame by at least 0.5; frames below that ratio are skipped entirely and excluded from every later stage.

Recoloring itself is a per-region composite color transform: one parameter set \(\Theta_r\) per region — combining ColorJitter, gamma correction, a PCA-based lighting offset, a fixed RGB-channel permutation, and optional grayscale conversion — is sampled once and then reused across all frames and views, so a given object is tinted identically from every angle and cross-view consistency is constructed rather than inferred. Where regions overlap, the per-pixel soft masks are renormalized to sum to one (with a small \(\epsilon\) in the denominator for numerical stability), and the recolored image is a soft blend in which the residual weight goes to the original pixels — with a single region this reduces to ordinary alpha blending. The background receives the same color transform to reduce distribution shift. On top of this, the SAM2 masks enable object-level augmentation (mask dilation/erosion, palette perturbation, limited background leakage) that narrows the gap between recolored training images and real edited images; removing that augmentation in the ablation visibly increases failures under strong edits. The decisive evidence that the design works is that a model trained only on recoloring transfers directly to the outputs of real editors such as IP2P and FLUX at inference — it has learned to fuse inconsistent observations and align them to the reference view, not to solve recoloring per se.

2. Asymmetric input: manufacturing cross-view conflict so the network learns to fuse

Consistent supervision alone is not enough. If every view were recolored consistently during training, the network would never encounter two views disagreeing about the same region and would never learn to handle conflict — yet real edited inputs are exactly that. Training therefore constructs an asymmetric pair on purpose: one recolored reference view \((I'_0, k_0)\) and one completely unprocessed view \((I_1, k_1)\). Edit semantics exist only in the reference view, so the network must propagate them to the auxiliary view through the decoder's cross-view attention, while the raw view's clean geometry keeps the scene structure anchored — content conflict becomes a training signal rather than noise. At inference the roles invert and all input views are edited, because each additional edited frame is extra evidence and the cross-view inconsistency is naturally smaller; having been trained under harsher conflict, the backbone does better in this "easier" setting. The ablation makes this concrete: switching inference back to asymmetric inputs drops CLIPt2i from 0.266 to 0.259 and raises C-FID from 171.3 to 189.7 and C-KID from 0.116 to 0.149.

3. Gaussian-center 3D regularization: anchoring the edited geometry to the original scene

Relying on 2D losses alone (CLIP + LPIPS + MSE) hides a hazard: the renders can look plausible while the underlying Gaussians have already drifted in 3D. Editing distorts the depth cues in the inputs, each view drifts in its own direction, and the result is separated depth layers and "floating" Gaussians — something 2D supervision cannot penalize, because every single view looks self-consistent. The remedy is a frozen pretrained LRM (NoPoSplat) that reconstructs reference Gaussian centers \(\mathcal{G}_{\text{ref}}\) from the unedited images, after which two complementary 3D regularizers constrain Edit3r's edited Gaussians. The first is a center-matching term, a Huber (SmoothL1) loss pulling each predicted center toward its reference location, which permits local deformation consistent with the edit but forbids wholesale drift. The second is a multi-view consistency term that randomly samples Gaussian centers per view and minimizes the pairwise Chamfer-L1 distance between them, suppressing depth-layer separation and view-dependent misalignment.

The two terms complement each other and neither substitutes for the other: center matching pins the edited scene to the base geometry but only compares against the reference view and says nothing about relative structure across views, whereas the Chamfer term governs exactly that cross-view configuration but provides no absolute anchor. The ablation confirms this is the single largest contributor — removing 3D regularization drops CLIPt2i from 0.266 to 0.237 and inflates C-FID from 171.3 to 278.4 while C-KID rises from 0.116 to 0.182, the worst degradation among all four ablation items.

4. Random drop of the reference view's Gaussians: keeping the edit style from being dominated by one view

Asymmetric inputs open a new shortcut: since the reference view already carries the edited content and appearance, the network can simply copy the reference view's Gaussians and render the auxiliary view irrelevant, overfitting to the reference viewpoint. The countermeasure is to discard, with probability 0.5, the Gaussians associated with the first (reference) input view when forming the supervision, forcing the network to actually propagate the edit semantics onto the auxiliary view's Gaussians. The drop probability has to be tuned: removing random drop lowers CLIPt2i to 0.252 and raises C-FID to 183.1 and C-KID to 0.130, confirming overfitting to the reference view, while too high a drop rate weakens the edit because too much of the semantic source is removed.

A Worked Example¶

Take an indoor video and walk it through both training and inference. On the training side, SAM2's automatic mask generator runs on frame 0 and yields candidate objects; a filter on area, stability, predicted IoU, aspect ratio, and border margin keeps a subset, each retained object receives a persistent ID and one sampled color transform, and the frame-0 boxes prompt the video segmentation predictor to propagate masks along the sequence. If some frame's active-ID set overlaps the last processed frame by less than 0.5 — the camera sweeps away and a target leaves the frame, say — that frame is dropped, and only frames passing the check proceed to recoloring, which writes the regions in by soft blending to produce cross-frame consistent reference images. The recolored frame 0 is then paired with the raw frame 1 as an asymmetric input: the backbone predicts dense Gaussians per view, concatenates them in the canonical frame, and renders them to be supervised against the recolored targets with CLIP/LPIPS/MSE, while the frozen LRM branch predicts reference centers from the unedited images and applies Huber anchoring plus a cross-view Chamfer constraint on the predicted centers; when supervision is formed, the frame-0 Gaussians are dropped with probability 0.5. On the inference side, the same scene with the prompt "Add a cactus garden" has both frames edited by IP2P and goes through Edit3r in one forward pass (about 0.5 s) to produce edited Gaussians and novel views. Here the paper highlights a behavior worth watching: when the second frame's edited pixels conflict with the first frame's, the Gaussians originating from the second view automatically lower their own opacity to defuse the conflict rather than painting the contradictory content into the scene, so the novel views synthesized from combining both inputs stay self-consistent.

Loss & Training¶

The training objective combines 2D supervision with 3D geometric constraints, summed over every view's rendered prediction \(\hat{I}_v\) and target \(I_v\):

\[\min_{\theta}\ \sum_{v=0}^{V-1}\Big[\mathcal{L}_{\text{CLIP}}(\hat{I}_v,I_v)+\mathcal{L}_{\text{LPIPS}}(\hat{I}_v,I_v)+\mathcal{L}_{\text{MSE}}(\hat{I}_v,I_v)\Big]+\lambda_1\mathcal{L}_{\text{center}}+\lambda_2\mathcal{L}_{\text{geom}}\]

(⚠️ refer to the original paper: this equation is assembled from the paper's prose, and the symbols and coefficients of the weights should be taken from the source.) The three 2D losses cover different frequency bands: the CLIP image–image loss aligns prediction and target in a shared semantic embedding space and provides a robust global signal; the VGG-based LPIPS handles mid- and high-frequency perceptual detail such as edges and textures; the low-frequency MSE enforces consistency of color, exposure, and illumination while staying tolerant of small reprojection errors. Dropping any of them unbalances the objective — MSE alone is blurry, CLIP alone loses detail. On the 3D side the center term takes the form \(\text{SmoothL1}(\hat{\mu},\mu_{\text{ref}})\), and the geometric term is the pairwise Chamfer-L1 between sampled per-view centers, normalized by \(V(V-1)\). The default configuration is SAM2 recoloring supervision, asymmetric inputs, all losses, random drop of the reference view's Gaussians at \(p=0.5\), and InstructPix2Pix as the inference front-end; all experiments run on a single NVIDIA RTX 6000, and neither training nor inference requires camera poses.

Key Experimental Results¶

Main Results¶

Evaluation covers two angles. Edit effectiveness uses CLIP image-text similarity (CLIPt2i), the absolute alignment between the edited render and the target description, where higher is better. Reconstruction quality and consistency cannot be captured by a single score, so the paper reports distribution-level C-FID and C-KID (the "C-" prefix means scene-conditioned: edited renders as a set are compared against real reference views of the same scene; lower is better), per-frame no-reference NIQE and BRISQUE plus variance-of-Laplacian sharpness, and geometric consistency through symmetric epipolar distance (SED) — SIFT correspondences between two views, the fundamental matrix from known intrinsics and poses, and SED mean/median statistics with 3-pixel inlier ratios.

Category	Method	Time (s) ↓	CLIPt2i ↑	C-FID ↓	C-KID ↓
Optimization	GaussCtrl	325.53	0.227	135.0	0.091
Optimization	EditSplat	584.46	0.241	174.1	0.122
Feed-forward	NoPoSplat	0.61	0.253	180.6	0.125
Feed-forward	Edit3r (Ours)	0.51	0.266	171.3	0.116

Edit3r is roughly 640 times faster than GaussCtrl and 1150 times faster than EditSplat while attaining the highest CLIPt2i. GaussCtrl's low C-FID and C-KID come from making only marginal updates, so its edit magnitude and text alignment suffer; EditSplat has moderate edit strength but stays overly conservative. NoPoSplat is designed for reconstruction and tends to average inconsistent multi-view evidence, which blurs its outputs and weakens every metric.

On novel-view geometric consistency (NoPoSplat → Edit3r): mean SED 30.260 → 23.752 px, median SED 3.783 → 3.595 px, the share of views with median SED below 3 px rises from 55.56% to 59.29%, and the 3-px inlier ratio from 42.54% to 47.06% — all four improve, indicating more coherent fusion of edited multi-view evidence. For no-reference quality among edited outputs, Edit3r gets the best NIQE (3.402) and BRISQUE (26.538) and remains competitive in sharpness (314.902); for reference, the original scene scores NIQE 2.982 / BRISQUE 15.402 / sharpness 1506.719, EditSplat 3.439 / 29.588 / 327.252, NoPoSplat 3.885 / 27.659 / 251.997, and GaussCtrl 7.081 / 66.384 / 13.718. Qualitatively, EditSplat partially succeeds but also alters regions that should stay unchanged (sky and ground); GaussCtrl fails almost entirely on scenes unseen during its training; NoPoSplat's quality is highly sensitive to cross-view inconsistency — blurry when the conflict is large, much better when the inputs are consistent; Edit3r is stable across all four scenes.

Ablation Study¶

Config	CLIPt2i ↑	C-FID ↓	C-KID ↓	Note
Edit3r (full)	0.266	171.3	0.116	Default: recoloring supervision + asymmetric input + full losses + R-Drop + IP2P
w/o Recolor	0.243	215.0	0.141	Trained directly on multi-view images from the 2D editor
w/o 3D Loss	0.237	278.4	0.182	Removes center anchoring and cross-view Chamfer
w/o SAM	0.248	179.6	0.127	Disables SAM-mask augmentation, plain recoloring only
w/o R-Drop	0.252	183.1	0.130	Removes random drop of the reference view's Gaussians

Config	CLIPt2i ↑	C-FID ↓	C-KID ↓	Note
IP2P	0.266	171.3	0.116	Default front-end; lowest perceptual quality, reliable localization
GPT-Image-1	0.261	166.2	0.102	Precise localization, but adds global artifacts
Gemini-2.5-Flash-Image	0.246	150.3	0.098	Lowest text alignment, best distribution-level metrics
FLUX.1 Kontext	0.276	169.9	0.112	Best visual fidelity and highest CLIPt2i, but frequently mislocalizes
Asymmetric inputs at inference	0.259	189.7	0.149	Only the reference view edited; less evidence, harder setting

Key Findings¶

3D geometric regularization contributes the most: removing it worsens C-FID from 171.3 to 278.4 (+62%) and drops CLIPt2i from 0.266 to 0.237, the largest degradation of any ablation item. This supports the authors' diagnosis — with 2D losses only, renders can look reasonable while Gaussian centers have already drifted in 3D.
Recoloring supervision is irreplaceable: training on real 2D edited images instead moves C-FID from 171.3 to 215.0 and C-KID from 0.116 to 0.141. Inconsistent labels act as noise and hinder convergence, and this negative result is one of the paper's most valuable pieces of evidence.
SAM augmentation and random drop are smaller but necessary: removing them lowers CLIPt2i to 0.248 and 0.252 respectively, less than the two items above, but each targets a specific failure mode — the distribution gap between training and inference for the former, overfitting to the reference viewpoint for the latter — so both are kept.
Editing more views at inference helps: switching inference inputs from asymmetric to all-edited moves CLIPt2i from 0.259 to 0.266 and C-FID from 189.7 to 171.3. This is a counter-intuitive by-product of asymmetric training — manufacturing conflict during training pays off in an "easier" setting at inference.
The method is decoupled from the editor: swapping in FLUX, GPT, or Gemini leaves inference speed and cross-view coherence unchanged, showing the reconstruction backbone is agnostic to the 2D front-end and scales with future 2D synthesis progress. Each editor has its own weakness — FLUX preserves fidelity but drifts spatially, GPT and Gemini localize well but add global artifacts, IP2P has the lowest perceptual quality — and that trade-off table is itself a practical guide for choosing a front-end.

Highlights & Insights¶

A controllable surrogate sidesteps the missing-ground-truth dead end: the paper's "aha" moment is that after admitting no cross-view consistent edited ground truth exists, it does not fabricate pseudo-targets but switches to a task that is consistent by construction (one color transform per object, reused across frames) to train fusion, betting that the capability transfers to real edits. The ablations and the cross-editor results show the bet pays off.
Treating inconsistency as a training signal rather than noise: the asymmetric input design deliberately gives the network a harder problem in training and an easier one at inference, inverting the usual "match train and test distributions" intuition — yet it works for tasks that must fuse conflicting observations into one coherent output. The same idea transfers to any feed-forward model facing contradictory multi-source evidence, such as multi-view 3D detection, cross-modal fusion, or video denoising.
A frozen model as a 3D teacher: extracting reference Gaussian centers from unedited images with a frozen NoPoSplat turns "the edited geometry must not drift" into a differentiable center-matching loss, which is far more precise than pixel-level regularization. Using a same-family pretrained model as a geometric anchor is cheap to reuse in other generation tasks that must preserve underlying geometry.
Conflict resolution emerges without explicit supervision: the paper shows a behavior no loss term asks for — when two views' edits conflict, the Gaussians from the conflicting view lower their own opacity instead of painting the contradiction in. Cross-view attention plus canonical-frame concatenation evidently yields a soft confidence mechanism on their own, and adding uncertainty-aware rendering (which the authors list as future work) should make it explicit.

Limitations & Future Work¶

Recoloring supervision cannot cover large geometric changes: recoloring is an appearance-level operation, whereas Add / Remove edits that genuinely restructure the scene (inserting a bulky new object, or deleting one that covers a large region) and extreme material or illumination shifts fall outside what recoloring can supervise. The authors acknowledge this and propose extending supervision with view-consistent generative augmentation.
The train/inference distribution gap is mitigated, not closed: SAM-mask augmentation, background leakage, and color transforms all push recolored images toward real edited images, but the structural difference — an editor conjuring new content in a local region — cannot be patched by color augmentation. The foreseeable failure mode is that when the editor's new content grossly disagrees with the original geometry, the network can only suppress it by lowering opacity, so those regions end up missing rather than correctly generated.
Edit granularity is per-view, not per-object: edit proposals come from a 2D editor on individual frames and the network only fuses them; there is no explicit per-object disentangled control. Precise 3D edits ("change this object, don't touch the one beside it") would need the per-object disentangled control the authors mention.
Evaluation scale and scene variety are limited: DL3DV-Edit-Bench has only 20 scenes and 100 instructions, evaluated with a fixed editor and a fixed random seed; dynamic scenes, long sequences, and heavier occlusion are not covered.
Dependence on 2D editor quality: since semantic content is generated entirely by the front-end, mislocalization (FLUX) or global artifacts (GPT/Gemini) propagate straight into the 3D result; the backbone can only limit the damage through opacity suppression and cannot correct semantic errors themselves.

vs Instruct-NeRF2NeRF / EditSplat / GaussCtrl (per-scene optimization): they run reconstruct–edit–refit and need a pre-reconstructed 3DGS, posed images, and hundreds of seconds of iterative fitting, during which inconsistency is baked into the representation. This paper does no test-time optimization at all, produces results in 0.51 s in one forward pass, and tackles cross-view conflict head-on. The price is that it consumes only two sparse unposed views and its semantic ceiling is bounded by the 2D editor, whereas they can in principle optimize a scene more thoroughly. Fairness caveat: the input configurations differ (two unposed frames here versus their full calibrated image set plus a pre-reconstructed scene), which also accounts for part of the order-of-magnitude speed gap.
vs NoPoSplat (feed-forward reconstruction): this paper adopts its pose-free backbone directly, but NoPoSplat averages inconsistent evidence under edited inputs and blurs; Edit3r trains fusion through asymmetric inputs and beats it across all SED and 3-px inlier metrics, making it the primary ablation reference.
vs video editing methods (TokenFlow, Video-P2P, etc.): they maintain temporal coherence through attention or token propagation but remain fundamentally 2D, lacking explicit 3D reasoning, and become view-inconsistent once lifted to 3D supervision; this paper resolves consistency in the canonical Gaussian space.
vs instant 3D editing work such as InstaInpaint: both pursue optimization-free one-pass editing, but this paper differs by requiring neither poses nor a pre-reconstruction, and by confronting the question of where training supervision comes from as its central problem.

Rating¶

Novelty: ⭐⭐⭐⭐ Reframing editing as "recoloring pretext + asymmetric input" for feed-forward fusion is clean, and feed-forward 3D editing was indeed unexplored; the backbone itself, however, is inherited from NoPoSplat.
Experimental Thoroughness: ⭐⭐⭐⭐ A self-built benchmark, four ablation groups spanning supervision/losses/augmentation/inference settings, and four 2D editors to verify decoupling; but the benchmark has only 20 scenes and 100 instructions, with no user study or human edit-quality rating.
Writing Quality: ⭐⭐⭐⭐ The motivation chain is clear and the negative result (training on edited images hurts) is reported honestly; several formulas are corrupted in the cached text.
Value: ⭐⭐⭐⭐ Half-second, pose-free 3D editing matters directly for interactive content creation and AR/VR, and both "manufacture conflict in training, enjoy an easier setting at inference" and "use a controllable surrogate to sidestep missing ground truth" are reasonably transferable lessons.