acceptodds
Under review as a conference paper at ICLR 2027

STAGE: Subspace-Targeted Affine Generative Erasure for Text-to-3D Models

Abstract

Concept erasure suppresses a target concept while preserving behavior on unrelated inputs. Existing closed-form methods were designed for 2D image diffusion and assume a single generative pathway, so one edit must cover geometry and texture at once. Native 3D generators, which synthesize structured 3D representations directly rather than by lifting 2D samples, violate this assumption. We show that shape and object concepts must be erased in the structural stage of the pipeline and material concepts in the appearance stage. We therefore formulate erasure in native text-to-3D as a stage-aware editing problem and introduce STAGE, a training-free, closed-form framework. STAGE confines each edit to the low-dimensional subspace spanned by the differences between erase and anchor embeddings, and relaxes the norm-preserving (orthogonal) constraint of prior editors into a least-squares affine correction that maps target activations onto safe anchors subject to a penalty on the displacement of retained prompts. The correction applies to the structural stage, the appearance stage, or both. We find that the stage an edit must reach is determined by concept type. On TRELLIS, the standard open native 3D generator, across 15 shape, material, and object concepts, STAGE reaches 66.7 on a composite score that balances forgetting the target concept against preserving everything else, aggregating CLIP-based semantic and physical metrics, versus 53.2 for the strongest adapted baseline.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Unlearning, Diffiusion Models, 3D

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2ZwgOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies concept erasure for native text-to-3D generative models. The authors argue that existing closed-form concept editing methods, originally developed for 2D diffusion models, overlook the multi-stage nature of native 3D generation, where geometry and appearance are produced by different stages. Based on this observation, the paper formulates concept erasure as a stage-aware editing problem and introduces STAGE, a training-free closed-form editor.

Critical strengths
  1. The motivation is clear and specific to native 3D generation. In TRELLIS, the structural and appearance stages play different roles, so applying the same edit indiscriminately across the pipeline may either fail to remove the target concept or introduce unnecessary collateral changes.
  2. The stage-locality claim is supported by direct experiments. The results directly support the central claim that different concept types depend on different stages of the native 3D generation pipeline.
Critical weaknesses
  1. Since the core stage-aware argument is closely tied to TRELLIS's explicit structural/appearance decomposition, it is unclear how well the proposed formulation generalizes to other native 3D generators with different architectures or representations. A second backbone, even with a smaller-scale evaluation, would substantially strengthen the generality of the claim.
  2. The quantitative comparison focuses mainly on UCE and OCE, both of which are closed-form editors originally developed for 2D generation. Adding a 3D baseline for comparison would be more convincing
  3. The preservation component is mainly computed using the remaining concepts within the same five-class semantic axis (as shwon in Table 4). For example, when erasing one object concept, preservation is evaluated primarily on the other object categories in that axis. This is useful, but it does not fully establish that the edited model preserves its general behavior on a broad distribution of unrelated prompts. A larger retain benchmark covering diverse categories and compositional prompts would provide stronger evidence against collateral degradation.
  4. The current stage assignment relies on predefined concept categories, rather than being automatically inferred by the model or data. It is still unclear how to determine which generation stage should be edited when the concept to be erased involves geometry, appearance, material, or object identity simultaneously.
Questions for authors

Same as weakness

5 comments
AuthorsOpenReview ↗

We thank the reviewer for the review and for noting that the stage-locality claim is supported by direct experiments. We address the concerns below.

W1: Second backbone

We ran the full benchmark on a second generator, DiffSplat [1], which repurposes Stable Diffusion 1.5 to generate 3D Gaussian splats as multi-view splat images. Its architecture and representation differ from those of TRELLIS. All three editors use the same concepts, prompts, seeds, evaluators, and hyperparameters as on TRELLIS. 3D-US on DiffSplat (higher is better):

MethodShapeMaterialObjectOverall
UCE70.770.065.868.8
OCE58.749.751.753.2
STAGE84.784.278.682.5

STAGE has the highest 3D-US on every axis and overall, with a margin of 13.7 points over the strongest baseline (13.5 on TRELLIS).

In the revised paper, this is in Section 4 (paragraph "Generalization to a Second Generator", Table 3a), with per-concept results in Appendix J (Table 18) and the generated assets in Appendix L.

W2: A 3D baseline

We are not aware of an erasure method for native text-to-3D generators. The only prior 3D unlearning method [2] fine-tunes the 2D multi-view diffusion model at the front of an image-to-3D pipeline, and has no counterpart in a native generator whose stages read the text. We therefore adapted the strongest closed-form editors, UCE [3] and OCE [4], under identical conditions: the same prompts and seeds, the same stages and layers, and the same one-at-a-time hyperparameter sweep, with each method at its best configuration. 3D-US on TRELLIS:

MethodShapeMaterialObjectOverall
UCE66.930.851.947.5
OCE67.941.853.253.2
STAGE77.756.467.766.7

STAGE also scores highest of the three editors under every edit mode that erases a concept (table in W4).

In the revised paper, the absence of a native 3D baseline is stated in Section 2 (second paragraph), and the adaptation of the baselines is in Appendix B.

W3: Preservation on a broad retain set

We added a retain set of 60 held-out TRELLIS-500K captions. They are not in the caption sample used for K_0, and they name none of the 15 concepts or their synonyms. They describe objects, characters, and scenes from many categories, each with several attributes, for example "Rusty container with blue lid window and a red triangle.". Each caption is generated with the same seed before and after every one of the 15 edits of Table 1, for every editor.

UCEOCESTAGE
Prompt alignment kept after the edit96.7%87.0%95.8%
CLIP image similarity, before vs. after0.8510.7880.840
Mean forgetting F on the 15 concepts0.470.640.68

STAGE preserves unrelated prompts at the level of UCE and above OCE, while forgetting the target concepts more than both.

In the revised paper, this is in Section 4 (paragraph "Preservation on Unrelated Prompts", Table 3e), with results per axis in Appendix I (Table 17).

W4: Stage assignment for concepts that combine several attributes

The per-axis assignment follows the architectural role of each stage and is tested directly by the mode crossover. 3D-US on TRELLIS:

MethodAxisMode-SMode-LMode-SL
STAGEShape77.70.077.2
Material31.456.472.9
Object66.72.767.7
OCEShape67.93.056.5
Material27.541.840.7
Object62.42.753.2
UCEShape66.90.056.3
Material24.730.848.4
Object45.60.051.9

For a concept that combines geometry, appearance, material, and identity, or whose type is unknown, Mode-SL edits every stage that reads the text and requires no inference about the concept type. Applied to all three axes, Mode-SL gives STAGE an overall 3D-US of 72.5, compared with 49.6 for OCE and 52.1 for UCE.

In the revised paper, this is in Section 4 (paragraph "Mode Assignment and Stage Locality", Table 2).

References

[1] Lin, Chenguo, et al. "DiffSplat: Repurposing image diffusion models for scalable Gaussian splat generation." ICLR'2025.

[2] Wang, Shixuan, et al. "Machine unlearning in 3D generation: A perspective-coherent acceleration framework." NeurIPS'2025.

[3] Gandikota, Rohit, et al. "Unified concept editing in diffusion models." WACV'2024.

[4] Sun, Yuhao, et al. "Orthogonal concept erasure for diffusion models." arXiv'2026.

Thank you for the detailed response and additional experiments. Most of my concerns have been adequately addressed. I have one remaining question regarding W4: Since Mode-SL achieves comparable or better performance across all three axes, what are the advantages of concept-dependent stage assignment over universally applying Mode-SL, particularly in terms of preservation and efficiency?

AuthorsOpenReview ↗

Response on concept-dependent stage assignment

We thank the reviewer for the follow-up. Mode-SL is a suitable default when the concept type is unknown. When the type is known, the concept-dependent assignment gives better preservation at nearly the same cost.

Preservation

Results of STAGE on TRELLIS. The last column is the fraction of assets whose attribute outside the edited axis keeps its top-1 label (material by CLIP / MINC-23 after shape edits, shape by CLIP after material edits).

AxisModeFP3D-USOther attribute kept
ShapeMode-S (assigned)0.700.9177.70.68 / 0.66
ShapeMode-SL0.710.8977.20.67 / 0.66
MaterialMode-L (assigned)0.450.9256.40.92
MaterialMode-SL0.720.8372.90.76

The object axis is assigned Mode-SL, so the two settings coincide there. On shapes, Mode-SL adds no 3D-US over Mode-S and lowers preservation. On materials, it raises forgetting at the cost of preservation, on the other concepts of the axis and on the shape of the edited assets. Mode-L also leaves the weights of the structural stage unchanged, so the same seed gives the same voxel support.

All three editors preserve the other concepts of the axis better under the assigned mode. Preservation P:

AxisEditorAssigned modeMode-SL
ShapeSTAGE0.910.89
ShapeUCE0.850.81
ShapeOCE0.670.45
MaterialSTAGE0.920.83
MaterialUCE0.840.71
MaterialOCE0.740.35

P is higher under the assigned mode in 29 of the 30 pairs of editor and shape or material concept.

Efficiency

The two options have nearly the same cost. Both use one closed-form solve and add no cost at generation time, and the assigned mode folds the correction into 28 cross-attention layers instead of 56. The advantage of the assigned mode is preservation.

Choice of mode

The assigned mode is the choice when the concept type is known and preserving the rest of the model matters most. Mode-SL is the choice when the type is unknown or when forgetting matters most. STAGE has the highest 3D-US of the three editors under both options, on every axis.

In the revised paper, the mode comparison is in Section 4 (paragraph "Mode Assignment and Stage Locality", Table 2), and the cross-attribute evaluation is in Appendix F (Table 14).

Reviewer_946rOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses concept erasure in native text-to-3D generation. The authors observe that, in TRELLIS, different concept types rely on different parts of the generation pipeline, and formulate erasure as a stage-aware model-editing problem. They propose STAGE, a training-free closed-form method that builds a low-dimensional contrast subspace from erase-anchor embedding differences, fits an affine redirection while penalizing changes to retained prompts, and folds the resulting correction into the cross-attention layers of the structural stage, appearance stage, or both depending on the concept type. The method is evaluated on TRELLIS using 15 shape, material, and object concepts and is compared against adapted UCE and OCE baselines.

Critical strengths
  1. The paper addresses an underexplored problem of concept erasure in native text-to-3D generation, and the proposed method is reasonably motivated by the staged structure of TRELLIS, using concept-dependent placement of the closed-form edit across the structural and appearance stages.
  2. The paper is generally clear and well organized.
  3. STAGE shows strong empirical performance, achieving the best overall 3D-US among the compared methods and outperforming the baselines on most evaluated concepts.
Critical weaknesses
  1. The paper lacks evaluation of simultaneous multi-concept erasure. Both UCE and OCE have been evaluated for simultaneous multi-concept erasure, whereas STAGE is only tested on one concept at a time. This is particularly relevant when multiple target concepts require edits to the same generation stage, since STAGE constructs its contrast subspace and affine correction from erase-anchor embedding differences, but the paper does not evaluate how such edits interact when applied jointly. It is therefore unclear whether STAGE remains effective and preserves unrelated concepts as the number of simultaneously erased concepts increases.
  2. The stage-specific editing setup does not guarantee attribute-specific preservation. Since the appearance stage is conditioned on the structure generated by the structural stage, editing G_S can change the final appearance even if G_L itself is left unchanged. This is also visible in the shape example of Fig. 4, where the appearance changes substantially after the shape edit. Therefore, the claims that Mode-S leaves appearance untouched and Mode-L leaves structure untouched (L166–169) are not sufficiently supported without direct cross-attribute preservation evaluation.
  3. The object-identity evaluation is relatively narrow. The paper evaluates only five coarse object categories—car, cat, chair, table, and teddy bear—even though object identity is the main setting used to motivate editing both the structural and appearance stages. This limited coverage makes it difficult to assess whether the observed benefit of Mode-SL generalizes to a broader range of object concepts.
Questions for authors
  1. How would STAGE be extended to simultaneous multi-concept erasure, particularly when multiple target concepts are assigned to the same generation stage?
  2. How do the authors justify the claim that Mode-S leaves appearance/material untouched, given that the appearance stage is conditioned on the structure produced by the structural stage?
  3. Have the authors evaluated additional object-identity categories beyond car, cat, chair, table, and teddy bear? If so, do the results support the same Mode-SL conclusion?
7 comments
AuthorsOpenReview ↗

We thank the reviewer for the detailed review and the positive assessment of the clarity and results of the paper. We address the concerns below.

W1/Q1: Simultaneous multi-concept erasure

The formulation of Section 3 is defined for a set of concepts C_erase = {c_1, ..., c_k}. Each concept is assigned to the stages of its mode. Every edited stage receives one correction, fitted on the stacked erase and anchor pairs of all concepts assigned to it, with the anchors of every concept and the null prompt in the retained set. This keeps one small linear solve per edited stage, independent of the number of concepts.

We evaluated this setting with two and three concepts that share a stage, for all three editors. 3D-US is the mean over the erased concepts, and P is computed on the concepts of the axis that are not erased.

Erased jointlyUCE 3D-USOCE 3D-USSTAGE 3D-USSTAGE P
cube, cellular lattice84.063.286.30.98
cube, cellular lattice, cylinder67.666.771.00.85
wood, glass23.124.653.50.95
wood, glass, stone40.340.147.50.98
car, chair39.132.474.20.82

STAGE has the highest 3D-US in all five settings, and its preservation of the remaining concepts stays at or above 0.82.

In the revised paper, the formulation is in Section 3 (paragraph "Erasing Several Concepts"), the results are in Section 4 (paragraph "Concept Coverage, Joint Erasure, and Cross-Attribute Effects", Table 3c), and the full results are in Appendix G (Table 15).

W2/Q2: Mode-S and Mode-L and cross-attribute preservation

We revised the wording. The modes now specify which weights change. Mode-S folds the correction only into the layers of G_S, "leaving the weights of G_L unchanged", and Mode-L folds it only into the layers of G_L, "leaving the weights of G_S, and therefore the sampled structure, unchanged". Mode-L cannot alter the coarse structure, since the unedited G_S samples the same voxel support for the same seed. Under Mode-S, G_L generates appearance on the edited structure.

We added the requested direct evaluation. For every shape and material edit, we predict the attribute of the other axis on the 150 evaluation assets of the edited axis before and after the edit, and report the fraction of assets whose top-1 label is unchanged (higher is better):

EditProbed attribute (evaluator)UCEOCESTAGE
Shape, Mode-Smaterial (CLIP)0.610.550.68
Shape, Mode-Smaterial (MINC-23)0.640.580.66
Shape, Mode-SLmaterial (CLIP)0.620.440.67
Shape, Mode-SLmaterial (MINC-23)0.650.540.66
Material, Mode-Lshape (CLIP)0.900.870.92
Material, Mode-SLshape (CLIP)0.750.580.76

STAGE preserves the other attribute best of the three editors in every setting.

In the revised paper, the wording is revised in Section 3 (paragraph "Stage Placement") and in the caption of Figure 1. The evaluation is in Section 4 (paragraph "Concept Coverage, Joint Erasure, and Cross-Attribute Effects", Table 3d), with all modes in Appendix F (Table 14).

W3/Q3: Additional object categories

We added five object concepts, airplane, boat, guitar, lamp, and shoe, each anchored to the next one in this list and shoe to airplane. Their erase and anchor prompts are built like those of the object axis, and every evaluation prompt is fixed to a seed on which both evaluators recognize the concept on the unedited model, as in the paper. All editors use Mode-SL, the assignment of the object axis, and their configurations of Table 1. 3D-US (higher is better):

ConceptUCEOCESTAGE
Airplane73.152.292.8
Boat79.847.259.6
Guitar56.042.887.5
Lamp13.415.046.1
Shoe45.347.980.9
Mean53.541.073.4

STAGE has the highest 3D-US on four of the five new concepts and a mean 19.9 points above the strongest baseline.

In the revised paper, this is in Section 4 (paragraph "Concept Coverage, Joint Erasure, and Cross-Attribute Effects", Table 3b), with forgetting and preservation per concept in Appendix H (Table 16).

Thank you for your response. For the ten object categories, STAGE performs worse than the strongest baseline on cat, teddy bear, and boat by 35.1, 31.6, and 20.2 points in 3D-US, respectively, which raises concerns about the sensitivity of the average gains to category selection. Therefore, I am not sure whether STAGE’s advantage can be maintained across a broader range of object categories.

AuthorsOpenReview ↗

Response on the dependence on category selection

Sensitivity of the object results to category selection

Over all ten object categories of the paper (Table 1 and the five added concepts), 3D-US:

UCEOCESTAGE
Mean52.747.170.6
Median58.549.280.2
Categories with the highest 3D-US307

To measure the dependence on category selection directly, we compared the mean 3D-US of the three editors on every subset of the ten categories and on 10,000 bootstrap resamples of the categories. Share of cases in which each editor has the highest mean:

Categories comparedUCEOCESTAGE
Every subset of 5 categories (252)6.0%0.0%94.0%
Every subset of 6 categories (210)2.9%0.0%97.1%
Every subset of 7 or more categories (176)0.0%0.0%100%
Bootstrap resamples of the 10 categories (10,000)5.2%0.0%94.8%

The advantage of STAGE on objects therefore does not depend on which categories are selected. Removing any three of the ten categories, including cat, teddy bear, and boat, leaves STAGE with the highest mean. On a second generator, DiffSplat, STAGE also has the highest object-axis 3D-US (78.6, compared with 65.8 for UCE and 51.7 for OCE).

Mapping the erased concept onto its anchor

STAGE is designed to map an erased concept onto a safe anchor rather than only suppress it (abstract and Section 3). We measure this with the redirect rate, the share of edited assets of the erased category that an evaluator assigns to its anchor (chance level 20%). Over the ten object categories:

UCEOCESTAGE
Redirect rate, CLIP31.7%32.7%85.7%
Redirect rate, Uni3D34.0%33.7%74.0%
Categories with the highest redirect rate0010

STAGE replaces the erased object with its anchor most often on every one of the ten categories (mean of the two evaluators), including the three mentioned by the reviewer:

Redirect rate (mean of CLIP and Uni3D)UCEOCESTAGE
cat → chair21.7%36.7%58.3%
teddy bear → table90.0%58.3%91.7%
boat → guitar15.0%18.3%61.7%

In the revised paper, this analysis is in Appendix H (Table 16), and the redirect rates on the axes of Table 1 are in Appendix D (Table 9).

Reviewer_K6XuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes new concept of attribute unlearning on text to 3d model

Critical strengths

The task itself is new and challenging.

Critical weaknesses

Although the proposed task is quite novel concept and interesting, there is a critical pitfall on this text to 3D generation based task.

There is only one available text to 3D model which is TRELLIS and this method is limited on that framework. Since recent 3D generation models use two staged pipeline such as text to image and image to 3D (if the shape and texture generation is unified representation such as TRELLIS) , or text to multiview and recon multiview to 3D (LGM)

I am curious whether this method can be applied to those methods which has different framework.

This current method is limited on specific model with sparse representation so it requires more experiments on whether this method can be robust on generation framework.

Questions for authors

See weakness

4 comments
AuthorsOpenReview ↗

We thank the reviewer for recognizing the task as new and challenging. We address the concerns below.

W1: The method is limited to TRELLIS and its sparse representation

STAGE does not use the 3D representation. It edits only the key and value projections of the cross-attention layers that read the text embedding, and its correction is computed from text embeddings alone. The 3D representation and its decoders stay frozen. The method therefore applies to any generator whose stages read the text through cross-attention.

To test this, we ran the full benchmark on a second generator, DiffSplat [1]. DiffSplat repurposes the image diffusion model Stable Diffusion 1.5 to generate 3D Gaussian splats as multi-view splat images, so both its architecture and its representation differ from those of TRELLIS. All three editors use the same concepts, prompts, seeds, evaluators, and hyperparameters as on TRELLIS.

3D-US on DiffSplat (higher is better):

MethodShapeMaterialObjectOverall
UCE70.770.065.868.8
OCE58.749.751.753.2
STAGE84.784.278.682.5

STAGE has the highest 3D-US on every axis and overall. Its margin over the strongest baseline is 13.7 points, compared with 13.5 on TRELLIS (66.7 vs 53.2).

In the revised paper, this is in Section 4 (paragraph "Generalization to a Second Generator", Table 3a), with per-concept results in Appendix J (Table 18) and the generated assets in Appendix L. The independence from the 3D representation is stated in Section 3 (paragraph "Preliminaries and Problem Formulation").

W2: Pipelines that go from text to image and then from image to 3D

In such a pipeline the 3D model is conditioned only on an image and never reads the text. A text-specified concept can enter only through the text-to-image model. Erasing it there is erasure in a 2D diffusion model, and STAGE applies to that model as an edit of its cross-attention layers.

In the revised paper, this is discussed in Section 2 (third paragraph).

W3: Pipelines that go from text to multiple views and then reconstruct 3D (e.g., LGM)

Here the text enters only the multi-view diffusion model (e.g., MVDream [2]), whose denoiser reads it through cross-attention, and the reconstructor (e.g., LGM [3]) receives only the generated views. STAGE folds into those cross-attention layers in the same way. DiffSplat belongs to this family, since it builds on a text-to-image diffusion backbone that produces several views, and the table in W1 shows that the edit transfers to it.

In native generators, the setting of the paper, the text conditions more than one stage, and the stage that receives the edit decides whether a concept is erased at all. 3D-US of STAGE on TRELLIS:

Edited stagesShapeMaterialObject
Structural stage only (Mode-S)77.731.466.7
Appearance stage only (Mode-L)0.056.42.7
Both stages (Mode-SL)77.272.967.7

STAGE has the highest 3D-US of the three editors under every edit mode that erases a concept.

In the revised paper, cascaded pipelines are discussed in Section 2 (third paragraph), and the stage analysis is in Section 4 (paragraph "Mode Assignment and Stage Locality", Table 2).

References

[1] Lin, Chenguo, et al. "DiffSplat: Repurposing image diffusion models for scalable Gaussian splat generation." ICLR'2025.

[2] Shi, Yichun, et al. "MVDream: Multi-view diffusion for 3D generation." ICLR'2024.

[3] Tang, Jiaxiang, et al. "LGM: Large multi-view Gaussian model for high-resolution 3D content creation." ECCV'2024.

Although the rebuttal has addressed my concerns partially, I still have major concerns on the proposed method.

As I commented in the early review, recent Text to 3D method for high-end mesh generation can be done with text-image-3D pipeline (Hunyuan, TRELLIS2, etc..) Since the content itself is decided with text to image stage, there are numerous alternative methods of text to image based methods (editing, disentangled generation, concept erasure .... etc) which can be directly combined with image to 3d pipeline to make high quality mesh with disentangled concept change.

I think task is interesting but I still don't think this 'method' is 'unique' and application to Diffsplat cannot justify the method robustness because diffsplat is a taylored text-to-gaussian method which is unified pipeline.

AuthorsOpenReview ↗

Response to the remaining concerns

1. Text-to-image-to-3D pipelines

Image-conditioned generators such as TRELLIS.2 and Hunyuan3D 2.x obtain text-to-3D by placing a text-to-image model in front of them. In such a pipeline, a text-specified concept enters only through the text-to-image model, and STAGE applies there as an edit of its cross-attention layers. We evaluated this case directly on the standard object-erasure benchmark of Stable Diffusion v1.4, erasing each of the ten CIFAR-10 classes in turn under the protocol of UnGuide [1]. Average over the ten classes (%):

MethodAcc_e ↓Acc_s ↑Acc_g ↓H_o ↑
ESD-u [2]18.2786.7616.2683.69
UCE [3]13.5498.4523.1885.48
OCE [4]10.1198.7721.8986.95
MACE [5]8.4997.3510.5392.61
UnGuide [1]6.5498.657.6794.77
STAGE5.8698.668.2894.56

STAGE has the lowest accuracy on the erased class, the highest H_o of the closed-form editors, and is on par with the strongest method of the benchmark, without any training. In a cascaded pipeline, STAGE is therefore a competitive editor for the stage that reads the text.

An edit at the front of a cascade leaves the image-to-3D model unchanged, and that model still generates the concept from any image that contains it. In a native text-to-3D generator, the text is read by the 3D stages themselves, and the edit has to be made inside the 3D generator. This is the setting of the paper, and it requires a decision that 2D editors do not make, namely which stage receives the edit. With the same editor on TRELLIS, editing only the appearance stage gives a 3D-US of 0.0 on shape concepts and 2.7 on object concepts, editing only the structural stage gives 31.4 on material concepts, and the stage assignment of STAGE gives 77.7, 56.4, and 67.7 (Table 2).

2. DiffSplat

The original concern was that the method is limited to a "specific model with sparse representation". DiffSplat differs from TRELLIS in each component the method could depend on:

TRELLISDiffSplat
Backboneflow transformersStable Diffusion 1.5 U-Net
Representationsparse voxel latentsmulti-view splat images
Stages that read the texttwoone

Without any retuning, STAGE keeps a nearly identical margin over the strongest baseline, 13.7 points on DiffSplat and 13.5 on TRELLIS. The denoiser of DiffSplat is a text-to-image diffusion backbone that reads the text through cross-attention, the same type of component that is edited at the front of a cascaded or multi-view pipeline. Together with the Stable Diffusion results above, this shows that STAGE transfers to the pipelines raised in point 1.

In the revised paper, the text-to-image results are in Appendix M (Table 19), and DiffSplat is in Section 4 (paragraph "Generalization to a Second Generator", Table 3a) and Appendices J and L.

References

[1] Polowczyk, Agnieszka, et al. "UnGuide: Learning to forget with LoRA-guided diffusion models." arXiv'2025.

[2] Gandikota, Rohit, et al. "Erasing concepts from diffusion models." ICCV'2023.

[3] Gandikota, Rohit, et al. "Unified concept editing in diffusion models." WACV'2024.

[4] Sun, Yuhao, et al. "Orthogonal concept erasure for diffusion models." arXiv'2026.

[5] Lu, Shilin, et al. "MACE: Mass concept erasure in diffusion models." CVPR'2024.

Thank you for the additional reply. Although the rebuttal has addressed partial concerns, there is still critical pitfalls remaining. The paper should have targetet on general unlearning or concept erasure for image generation and 3D generation if it should show the erasure on stable diffusion. Still the target task : text to 3d concept erasure is ambiguous.

Still there are so many possible alternative approaches such as using state of the are generative editing method (Nano Banana, QWEN Image) to erase the concept from image then generate 3D. If we think about many recent commercial models, this concept erasure on text to 3D is limited on its task.

I'll leave the decision to AC, as this is on the borderline.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.