acceptodds
Under review as a conference paper at ICLR 2027

DreamBridge: Bridged Cross-Modal Distillation for Text-to-3D Generation

Abstract

Existing text-to-3D generation models can produce semantically consistent 3D ob- jects from natural-language prompts, yet they still struggle to capture the geometry and spatial layout of complex objects. Common approaches either train native 3D generators on large paired datasets or combine text-to-image synthesis with image- to-3D generation at inference time. Inspired by the latter, we propose DreamBridge, a bridged cross-modal distillation framework that transfers sparse-structure priors from a frozen image-to-3D teacher to a text-to-3D student, where synthesized im- ages bridge the conditioning modalities during training. DreamBridge comprises three components: (i) Bridged Triplets Synthesis, which constructs triplets of a prompt, a filtered FLUX-generated image, and a cached teacher sparse-structure la- tent; (ii) Native Latent Distillation, which adapts only the student’s sparse-structure flow with LoRA through conditional flow matching on cached teacher latents; and (iii) Anchor Replay, which supplements teacher supervision with structural samples from the frozen original text model. The overall adaptation procedure uses model-generated supervision without additional ground-truth 3D training data and retains text-only inference. On BridgeStruct-1.5K, our evaluation suite of 1,536 held-out prompts, DreamBridge improves structural agreement with the teacher over the pretrained baseline and achieves the highest reported text–3D semantic alignment scores among the evaluated native text-to-3D models.

Primary area:
generative models
Keywords:
Text-to-3D Generation, Image-to-3D Generation, Cross-Modal Knowledge Distillation, 3D Latent Distillation, Conditional Flow Matching, Structured 3D Latents, Semantic Preservation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KT6hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DreamBridge transfers structural knowledge from an image-to-3D teacher to a native text-to-3D student using synthetic supervision. The method generates images from text prompts with FLUX, extracts sparse-structure latents with a frozen TRELLIS-image teacher, and adapts the sparse-structure flow of TRELLIS-text using LoRA and standard conditional flow matching. It also incorporates samples from the original text model through Anchor Replay. The resulting model retains text-only inference and improves teacher-referenced structural agreement on BridgeStruct-1.5K, with additional evaluations on Toys4K, GSO, and human preferences.

Critical strengths

The shared sparse-structure interface provides a practical way to transfer image-conditioned structural knowledge while reusing the frozen downstream generation modules and retaining text-only inference. The teacher-referenced structural improvements are supported by prompt-level and split-wise analyses: SS F1 increases from 0.1512 to 0.2069, with improvements on 81.8% of valid paired prompts. The paper provides detailed descriptions of data construction, image filtering, sampling, validity handling, and evaluation protocols, and acknowledges important limitations of its results.

Critical weaknesses

The advantage over a text-to-image-to-3D pipeline is insufficiently established. High-quality text-driven 3D generation can already be achieved by combining text-to-image and image-to-3D models. Although the paper includes Teacher-II as a reference, it does not establish a compelling quality–inference-cost tradeoff relative to the complete cascade. More importantly, text can specify back-facing, internal, and occluded components that are not visible in a single image. Since DreamBridge obtains supervision through a single synthesized image, these constraints may be lost. Targeted evaluations are needed to demonstrate that the method strengthens text-based control over complete 3D structure rather than primarily reproducing the image teacher’s preferences.

The qualitative comparisons should more directly establish fidelity to ground-truth geometry and the complete cascade. For examples with available ground-truth assets, the paper should show GT meshes, BASE, DreamBridge, and the complete text-to-image-to-3D pipeline side by side under consistent viewpoints, including views that expose fine components, openings, and occluded regions. The Toys4K qualitative figure already includes GT and Teacher-II, but the paper should clarify whether Teacher-II represents the complete cascade generated from the same text prompt and explicitly analyze these comparisons. This would help determine whether DreamBridge moves closer to the true structure and what details it preserves, improves, or loses relative to directly running the cascade. Greater agreement with the teacher alone does not establish greater fidelity to the ground-truth asset or the structural requirements in the text.

Questions for authors

please refer to weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer KT6h

Thank you for recognizing the practical use of the shared sparse-structure interface, the evidence supporting structural improvements, and the detailed evaluation protocols, and for suggesting a direct quality–cost comparison and closer inspection against GT geometry.

W1: Quality and inference cost relative to the complete cascade.

We measured the following operating points on the same 24 GB RTX 4090 with offloading, using 48 GSO objects and two fixed inference repeats per method. Every Teacher-II request includes a fresh FLUX image.

MethodCLIP ×100 ↑GT CD ↓GT NC ↑Service time ↓ (s)
BASE30.480.02280.621113.48
DreamBridge31.180.02190.639814.17
Full cascade, Teacher-II32.300.01850.6680124.02

All three rows use 25 steps. Full service time includes offloading, mesh export, and request bookkeeping.

At 25 steps, DreamBridge saves approximately 110 seconds per asset relative to Teacher-II, an 88.6% reduction in measured service time. This moves image synthesis into the offline supervision pipeline while preserving direct text inference.

W2: Geometry against GT assets and text-specified structure.

We evaluated the 744-object GSO expansion against its scanned meshes. The 743 valid BASE–DreamBridge pairs give the following results.

GT metricBASEDreamBridgePaired Δ95% CI
Chamfer distance ↓0.01920.0183−0.00083[−0.00176, +0.00010]
Normal consistency ↑0.64450.6522+0.0077[+0.0012, +0.0144]

The normal-consistency gain provides an independent measure of surface fidelity to the scan. Intervals use paired-asset bootstrap resampling at the fixed checkpoint and inference seed, retaining the failed generation in the 744-object coverage accounting.

The two new appendix figures examine complementary aspects of the generated structures. Multi-view detail comparisons with ground-truth 3D assets places GT, BASE, DreamBridge, and Teacher-II side by side under the same prompts. The complementary views make supporting structures, part boundaries, proportions, and object count directly comparable with the GT geometry.

Multi-view structural details on BridgeStruct compares BASE, DreamBridge, and Teacher-II using reverse views and enlarged views of part connections, openings, and accessible cavities. Across these displayed comparisons, DreamBridge produces more coherent connections and cavity boundaries than BASE, with the complementary views making the spatial organization directly inspectable. Together, the figures connect the aggregate structural results to visible 3D part geometry.

W3: Definition of Teacher-II and presentation of the comparisons.

Teacher-II is the complete same-prompt cascade. The FLUX image generated from the prompt conditions both sparse-structure prediction and downstream generation. We checked the prompt, seed, and input-image records for all 1,250 Toys4K cases and will state this definition in the original qualitative caption and both new appendix captions.

The new figures show complementary viewpoints for every method and enlarge the relevant structural regions. Assets are rigidly reoriented for comparable presentation, and the captions describe the viewing and scaling conventions. The discussion summarizes the structural patterns visible across the displayed comparisons, with cohort-level conclusions supported by the quantitative tables above.

AuthorsOpenReview ↗

Genuinely looking forward to your reply

Dear Reviewer KT6h,

Thank you for taking the time to review our paper and for your valuable feedback. With only two days remaining in the author–reviewer discussion period, we would be happy to address any remaining questions or concerns you may have about our paper.

Best regards,
Authors of Submission 1017

Thanks to the authors for your rebuttal.

I think a full generation pipeline is necessary in 3D generation. Some technical implementations, such as placing Flux on different GPUs or applying acceleration in multi-stage generation models, would be better solutions. In industrial applications, frameworks such as TIGON would be preferable (using Flux to generate images, then executing the parallel pipeline).

Therefore, I will keep my initial score.

AuthorsOpenReview ↗

New Response to Reviewer KT6h

Thank you for the suggestion. Native end-to-end text-to-3D generation serves important practical needs, including rapid construction of large synthetic 3D datasets and workflows driven by explicit textual specifications. DreamBridge strengthens this existing pathway through a small parameter update—1.33% of the sparse-structure flow—without collecting additional real 3D training assets or adding inference stages. At the reported 25-step operating points, BASE and DreamBridge take 13.48 and 14.17 seconds per asset, respectively.

Multi-GPU serving can improve cascade execution. DreamBridge addresses the complementary objective of improving the native text-conditioned generator itself. For large-scale dataset construction under a fixed compute budget, the additional image-generation work per asset can represent a substantial aggregate cost.

Our preceding response already presents the quality–cost evaluations and complementary multi-view comparisons. These results support assessing quality gains together with generation cost, and establish the practical value of a stronger native text-to-3D model alongside cascaded and multimodal approaches such as TIGON.

Best regards,
Authors of Submission 1017

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

The article proposes DreamBridge, where TRELLIS-image-large acts as the teacher and TRELLIS-text-xlarge serves as the student. During the training phase, FLUX is used to synthesize images as a bridge. During student training, the cached teacher end points' latents are fitted, and during inference, pure text is maintained, eliminating the need for FLUX in the inference phase. DreamBridge can improve the generation quality of text-to-3D conversion with a small additional amount of LoRA parameters. However, the DreamBridge scheme is relatively naive. Sharing the same SS VAE latent space between the teacher and the student is the prerequisite rather than the outcome of this work, and the framework name is larger than the actual demonstration content.

Critical strengths
  1. The paper is easy to understand.

  2. Only a small LoRA needs to be trained, and the inference does not affect the downstream process at all. It is extremely lightweight in terms of engineering. Such "plug-and-play minor modifications" have low costs.

Critical weaknesses
  1. This method is relatively simple and lacks novelty. It is merely a minor continuation training scheme for Trellis. The so-called cross-modal distillation originates from the spatial consistency of the two models in Trellis. Essentially, it is no different from other distillation schemes.

  2. Listing "cascade will propagate image synthesis errors" as the reason against cascade, but the teacher supervision in Section 4.1 is precisely generated offline by the complete cascade from FLUX→TRELLIS-image. Distillation only solidifies this error rather than eliminates it. The smaller filtering ratio indicates that the screening hardly changes the target distribution. Teacher supervision essentially is the unfiltered product of the cascade. If this error can be ignored, the motivation paragraph does not hold; if it is significant, the teacher's objective itself is not clean, and its impact needs to be quantified. Additionally, why not directly use qualified rendered images via VLM caption, or select the original rendered image-caption data pair? These data should be easier to obtain and cleaner than those generated by Flux.

  3. I have not seen the advantage of DreamBridge over generating 3D from text first and then from image. Because its capability is weaker than the teacher, and the teacher is only 1.2B while the student is 2B.

  4. In Table 1, the student on Uni3D has 0.1930 > the teacher Teacher-II 0.1918. For a distillation method, it is likely that it is noise or the indicator is unreliable. The paper does not explain this.

Questions for authors
  1. Image filtering indicates which images were eliminated. Are there any examples? With the powerful capabilities of current image generation models, I am curious about the failure rate of image generation and the main reasons for those failures.
4 comments
AuthorsOpenReview ↗

Response to Reviewer qxcd

Thank you for recognizing the paper’s clarity and lightweight LoRA adaptation with an unchanged downstream pipeline, and for your questions about the shared representation, filtering, and inference cost.

W1: This method is relatively simple and lacks novelty.

TRELLIS provides the common coordinates. We use them to teach the text branch from an image it never receives as input. For each accepted bridge image, we save the frozen image teacher’s sparse latent and pair it with the prompt that generated the image. Only the text model’s sparse-structure flow is adapted. The 13.0M trainable parameters bring teacher-referenced occupancy F1 from 0.1512 to 0.2069 while preserving the native text-only inference path. The paired structural results are given in our response to Reviewer 68Na.

W2 and Q1: Which images are filtered, and how much does the choice of bridge image change the target?

The 98.4% figure in the paper is prompt retention. At the candidate-image level, 4,680 of 19,566 images fail at least one technical check.

Technical checkFailed candidatesFraction of all candidates
Foreground-area change after background removal4,56023.31%
Largest-component fraction1910.98%
Crop or border contact1260.64%
Foreground fraction440.22%
Multiple foreground components140.07%
Foreground retained after preprocessing20.01%
At least one technical check, unique candidates4,68023.92%

Failure categories overlap. A prompt remains eligible when another candidate passes, giving 8,062 retained prompts out of 8,192. We will report both denominators and add rejected-image examples illustrating foreground-area changes and border contact.

Synthesizing the bridge from the training prompt lets us choose the object attributes and part relations first, then obtain a corresponding native structural target without sourcing a new GT 3D asset. The presentation-suffix ablation further increases CLIP ×100 from 30.06 to 30.17 and Uni3D from 0.1955 to 0.1986 in the 2K setting. The student receives the original prompt throughout. Rendered image–caption pairs remain another possible supervision source.

W3: Advantage over serving the image teacher.

A text request to the image teacher also requires image synthesis. On the same 24 GB RTX 4090 with offloading, DreamBridge takes 14.17 s per asset and the fresh FLUX-to-3D cascade takes 124.02 s at 25 steps. This is an 88.6% reduction in full service time over 48 GSO objects with two repeats. Our response to Reviewer KT6h gives the compact quality–cost comparison.

W4: How should we read the Uni3D scores of 0.1930 and 0.1918 in Table 1?

On 1,522 prompts valid for both models, DreamBridge minus Teacher-II is +0.0010 in Uni3D, with 95% CI [−0.0013, +0.0033]. We will report this interval alongside the point estimates. The structural-transfer claim rests on the paired gains over the original text model in occupancy, normals, and depth.

AuthorsOpenReview ↗

Genuinely looking forward to your reply

Dear Reviewer qxcd,

Thank you for taking the time to review our paper and for your valuable feedback. With only two days remaining in the author–reviewer discussion period, we would be happy to address any remaining questions or concerns you may have about our paper.

Best regards,
Authors of Submission 1017

I appreciate the authors' response, which improved my understanding of the paper and resolved most of my concerns. However, I remain of the view that the proposed approach is relatively straightforward, with insufficient novelty and limited engineering contribution; in addition, the resulting solution applies only to the specific case of TRELLIS. I will retain my score. I consider this work to be borderline, with a slight lean towards rejection.

AuthorsOpenReview ↗

New Response to Reviewer qxcd

Thank you for confirming that our response resolved most of your concerns. DreamBridge contributes a supervision mechanism that transfers image-conditioned structural priors into an existing native text-to-3D generator, without collecting additional GT 3D training assets. With the representation and downstream generator held fixed, teacher-referenced occupancy F1 increases from 0.1513 to 0.2070 on 1,518 paired prompts, improving on 81.8% of them.

TRELLIS supplies the shared structural interface through which this transfer is demonstrated. This enables a controlled study of how cross-modal supervision improves the text model itself. We respectfully ask that the novelty assessment consider this supervision construction and demonstrated improvement, alongside the backbone used to implement it.

Best regards,
Authors of Submission 1017

Reviewer_68NaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DreamBridge, a cross-modal distillation framework for text-to-3D generation. The method transfers geometric priors from a pretrained image-to-3D teacher to a text-to-3D student through a shared sparse-structure latent space. DreamBridge consists of three components: Bridged Triplets Synthesis, Native Latent Distillation, and Anchor Replay. By adapting only the student’s sparse-structure flow with LoRA, the framework preserves text-only inference without requiring additional ground-truth 3D data. In the experiment, this paper provides qualitative and quantitative experiments.

Critical strengths
  1. Leveraging cross-modal information to improve the quality of text-to-3D generation is a well-motivated direction.

2.The paper provides qualitative visualizations comparing the proposed method with existing baselines.

Critical weaknesses

1.Although the paper has a good motivation for using cross-modal information, the proposed method is not very novel. The proposed method mainly combines existing techniques, including LoRA fine-tuning, knowledge distillation.

2.The experimental comparisons lack recent text-to-3D methods published in 2026, such as AR3D-R1 and ReConText3D. Including more recent baselines and discussing related replay-based approaches would strengthen the evaluation. [1]Are We Ready for RL in Text-to-3D Generation? A Progressive Investigation [2]ReConText3D: Replay-based Continual Text-to-3D Generation

3.Table 1 uses too many colors, making the results difficult to read. I suggest highlighting only the best and second-best results in each column.

4.Proposed DreamBridge achieves leading results on the constructed BridgeStruct-1.5K benchmark but does not consistently outperform existing methods on Toys4K and GSO. The paper may analyze the reasons.

Questions for authors

See the weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer 68Na

Thank you for recognizing the motivation for cross-modal transfer and the qualitative comparisons with existing baselines, and for suggesting recent baselines and encouraging us to clarify our contribution.

W1: Although the paper has a good motivation for using cross-modal information, the proposed method is not very novel.

DreamBridge contributes a way to train a text-conditioned generator from an image-conditioned structural prior. A prompt of our choosing generates a bridge image, the frozen image teacher converts that image into a native sparse-structure target, and the student learns from the original prompt and that target. The shared TRELLIS representation makes this transfer possible while keeping the downstream generator frozen.

Bridged Triplets Synthesis supplies the matched text–structure supervision, Native Latent Distillation adapts the sparse-structure flow, and Anchor Replay incorporates samples from the original text model during adaptation. This yields supervision for 8,062 prompts without collecting additional GT 3D assets, using 13.0M trainable parameters, or 1.33% of the sparse-structure flow. The resulting model generates directly from text.

The transfer produces the following changes relative to the original text model. All means below use the same 1,518 valid paired prompts.

Teacher-referenced metricBASEDreamBridgePaired Δ95% CI
Occupancy F1 ↑0.15130.2070+0.0557[+0.0523, +0.0591]
Normal cosine ↑0.55710.6095+0.0524[+0.0485, +0.0564]
Normalized depth error ↓0.14070.1184−0.0223[−0.0242, −0.0203]

Occupancy F1 improves on 81.8% of these prompts. Intervals are pointwise 95% paired-prompt bootstrap intervals. The appendix also reports the 2K/8K training configurations and their structural results.

W2: The experimental comparisons lack recent text-to-3D methods published in 2026, such as AR3D-R1 and ReConText3D.

We evaluated the released AR3D-R1 and ReConText3D models on BridgeStruct. DreamBridge improves all six reported measures against both methods.

Metricvs. AR3D-R1, Δ [95% CI]Paired nvs. ReConText3D, Δ [95% CI]Paired n
Occupancy F1 ↑+0.0371 [+0.0325, +0.0417]1,472+0.1042 [+0.0991, +0.1094]1,522
Normal cosine ↑+0.0302 [+0.0258, +0.0347]1,472+0.0396 [+0.0339, +0.0455]1,520
Normalized depth error ↓−0.0135 [−0.0159, −0.0110]1,472−0.0259 [−0.0289, −0.0228]1,520
12-view CLIP ×100 ↑+1.94 [+1.82, +2.06]1,473+5.73 [+5.55, +5.90]1,523
Uni3D ↑+0.0296 [+0.0270, +0.0322]1,473+0.0581 [+0.0552, +0.0610]1,523
ULIP-2 ↑+0.0516 [+0.0487, +0.0544]1,473+0.0994 [+0.0957, +0.1030]1,523

Δ is DreamBridge minus the comparison method. The first three rows use frozen Teacher-IT as the geometric reference.

On GSO, we recomputed all four methods on the same 937 objects valid across all five metrics. Each row below uses exactly this common set.

MethodGT CD-24 ↓GT NC-24 ↑CLIP ×100 ↑Uni3D ↑ULIP-2 ↑
BASE0.019440.639931.230.16030.1809
DreamBridge0.018730.649231.240.16290.1920
AR3D-R10.018250.635329.620.14710.1619
ReConText3D0.033140.651027.160.14610.1441

Relative to AR3D-R1, DreamBridge improves GT normal consistency by +0.01383 (95% CI [0.00613, 0.02143]) and CLIP ×100 by +1.620 (95% CI [1.433, 1.809]). Uni3D and ULIP-2 also improve, by +0.01573 and +0.03016, with paired 95% intervals excluding zero. Against ReConText3D, DreamBridge has lower GT Chamfer distance and higher scores on all three text-alignment measures.

ReConText3D uses replay to retain earlier categories during continual learning. Our replay source is the frozen original text generator, used alongside image-teacher targets during adaptation. In the 2K development ablation, 10% replay raises guidance-direction cosine from 0.5431 to 0.5785. The matched 2K replay control is included in the appendix alongside the complete ablation results.

W3: Table 1 is difficult to read.

We will remove the color shading, bold the best result, and underline the runner-up in each column, retaining the teacher as a separately identified reference.

W4: Proposed DreamBridge achieves leading results on the constructed BridgeStruct-1.5K benchmark but does not consistently outperform existing methods on Toys4K and GSO.

BridgeStruct measures agreement with the frozen teacher's structure for a prompt. Toys4K and GSO measure fidelity to a particular GT asset, whose exact geometry can be underspecified by its caption. On Toys4K, ULIP-2 increases from 0.2850 to 0.2888. On the GSO expansion, normal consistency with the scan increases from 0.6445 to 0.6522 over 743 valid pairs. We will make the reference used by each metric explicit in the tables and discussion.

AuthorsOpenReview ↗

Genuinely looking forward to your reply

Dear Reviewer 68Na,

Thank you for taking the time to review our paper and for your valuable feedback. With only two days remaining in the author–reviewer discussion period, we would be happy to address any remaining questions or concerns you may have about our paper.

Best regards,
Authors of Submission 1017

Thank you for the detailed rebuttal and additional experiments. While the response partially addresses my concerns, my main concerns regarding methodological novelty (W1) and the results on Toys4K and GSO (W4) remain unresolved. Therefore, I maintain my original score.

AuthorsOpenReview ↗

New Response to Reviewer 68Na

Thank you for the follow-up. We would like to clarify the methodological contribution and the external-dataset evidence separately.

W1 — Methodological contribution.
DreamBridge contributes prompt-aligned native structural supervision across conditioning modalities: the originating text is paired with structural targets obtained from a frozen image-conditioned 3D teacher. This enables the existing text generator to learn from image-conditioned structural knowledge while keeping its representation, downstream generator, and decoders fixed. The adaptation updates only 1.33% of the sparse-structure flow parameters and requires no additional GT 3D training assets.

W4 — Toys4K and GSO.
On Toys4K, all three text-alignment metrics—CLIP, Uni3D, and ULIP-2—have higher point estimates than BASE. GSO additionally provides GT-referenced geometric evidence. On the 743 valid pairs from the 744-object expansion, normal consistency increases from 0.6445 to 0.6522, with a paired gain of 0.0077 (95% CI: [0.0012, 0.0144]).
In the common 937-object comparison, DreamBridge improves GT normal consistency and all three text-alignment measures over AR3D-R1. Against ReConText3D, it achieves lower GT Chamfer distance and higher CLIP, Uni3D, and ULIP-2 scores. These external evaluations provide evidence of useful transfer beyond agreement with the image teacher.

We have also revised and re-uploaded the manuscript PDF, adding two new qualitative comparison figures to the appendix covering BridgeStruct and the external datasets GSO and Toys4K.

We sincerely appreciate your time and thoughtful feedback, which have helped us strengthen the paper.

Best regards,
Authors of Submission 1017

Comments on the paper

1 comment
Area_Chair_j4JbOpenReview ↗

Participate in the Discussion

The authors have provided their responses to the reviewers' comments. We would appreciate it if the reviewers could review the responses and share any further thoughts during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.