acceptodds
Under review as a conference paper at ICLR 2027

BTC3D: Blended Tile Conditioning for Detail-Enhancing Image-to-3D Generation

Abstract

Recent diffusion-based pipelines have achieved promising progress in image-to-3D synthesis. However, generating high-fidelity details remains challenging, especially when the input image contains rich details. Existing approaches often rely on globally encoded conditioning features, which compress spatial information and limit the model to reproduce fine-grained details. This common design often leads to a phenomenon we term detail attenuation. Moreover, improving image-to-3D synthesis quality typically requires retraining or fine-tuning large diffusion models, which can be computationally expensive and impractical for complex 3D pipelines. In this work, we present Blended Tile Conditioning for image-to-3D generation (BTC3D), a training-free inference time framework that enhances fine-grained detail preservation in image-to-3D diffusion pipelines. To alleviate detail attenuation, we first examine the image feature additivity in image-to-3D models. Based on this property, we introduce a blended tile embedding that extracts local conditioning signals from split image regional patches, allowing the diffusion model to better preserve fine-grained visual details. To integrate the global and local conditioning guidance stably, we propose a dynamic conditioning schedule that gradually increases the influence of tile-level conditioning during later low-noise stages of diffusion. Our proposed method BTC3D operates entirely at inference time and can be seamlessly integrated into existing image-to-3D diffusion pipelines. Experimental results demonstrate that the proposed approach significantly improves texture quality and visual fidelity of the base model while maintaining global structural consistency in a training-free manner.

Primary area:
generative models
Keywords:
Image-to-3D Generation

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

The paper suggests a training-free method to enhance pre-trained 3D flow-matching models' texture handling. In practice, they suggest averaging features obtained from patches of images and combining these with the original image features in a dynamically scheduler-conditioned way. The suggested method achieves a boost in performance across 3 tested models.

Critical strengths
  • The paper is mostly well-written and easy to follow.
  • The qualitative results are impressive, and the suggested method looks practical.
Critical weaknesses
  • Novelty - the suggested method is too simple. While quantitatively effective, the suggested observation is neither deeply explored nor explained. And while functional, its novelty is very limited.

  • Image feature additivity - while interesting and a bit counterintuitive, the authors mention these observations with limited experiments backing them up. The qualitative visuals, while convincing, don't really explain the TSNE plot shown, which does show a difference in the semantic space. And the authors don't try to further explain this or explore it deeply.

  • Limited ablation studies - The authors show general component ablation, but do not explore different design choices like weighting the patches according to masks, different scheduling schemes for the dynamic integration, or even different aggregation methods that are not a simple weighted average.

Questions for authors

Please see weaknesses

5 comments

We appreciate your feedback and recognition about our clear presentation, impressive qualitative results, and the practical value of our method. Below, we address each of your concerns.

W1: BTC3D provides a training-free conditioning framework for improving fine-grained detail preservation in pretrained image-to-3D pipelines. It combines local crop encoding, foreground-based feature aggregation, and dynamic conditioning while maintaining compatibility with the pretrained generator and keeping its parameters unchanged.

In our paper, global conditioning refers to the full-image DINO representation, including spatial patch tokens. BTCemb independently encodes local crops and aggregates their features using foreground-based weights, providing fine-grained local information that complements the global condition. This design is motivated by the empirical observations of image feature additivity and compatibility in Sec. 3.1, with further supporting analyses in Appendix C.1–C.2.

DyCond further adjusts the normalized tile weights during texture generation, changing the relative contributions of individual tiles. BTCemb and DyCond combine local crop encoding, foreground-based feature aggregation, and dynamic tile weighting to enhance detail preservation in pretrained image-to-3D pipelines without retraining. To further evaluate the contributions of these components, we have added controlled comparisons in Appendix D.6 of the revised manuscript. The results show that foreground-mask weighting improves generation quality over uniform averaging, while dynamic conditioning provides further improvements across PSNR, SSIM, and LPIPS. These additional results support the effectiveness of the proposed combination.

W2: In this work, image feature additivity refers to the empirical compositional compatibility of independently encoded tile features within the pretrained conditioning space. Fig. 2(a) illustrates the correspondence between global and averaged tile embeddings. The differences between them do not contradict this observation, since compositional compatibility does not require the aggregated tile embedding to be identical to the global embedding.

To complement the t-SNE visualization, Appendix C.2 reports a quantitative evaluation of global-local feature compatibility on 210 Toys4K objects from 105 categories, with two objects per category. We compute the cosine similarity between each global feature and its corresponding aggregated tile feature. We additionally compare against a non-matching global feature from the same category. The results are summarized below as mean [95% bootstrap confidence interval].

RepresentationCosine similarity ↑Same-category cosine margin ↑
Average tiles0.7466 [0.7179, 0.7758]0.1728 [0.1433, 0.2041]
BTCemb0.8242 [0.8035, 0.8434]0.2195 [0.1894, 0.2517]

The averaged tile embedding remains aligned with its corresponding global feature, and BTCemb further improves this alignment. The positive same-category margins support input-specific semantic correspondence beyond coarse category separation. These results provide quantitative evidence for the empirical compositional compatibility of global and local image features.

At the generation level, Fig. 2(b) shows that averaged tile embeddings can directly condition the pretrained generator. Appendix C.2 further provides a complementary example in which the averaged conditioning embeddings of an original image and its color-modified version produce a result combining color characteristics from both inputs. Together, the feature comparisons and generation examples support the empirical observation motivating our method.

W3: We have added controlled comparisons of tile weighting and dynamic conditioning on all 101 images of 3D-Arena using TRELLIS.2. The tile-based variants share the same crops, encoder resolution, generation seeds, and fixed global/local fusion ratio. All configurations use the same total number of sampling steps.

MethodPSNR ↑SSIM ↑LPIPS ↓
Baseline20.67410.84800.4693
Uniform averaging20.18850.82610.4806
Static foreground-mask weighting20.91170.83340.4725
Ours21.31490.87210.4597

Our method improves all three metrics over uniform averaging, and dynamic conditioning provides further improvements, achieving the best results among the evaluated configurations. These quantitative results and corresponding visual comparisons have been added to Appendix D.6.

The schedule ablations in Appendix D.2–D.3 further support the local-to-global tile expansion direction and the choice of β=2\beta=2 among the evaluated settings. Appendix D.4 additionally provides intermediate renderings, detail crops, and tile weights to illustrate differences between static and dynamic conditioning.

I thank the authors for their response. I acknowledge the empirical usefulness of the suggested method. But, I still hold the opinion that the observation in the paper is technically simple and under-explored (theoretically or in-depth empirically under various conditions) which really limits the contribution of the paper. While its hard to asses novelty, I'm still under the impression that the contribution in the paper does not qualify for a main conference paper.

Thank you for acknowledging the empirical usefulness of our method. We would like to clarify the methodological contribution and the evidence supporting it.

BTC3D introduces a training-free conditioning framework for preserving fine-grained details in pretrained image-to-3D generation. Its design builds on the empirical compositionality and compatibility of independently encoded local image features examined in Sec. 3.1. BTCemb uses these properties to construct a local condition through foreground-based tile aggregation, while DyCond adjusts the relative contributions of individual tiles during sampling, keeping the global/local fusion ratio fixed. Together, these components enable complementary local information to improve detail preservation through the pretrained generator’s existing conditioning interface.

The empirical investigation includes controlled frequency-response analysis (Appendix C.1) and global/local feature comparisons with same-category negatives across 210 objects from 105 categories (Appendix C.2). We have further added controlled comparisons in Appendix D.6 of the revised manuscript, using matched crops, encoder resolution, generation seeds, and global/local fusion coefficients for the tile-based variants. Foreground-mask weighting improves PSNR, SSIM, and LPIPS over uniform averaging, and dynamic conditioning further improves all three metrics. These analyses examine both the underlying feature properties and the contributions of the proposed components.

Regarding practical applicability, we demonstrate consistent improvements across TRELLIS, TRELLIS.2, and Hunyuan3D-2.1, covering DINOv2 and DINOv3 encoders on both 3D-Arena and Toys4K. This supports the applicability of the conditioning framework across different pretrained pipelines without retraining or modifying model parameters.

We would appreciate clarification on which specific aspect of the proposed framework you consider insufficiently novel, or which experimental condition remains insufficiently examined in light of these analyses. We would be happy to discuss these points further and provide additional clarification as needed.

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

This paper proposes BTC3D, a training-free inference-time framework for enhancing detail preservation in image-to-3D generation. The method extracts conditioning features from image tiles using a frozen visual encoder, aggregates them according to foreground coverage, and blends the resulting representation with full-image conditioning. A dynamic conditioning mechanism adjusts the relative contributions of individual tiles during sampling. The framework is integrated into pretrained generation pipelines without updating their parameters. Experiments on 3D-Arena and Toys4K, using TRELLIS, TRELLIS.2, and Hunyuan3D-2.1, demonstrate improvements over the corresponding baselines. Component and hyperparameter ablations, feature analyses, region-level evaluation, and a human preference study provide complementary evidence for its effectiveness.

Critical strengths
  1. Clear motivation and presentation. The paper has a coherent narrative, and the method is developed around its stated motivation. The clear and accessible writing makes the proposed framework and its design choices easy to follow.

  2. A practically useful inference-time improvement validated across multiple backbones. BTC3D improves detail preservation without retraining the underlying models. Its evaluation on multiple pretrained pipelines supports its practical applicability, while the user study provides complementary evidence of improved perceived quality.

Critical weaknesses
  1. The description of dynamic conditioning does not accurately reflect the formulation. The paper describes DyCond as progressively increasing the influence of local conditioning during later diffusion stages. However, the global/local fusion coefficient remains fixed; the time-dependent changes occur within the relative weights of the BTCemb tiles. This distinction should be clarified because it affects the interpretation of the proposed mechanism.

  2. Some targeted ablations and process-level visualizations are missing. Although the paper includes component ablations and static-versus-dynamic comparisons, these do not fully isolate the effects of the different tile aggregation strategies or establish the necessity of retaining global conditioning. Intermediate denoising visualizations are also needed to better support the explanation of how dynamic conditioning enhances details.

Questions for authors
  1. Clarification of DyCond. The introduction states that DyCond “progressively increases the influence of BTCemb during the later diffusion stage.” However, Eq. (8) uses a fixed fusion coefficient α\alpha, so the nominal global and local mixing coefficients remain 1−α1-\alpha and α\alpha, respectively. The dynamic operation instead changes the normalized weights within BTCemb. Could the authors clarify this distinction and revise the relevant descriptions.

  2. Comparison of fusion strategies. Beyond the existing component ablations, could the authors provide a controlled comparison of the original global-only baseline, uniform tile averaging, foreground-weighted tile averaging, and dynamic tile weighting? For the tile-based variants, please use the same crops and global/local fusion coefficient, and report both quantitative results and visual comparisons. In addition, Figure 2 provides a qualitative example of replacing global conditioning with averaged tile features, but the fusion-ratio ablation does not include α=1\alpha=1. Could the authors evaluate this local-only setting with the proposed weighting and schedule to clarify the role of global conditioning in preserving structure?

  3. Intermediate visualizations of DyCond. The reported results show that dynamic conditioning outperforms static conditioning, but the mechanism would be clearer with intermediate denoising visualizations. Could the authors compare the two variants at representative sampling timesteps? Renderings or decoded clean predictions, together with corresponding detail crops and tile weights, would help show when and where the improvement emerges.

4 comments

We thank the reviewer for recognizing the clear motivation and presentation of our work, as well as the practical applicability of our training-free approach across multiple backbones. Below, we address each of your questions.

W1&Q1: Thank you for pointing out this distinction. We agree that the original wording was imprecise. In Eq. (8), α\alpha is a fixed hyperparameter controlling the fusion ratio between the global feature and the aggregated tile feature throughout sampling. DyCond progressively adjusts the normalized tile weights by β\beta within BTCemb, changing the relative contributions of individual tiles while keeping the global/local fusion ratio fixed. We have revised the relevant descriptions in the Introduction and Method sections with color blue.

W2&Q2: Global conditioning captures holistic semantic information and helps maintain overall structural consistency during generation, while local conditioning provides visual details. As shown in Fig. 2(b), replacing global conditioning with the average tile embedding can improve local details while preserving the rough structure, but may disturb global structure. We therefore retain global conditioning and combine it with local conditioning to balance global coherence and detail enhancement.

We have extended Figure 16 of Appendix D.5 in blue color of the revised manuscript to include results for local-only conditioning (α=1\alpha=1). This setting retains the proposed tile weighting and dynamic schedule. In the illustrated examples, both PSNR and ULIP-2 are lower than with α=0.4\alpha=0.4, and some outputs show reduced color consistency with the input image. These results support the benefit of retaining global conditioning.

As the table below, we added a controlled comparison of the original global-only baseline, uniform averaging, static foreground-mask weighting, and our dynamic conditioning on all 101 images of 3D-Arena using TRELLIS.2. The three tile-based variants share the same candidate crops, encoder input resolution, generation seeds, and fixed global/local fusion coefficient. Both quantitative results and visual comparisons are provided in the appendix subsection Appendix D.6 Ablation of the Fusion Strategy in blue color. Our method BTC3D achieves the highest PSNR and SSIM and the lowest LPIPS among the evaluated configurations. The accompanying visual comparison shows more distinct fine-grained surface textures in the illustrated example.

MethodPSNR ↑SSIM ↑LPIPS ↓
Baseline20.67410.84800.4693
Uniform averaging20.18850.82610.4806
Static foreground-mask weighting20.91170.83340.4725
Ours21.31490.87210.4597

W2&Q3: Thank you for this suggestion. We have added intermediate visualizations to Appendix D.4 Intermediate Visualizations of Dynamic Conditioning in blue color. We compare static and dynamic conditioning at three representative timesteps of a 12-step sampling trajectory, with β=2\beta=2 for dynamic conditioning. The figure shows decoded and rendered clean predictions, corresponding detail crops, and selected input tiles with their weights. In the illustrated example, dynamic conditioning produces more distinct details in the highlighted mouth region than static conditioning and achieves a higher final PSNR.

AuthorsOpenReview ↗

Genuinely looking forward to your reply

Dear Reviewer CDdY,

Thank you for taking the time to review our paper and for your valuable feedback. We submitted our response a few days ago and would appreciate any further comments you may have. We would be happy to clarify any remaining questions or concerns.

Best regards,

Authors of Submission 1060

Thank you for the response. My concern regarding the description of DyCond has been addressed. The additional comparisons of fusion strategies and visualizations of the DyCond mechanism are also helpful. I have no further questions.

Thank you for your thoughtful feedback and careful consideration of our responses. Your constructive comments have helped us clarify the description of DyCond and strengthen our experimental analysis. We sincerely appreciate your time and engagement throughout the review process.

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

The paper proposes BTC3D, a training-free way to improve detail preservation in image-to-3D generation. It encodes image crops separately, combines their features using foreground-based weights, and blends them with the full-image condition. A timestep-dependent schedule changes the relative weights of the crops during sampling.

Critical strengths
  1. The method is practical. It works with frozen models and only changes how the image condition is constructed and used during inference. For users who cannot afford to fine-tune a large 3D generator, this is a useful direction.
  2. The results cover three backbones and both DINOv2 and DINOv3 encoders. This gives the approach more credibility than a demonstration on a single pipeline. The gains in Table 1 are also consistent across the tested backbone dataset combinations.
  3. Several visual examples show a noticeable improvement in surface patterns and small appearance details.
Critical weaknesses
  1. It is unclear how much of the improvement comes from simply giving the image encoder access to higher-resolution details. Each crop is resized to the encoder’s input resolution, so small patterns become easier to resolve even before the proposed weighting or scheduling is applied. The frequency analysis in Appendix C.1 is consistent with this explanation. A comparison against higher-resolution full-image encoding, where supported, would help establish what BTC3D adds beyond this effect.
  2. The ablations do not sufficiently isolate the value of foreground weighting and the particular dynamic schedule. Table 2 compares the full method against static conditioning, but does not include uniform crop averaging with a simple late-stage blending schedule. Table 6 compares uniform and foreground-weighted features only in feature space, without showing the corresponding generation quality.
  3. The evaluation does not yet clearly distinguish faithful detail preservation from an overall improvement in appearance. Comparing multiple rendered views against one conditioning image mixes texture differences with viewpoint and shape differences, while ULIP-2 and Uni3D mainly measure asset-level alignment. Table 9 could provide stronger evidence, but the alignment procedure behind its matched views and spatially paired regions is insufficiently explained.
Questions for authors
  1. With the same crops, encoder resolution, and generation seeds, how does BTC3D compare against uniform crop averaging combined with a simple late stage global/local blending schedule?
  2. How are the matched views and spatially paired regions in Table 9 obtained? Please clarify the reference images, camera settings, asset alignment, and valid-region selection. Are baseline and BTC3D evaluated using exactly the same alignment?
  3. What happens when the full-image encoding resolution is increased, where supported by the backbone? A comparison at a similar encoding cost would help establish how much benefit comes from the proposed aggregation rather than access to finer image information.
4 comments

We thank the reviewer for recognizing the practicality of our method, the credibility of our evaluation across backbones, the consistent gains across tested settings, and the noticeable improvements in surface details in several examples. Below, we address each of your concerns.

Q1&W2: As the table shown below, we added the performance comparisons on all 101 images of 3D-Arena using TRELLIS.2. The tile-based variants share the same crops, encoder input resolution, and generation seeds. Our proposed BTC3D achieves higher PSNR and SSIM and lower LPIPS than your mentioned uniform averaging mechanism.

MethodPSNR ↑SSIM ↑LPIPS ↓
Baseline20.67410.84800.4693
Uniform averaging20.18850.82610.4806
Ours21.31490.87210.4597

Q2&W3: We use the original GT-rendered conditioning image as the fixed reference. All images are rendered at 1024 × 1024 with a camera radius of 2.2, elevation of 15°, and focal length of 50 mm. The reference azimuth is 45°. Baseline and BTC3D predictions are independently registered to their paired GT assets using the same geometry-only similarity registration procedure (PCA initialization followed by trimmed ICP). Each prediction has its own estimated transform, which remains fixed throughout evaluation.

Both methods share the same selected cameras and valid regions. For each object and backbone, we select the two baseline views with the highest CLIP-I similarity to the reference from 12 candidate azimuths ranging from 15° to 345° at 30° intervals, and reuse these cameras for BTC3D. Both views are compared against the same fixed reference. On a fixed 4 × 4 grid, a region is retained when the intersection of the reference and baseline foreground masks covers at least 20% of its area. The same region coordinates are used for both methods, without separate selection for BTC3D. PSNR and SSIM are not used for registration or view selection.

We have added the full evaluation protocol to Appendix C.4 (Table 9). For more detailed information, please refer to this section.

Q3&W1: We first clarify that both the baseline and BTC3D follow the backbone's official image preprocessing and native conditioning resolutions. Images are preprocessed and resized before encoding: e.g. the image is rescaled to 1024 × 1024 by default for TRELLIS.2. Consequently, increasing the source-image resolution alone does not automatically increase the actual conditioning resolution under the default pipeline.

Directly increasing the encoder-input resolution requires modifying this default conditioning configuration. We therefore evaluate it separately as an additional resolution ablation, while keeping the pretrained weights and 3D generation resolution unchanged.

We further conduct a resolution ablation with TRELLIS.2 as the backbone on 100 randomly selected objects from the GSO dataset (seed 2026). We render images at 512 × 512, 1024 × 1024, and 1536 × 1536 resolutions. As shown below, increasing the full-image encoding resolution beyond 1024 × 1024 yields diminishing returns, with no further improvement observed for BTC3D.

MethodEncoder inputRegion PSNR ↑Region SSIM ↑
Baseline512 × 51219.77210.8082
Baseline1024 × 102419.85680.8113
Baseline1536 × 153619.85710.8128
Ours512 × 51220.31270.8359
Ours1024 × 102420.47110.8470
Ours1536 × 153620.47100.8467
AuthorsOpenReview ↗

Genuinely looking forward to your reply

Dear Reviewer GXjZ,

Thank you for taking the time to review our paper and for your valuable feedback. We submitted our response a few days ago and would appreciate any further comments you may have. We would be happy to clarify any remaining questions or concerns.

Best regards,

Authors of Submission 1060

Reviewer_GXjZOpenReview ↗

Thanks for your reply.

Thanks for the authors’ response. I read the paper again and also considered the comments from the other reviewers. Based on all of this information, I have reached my final decision. I think the paper tackles a meaningful problem, since models like TRELLIS may keep the overall object structure but still lose some fine local details. The paper proposes a simple inference-time method to improve this, and I find the idea of using tile-level DINO features as plug-and-play conditioning for several existing image-to-3D models reasonable and practically useful. It's especially nice that the method doesn't need any training or fine-tuning, adds only modest absolute inference overhead, and still gives consistent improvements across several metrics. Overall, I see this as a strong engineering work with a well-designed system. However, in terms of novelty, I feel that a lot of the main ideas are already pretty common, like dividing an image into local patches or tiles, enlarging crops to preserve details, and combining global and local features. Because of this, the paper feels more like transferring ideas from 2D high-resolution diffusion to image-to-3D generation. I do recognize that making these ideas work reliably in a 3D generation pipeline may involve non-trivial technical challenges, so I still appreciate the engineering effort. The paper also gives readers a useful takeaway that details lost through global encoding may be recovered at inference time without retraining the model. My main remaining concern is that the experiments don't fully show whether the added details are actually correct. Producing sharper or more detailed textures doesn't necessarily mean those details match the input or show up in the right places on the 3D object. Overall, I think the method is simple and useful, and the experiments are relatively complete, but the methodological novelty and broader takeaway still feel somewhat limited and incremental. So I've decided to keep my current score.

Thank you for revisiting our paper and for recognizing the importance of the problem and the practical value of our approach.

The methodological contribution of BTC3D lies in a training-free conditioning framework built on the observed compositionality and compatibility of local DINO features with pretrained image-to-3D generators. BTCemb aggregates independently encoded tiles using foreground-based weights, while DyCond adjusts their relative contributions during texture generation. The feature analyses and controlled comparisons in Appendix C.1–C.2 and D.6 examine the empirical basis and contributions of this design.

Regarding detail fidelity, the paired region-level evaluation in Appendix C.4 compares baseline and BTC3D outputs against the same reference image using shared selected cameras and valid regions. BTC3D achieves higher region PSNR and SSIM across all three backbones, providing quantitative evidence of improved local appearance fidelity under this protocol.

We sincerely appreciate your careful assessment and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.