- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper presents LASAGNA, a unified framework for layered image generation with visual effects. The authors argue that existing layered methods produce transparent foregrounds without realistic shadows/reflections, forcing a harmonization model after every edit and thereby reintroducing identity drift. LASAGNA generates a photorealistic background and an RGBA foreground with object-associated visual effects in a single forward pass, treating effects as part of the foreground layer so that consumer edits (translation, scaling, recoloring, duplication) can be done by alpha compositing alone without model re-inference. A single DiT-based architecture with type/IO/position embeddings handles three modes—background-conditioned foreground generation, foreground-conditioned background generation, and text-to-all layer generation—under diverse conditioning inputs (text, FG, BG, masks). The authors also release LASAGNA-48K, a 48K layered triplet dataset filtered by a VLM-based curator, and LASAGNABENCH, a 242-sample expert-annotated benchmark across six sources. Experiments compare against FLUX.1, Qwen-Image-Edit, gpt-image-1, and LayerDiffuse, plus ablations, user study, and additional benchmarks.
Critical strengths
- The core idea—binding scene-consistent visual effects into the foreground alpha layer—is simple, well-motivated, and practically valuable, enabling zero-inference post-editing and eliminating cascade drift.
- The unified architecture handles three generation modes and heterogeneous conditioning within one model, contrasting favorably with prior per-task expert models; the type/IO embedding interface is a clean design.
- Community contributions are substantial: the first public layered dataset with visual effects and the first standardized layer-centric benchmark, both with documented construction pipelines and human verification.
- Experiments are broad: comparisons against both general models and the closest layered baseline, three modes, GPT-based identity/instruction metrics, a user study, and a paradigm-level necessity analysis (instruct vs. cascaded layer vs. layer+effects).
- The 2B model outperforms much larger general models on several metrics, and the runtime/editing advantage is clearly demonstrated.
Critical weaknesses
- The dataset is built by filtering LayerDecomp outputs with a learned curator; residual decomposition artifacts and the curator's 72.3% recall may propagate into training and benchmark labels, and the benchmark's "ground-truth" BG/FG for public sources is itself model-generated rather than captured.
- Evaluation relies heavily on FID/CLIP-FID and GPT-4o-based scores; the GPT judge may be biased toward the authors' outputs, and no human evaluation of visual-effect physical plausibility is reported beyond a preference study.
- The method is evaluated mainly in the "FG over BG" single-foreground setting; multi-object interactions, occlusion, and complex scene geometry are acknowledged as future work, limiting claims of generality.
- Comparisons with general models use their task-specific expert variants, which is reasonable but means LASAGNA's advantage in "unified" capability is not isolated from raw quality; LayerDiffuse comparison is limited to CLIP-FID.
- Training details (20K iterations, 2B DiT, RGBA-VAE fine-tuning) and data scale are modest; ablation of training data, prompt formulation, and joint training is relegated to the appendix, and sensitivity to curator/filter thresholds is unclear.
Questions for authors
- How much do LayerDecomp artifacts and curator errors propagate into LASAGNA-48K and the benchmark labels? Can you report metrics on human-captured in-house pairs separately?
- How reliable is the GPT-4o judge, and can you provide human evaluation of shadow/reflection physical plausibility and identity preservation?
- How does LASAGNA behave with multiple foreground objects, occlusions, and interacting shadows? What breaks first?
- Since baselines use expert variants, can you report a unified-model comparison or a compute-matched comparison to isolate the benefit of the unified interface?
- How sensitive are results to the RGBA-VAE fine-tuning and the 20K-iteration training budget? Would a larger data scale change the conclusions?
2 comments
Q4: Since baselines use expert variants, can you report a unified-model comparison or a compute-matched comparison to isolate the benefit of the unified interface?
R: Yes, this is already isolated in Table 9 (third row, "Separate task"): under the same training samples, the unified model outperforms separately trained single-task models in all three modes. The gain therefore comes from the unified interface itself.
Q5: More evaluation metrics for the LayerDiffuse comparison (currently limited to CLIP-FID).
R: Our model also significantly outperforms LayerDiffuse on GPT scores. Following the protocol in our paper, we additionally report them on FG_Gen and BG_Gen:
| Model | FG_Gen IF ↑ | FG_Gen IP ↑ | FG_Gen Avg | BG_Gen IF ↑ | BG_Gen IP ↑ | BG_Gen Avg |
|---|---|---|---|---|---|---|
| LayerDiffuse | 7.9 | 6.2 | 7.1 | 5.8 | 5.3 | 5.6 |
| Ours | 9.27 | 9.29 | 9.3 | 9.34 | 8.67 | 9.0 |
This is consistent with Figure 6 and Table 4. We additionally compare with BFS (SIGGRAPH 2026), a recent dedicated FG_Gen model, where our FG layers are also clearly better (FG FID 151.7 vs. 234.9); see Q2 of Reviewer nwuN and Fig. S3 in the newly updated supplementary material.
Q6: How sensitive are results to the RGBA-VAE fine-tuning and the 20K-iteration training budget? Would a larger data scale change the conclusions?
R: Our conclusions are stable: the reported results have converged within the 20K-iteration budget, and a larger data scale would only strengthen them. The RGBA-VAE is important, but it is a general component well validated by the community and outside the scope of our paper, so we did not ablate it specifically. Larger and more diverse data should make the model stronger and may further benefit the challenging cases discussed above.
We hope this addresses your concerns; please feel free to ask any further questions.
We thank the reviewer for recognizing that our core idea (binding scene-consistent visual effects into the foreground alpha layer) is simple, well-motivated, and practically valuable, that the Type/IO interface is a clean design unifying three modes in one model, that our dataset and benchmark are substantial community contributions, that our experiments are broad, and that our 2B model outperforms much larger general models.
Q1: How much do LayerDecomp artifacts and curator errors (72.3% recall) propagate into LASAGNA-48K and the benchmark labels?
R: Very little. For the curator, the 88.8% precision matters more than the 72.3% recall: recall is the fraction of good cases retained, whereas precision is the probability that a sample admitted into the dataset is good. We deliberately trade a little recall for higher precision, so curator errors mainly discard good samples rather than admit bad ones.
Moreover, the benchmark labels of the public sources are not all model-generated: all FGs are human-extracted and the BGs of the two camera sets are real captures; only the BGs of the public sources are model-generated, and all of them passed human verification. Overall, 66.5% of the benchmark's FG/BG layers are produced entirely by humans without any model involvement (all 242 FGs plus the 80 camera-captured BGs, out of 484 layers; Table 10). This high-quality benchmark ensures that the model is iterated in the right direction even if the training set contains slight noise. For the training data, strict filtering (a curator trained on 30K human annotations with hard negatives, plus a 32B VLM) leaves >90% of samples judged good; see Q1 of Reviewer fZAY. Results on the human-captured in-house pairs are reported in Q2 below.
Q2: Reliability of the GPT-4o judge; human evaluation of shadow/reflection physical plausibility and identity preservation; metrics on human-captured in-house pairs.
R: Human evaluation confirms our GPT-4o-based results, both overall and on the human-captured in-house pairs.
Human evaluation. We conducted a comprehensive blinded user study with 1,878 criterion-level votes (Appendix D, Table 6) along three dimensions:
HQ evaluates shadow/reflection physical plausibility, and IP evaluates identity preservation. Our model is the most favored, exceeding the best baseline in Task Avg by 35.6 points in FG_Gen and 38.2 in BG_Gen (see also Q3 of Reviewer nwuN). This ranking is consistent with the GPT-4o scores, showing that the GPT-4o judge is reliable and not biased toward our outputs.
Metrics on human-captured in-house pairs. From the user study, we separately extract the statistics of the in-house samples: 40 camera-captured samples across the two tasks, with 549 valid votes.
Note: BWS stands for Best-Worst Scaling, where scores range from -100 to +100, and higher scores indicate stronger human preference (Appendix D, Table 6) .
FG_Gen (259 votes; IF/IP/HQ: 94/84/81)
BG_Gen (290 votes; IF/IP/HQ: 98/97/95)
Our model achieves the highest Task Avg in both tasks by a wide margin (+46.5 vs. -4.4 and +35.4 vs. -0.6 for the best baseline); the conclusions on real-captured data remain consistent with our original findings.
Q3: How does LASAGNA behave with multiple foreground objects, occlusions, and interacting shadows? What breaks first?
R: LASAGNA supports multiple foreground objects, occlusions, and interacting shadows; see Fig. S1 and Fig. S2 in the newly updated supplementary material. Occlusions and interacting shadows are harder than multiple foreground objects, yet the results remain strong although such cases were never deliberately collected for training. Please also see Q1 of Reviewer XXLt (multi-object layers in a multi-turn manner, Figure 8) and Q2 of Reviewer fZAY (fine hair, fur, and semi-transparency are handled well).