acceptodds
Under review as a conference paper at ICLR 2027

AlphaAnything: Extending Pretrained RGB Generators to RGBA

Abstract

Existing RGBA generation methods commonly learn a transparency-specific representation through reconstruction, introducing an additional training stage and potentially altering the RGB decoding function inherited from pretrained models. We present AlphaAnything, a lightweight framework that extends pretrained RGB generators to RGBA directly in their original latent space, without RGBA reconstruction pretraining. AlphaAnything inserts LoRA modules into the generative transformer and attaches an alpha branch to the frozen RGB VAE decoder. Unidirectional RGB-to-alpha feature sharing allows opacity supervision to shape the generative latent while guaranteeing that, for any fixed latent, RGB decoding remains identical to that of the pretrained decoder. We further introduce Endpoint-Aware Alpha Loss (EA Loss), which derives opacity supervision from the complementary foreground–background weights used in compositing. It supports exact opacity endpoints and continuous partial transparency while emphasizing localized alpha errors. Experiments on text-to-RGBA and image-to-RGBA tasks across multiple pretrained RGB backbones demonstrate competitive generation quality, and controlled ablations validate the shared-latent training design and EA Loss.

Primary area:
generative models
Keywords:
Diffusion Models, Flow Matching, RGBA Generation, Visual Generation

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_BUm2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AlphaAnything presents a method for extending pretrained RGB generators to RGBA generation. The method retains and freezes the original RGB VAE, adapts the generative Transformer using LoRA, and adds a decoder-side alpha branch to produce RGB and opacity from the same generated latent, without separately training an RGBA VAE. The paper also introduces Endpoint-Aware Alpha Loss (EA Loss), which combines a Fisher–Rao opacity coordinate, Fenchel–Young score supervision, and separate square-root aggregation of foreground- and background-weighted contributions. Experiments cover text-to-RGBA and image-conditioned RGBA generation across multiple backbones. The results are competitive overall, with particularly notable improvements on the image-conditioned task.

Critical strengths
  1. The motivation to avoid training an RGBA VAE is reasonable and practically relevant. Adapting an RGB VAE to RGBA requires additional representation learning and training design, and may alter its original RGB reconstruction capabilities. By freezing the RGB VAE and adding an alpha branch that receives RGB features unidirectionally, the paper offers a feasible approach to reusing pretrained representations.

  2. EA Loss is motivated by the specific properties of opacity. The alpha channel includes exact transparency and opacity as well as continuous partial transparency. Compared with directly applying MSE, EA Loss more explicitly accounts for small residual errors near the endpoints, incorrect saturation, and localized detail errors. Its theoretical design is therefore well motivated.

  3. The method achieves promising results and demonstrates applicability across backbones. Experiments with multiple pretrained models and both conditional generation tasks support the feasibility of the framework. The quantitative results and qualitative examples on the image-conditioned task also demonstrate good opacity prediction and compositing quality.

Critical weaknesses
  1. A direct comparison with an alternative that also avoids training an RGBA VAE is missing. A central motivation of the paper is to preserve the original RGB VAE, but its implementation still introduces a trainable alpha branch. Another approach worth evaluating is to reuse the same frozen VAE to encode RGB and alpha separately, after replicating alpha into three channels, concatenate the two latent representations along the DiT's sequence dimension, and adapt the model for joint generation using LoRA (to my knowledge, this training approach works well). This approach is related to Zippo's dual-latent formulation. However, Zippo uses the Stable Diffusion 2.1 UNet and concatenates the latents along the channel dimension. Its architecture differs from sequence concatenation in a modern DiT, so its results do not directly establish the effectiveness or training difficulty of the latter approach.

  2. The computational overhead of EA Loss during diffusion training and its effect on the final sampled outputs require further evaluation. Each evaluation of EA Loss requires converting the DiT's velocity prediction into an estimated clean latent z^0\hat z_0, then processing this estimate through the RGB decoder's feature hierarchy and the alpha branch to compute the supervision. Although the original RGB decoder parameters are frozen, gradients must still propagate through the decoder to update the DiT LoRA. This supervision therefore adds decoding and gradient-propagation costs. The small reported inference overhead does not establish the additional training cost; comparisons of training time and GPU memory against training with only the latent-domain generative loss are needed. Furthermore, EA Loss is applied to a one-step estimate of z^0\hat z_0 at a randomly sampled noise level, rather than to the final output of the full multistep sampling process. The pixelwise properties of EA Loss still hold for this estimate, but those properties alone do not guarantee corresponding improvements in the final sampled outputs.

Questions for authors
  1. Please add an ablation experiment as discussed in Critical Weakness 1.
  2. Please explain why EA Loss remains appropriate for diffusion training.
6 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer BUm2 (Q1)

We sincerely thank the reviewer for your time and effort in reviewing our work, as well as for your constructive and valuable feedback.

Zippo presents a compelling approach to RGBA generation by reusing a pretrained RGB VAE to encode alpha maps as three-channel grayscale images. It jointly generates an RGB image and a grayscale alpha map through separate latent representations, enabling the modeling of color and opacity.

AlphaAnything explores a different perspective: pretrained RGB generators can learn competitive RGBA generation capabilities without RGBA reconstruction pretraining or a separate alpha latent. In AlphaAnything, RGB and opacity are generated from a single shared latent, with an alpha branch reading features from the frozen RGB decoder. During training, EA Loss trains the alpha branch and propagates gradients through the frozen decoder into the DiT LoRA, allowing opacity supervision to shape the shared generative latent.

Following your suggestion, we implemented a Zippo baseline with a separate alpha latent on the same DiT backbones, concatenating RGB and alpha latents along the token dimension. For the Zippo baseline, we use the same training settings as AlphaAnything, including the LoRA configuration. We evaluate both methods on RGBA-Text, RGBA-Image, and RefMatte-RW100 (Referring Image Matting, CVPR 2023), using the same evaluation settings and metrics as in our paper.

Table A. T2RGBA evaluation on RGBA-Text.

BackboneMethodCLIPScore ↑Aesthetic ↑ImageReward ↑
FLUX.1Zippo0.83625.54230.7925
AlphaAnything0.83535.52560.7588
FLUX.2Zippo0.86755.49030.9553
AlphaAnything0.86565.51250.9752

Table B. I2RGBA evaluation on RGBA-Image and RefMatte-RW100.

DatasetMethodRGB L1 ↓Soft IoU ↑SAD ↓LPIPS ↓Alpha Grad ↓Alpha Conn ↓
RGBA-ImageFLUX.2 Zippo0.11280.805250.93780.021071.933748.5306
RGBA-ImageFLUX.2 AlphaAnything0.11550.829646.35720.020769.594746.0840
RefMatte-RW100FLUX.2 Zippo0.17220.5850167.58610.158438.6026165.5608
RefMatte-RW100FLUX.2 AlphaAnything0.14490.6469158.35930.149742.1427157.3270

AlphaAnything achieves better results on five of six metrics on both RGBA-Image and RefMatte-RW100, while obtaining comparable results on RGBA-Text. These findings support our central insight that competitive RGBA generation does not require a separate latent representation for alpha. We also provide qualitative comparisons with the Zippo baseline for T2RGBA and I2RGBA at anonymous link (T2RGBA) and anonymous link (I2RGBA), respectively.

Table C. Inference cost for T2RGBA and I2RGBA. We report inference time per image and peak memory using the same inference setting on a single GPU.

BackboneMethodTime / image (s)Peak memory (GiB)
FLUX.1Original T2RGB12.35933.573
Zippo T2RGBA26.94833.585
AlphaAnything T2RGBA12.45433.664
FLUX.2Original T2RGB41.00334.468
Zippo T2RGBA85.97234.481
AlphaAnything T2RGBA41.11534.558
FLUX.2Original I2RGB87.53834.468
Zippo I2RGBA140.64034.481
AlphaAnything I2RGBA87.63834.558

Additionally, during inference, AlphaAnything's can generate RGBA images while remaining nearly identical time cost of the original RGB generator, and is 2.16×, 2.09× and 1.60× faster than Zippo on FLUX.1 and FLUX.2 T2RGBA and FLUX.2 I2RGBA. These results demonstrate that generating RGB and alpha from a shared latent achieves competitive quality while substantially reducing inference latency, making it particularly attractive for large scale applications.

AuthorsOpenReview ↗

Rebuttal to Reviewer BUm2 (Q2)

EA Loss.

We appreciate the reviewer’s question about our opacity supervision based on a clean latent estimated in one step. The estimate z^0\hat z_0 can indeed differ from the final latent obtained through multistep sampling, but applying opacity supervision to the multistep final output at every training update would require running a multistep sampling and propagating gradients through multiple denoising steps. This can substantially increase computation and requiring additional activation storage or recomputation.

Our design already accounts for the varying reliability of this estimate by loss reweighting, while keeping opacity supervision computationally practical.

Given the model’s velocity prediction v^\hat v, the clean latent estimate is z^0=zσ−σv^\hat z_0=z_\sigma-\sigma\hat v, and we apply EA Loss to the alpha decoded from this estimate. We explicitly reduce the weight of this opacity supervision at high noise levels, where z^0\hat z_0 could be less reliable. The difference between z^0\hat z_0 and z0z_0 depending on noise level follows directly from the formulation: zσ=(1−σ)z0+σϵ, v⋆=ϵ−z0z_\sigma=(1-\sigma)z_0+\sigma\epsilon,\ v^\star=\epsilon-z_0, thus ∥z^0−z0∥22=σ2∥v^−v⋆∥22\|\hat z_0-z_0\|_2^2=\sigma^2\|\hat v-v^\star\|_2^2.

So for a given error in the predicted velocity v^\hat v, the error in the estimated clean latent z^0\hat z_0 decreases as the noise level decreases. We therefore use loss reweighting (Sec. 3.4, Eq. (12)) to retain stronger opacity supervision at moderate and low noise levels, where z^0\hat z_0 is expected to be more accurate and closer to the final latent obtained through multistep sampling, and reduce the weight at high noise levels.

Similar strategies also appear in SpeeDiff (CVPR 2026) and ImageReward (NeurIPS 2023). Specifically, SpeeDiff applies losses to images decoded from z^0\hat z_0, while the ReFL method in ImageReward applies reward feedback to clean sample predictions at selected late denoising steps.

The effect on final sampled outputs is evaluated directly by the stop-gradient ablation in Table 4: allowing EA Loss gradients to update the DiT through z^0\hat z_0 improves opacity accuracy and composite fidelity after multistep inference on both backbones.

Training overhead.

Table D quantifies the additional cost of extending backbone training to RGBA generation. “Backbone LoRA (reference)” uses the standard latent flow matching objective without alpha prediction or opacity supervision. It establishes the computational cost of backbone training under matched settings, allowing us to measure the additional cost of learning opacity.

Table D. Training cost under matched settings. Time per optimizer update is reported as mean ± standard deviation.

Backbone / TaskTraining configurationTime / update (s)Peak memory (GiB)
FLUX.1 T2RGBABackbone LoRA (reference)2.925 ± 0.07136.84
Zippo baseline6.435 ± 0.04139.19
AlphaAnything3.840 ± 0.06344.65
FLUX.2 T2RGBABackbone LoRA (reference)2.629 ± 0.04236.86
Zippo baseline5.768 ± 0.05539.34
AlphaAnything3.588 ± 0.05943.78
FLUX.2 I2RGBABackbone LoRA (reference)5.750 ± 0.00239.35
Zippo baseline9.872 ± 0.00141.83
AlphaAnything6.694 ± 0.00244.86

Relative to this reference, AlphaAnything adds 0.915 s-0.959 s per optimizer update across the three settings, with 5.51–7.81 GiB of additional peak memory. This manageable additional cost enables training for RGBA generation capability on the pretrained backbones.

For the comparison between RGBA generation methods, AlphaAnything requires much less time per optimizer update than the Zippo baseline in all three settings. AlphaAnything generates color and opacity from one shared latent, retaining the original generated sequence length, whereas Zippo adds a separate sequence of alpha latent tokens that the DiT processes at every update. As reported in our response to Q1, our training and samping time savings are achieved while maintaining competitive T2RGBA performance and achieving better results on five of six metrics on both RGBA-Image and RefMatte-RW100.

We hope our responses have addressed your concerns. If you have any remaining questions or concerns, we would be glad to provide further clarification and address them.

Dear Reviewer BUm2,

We sincerely thank you for your time and effort in reviewing our work and for your valuable feedback.

As the discussion period draws to a close, we would like to briefly summarize how our rebuttal addresses your main concerns:

  1. Following your suggestion, we added a Zippo baseline that reuses the same frozen RGB VAE to encode RGB and alpha separately and concatenates their latents along the DiT sequence dimension; AlphaAnything achieves competitive T2RGBA performance, better results on five of six metrics on both RGBA-Image and RefMatte-RW100, and lower inference latency than this baseline.
  2. We clarified the rationale for EA supervision on one-step clean-latent estimates and its noise-dependent weighting, and highlighted the stop-gradient ablation showing that propagating this supervision to the DiT improves opacity accuracy and composite fidelity after full multistep sampling.
  3. We quantified the training overhead against the backbone LoRA reference: 0.915–0.959 s per optimizer update and 5.51–7.81 GiB of additional peak memory, and showed that AlphaAnything requires less training time per update than Zippo.

We would be happy to provide further clarification and answer any remaining questions. If you feel our responses have addressed your concerns, we would appreciate it if you could consider raising your score.

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

AlphaAnything adds RGBA generation to pretrained RGB generators without a separate RGBA reconstruction-pretraining stage or a separately sampled alpha latent. It trains transformer LoRA modules and an alpha branch attached to the frozen RGB VAE decoder. Features flow from RGB to alpha, leaving RGB decoding unchanged for a fixed latent; alpha-loss gradients also update the generator. Its Endpoint-Aware Alpha Loss combines a Fisher-Rao opacity coordinate, a Fenchel-Young score loss that can represent exact endpoints, and foreground/background-weighted root aggregation. Experiments cover text-to-RGBA and image-to-RGBA on several backbones. On RGBA-Image, the Qwen-Image-Edit variant reaches Soft IoU 0.9223 and composite LPIPS 0.0136. The appendices report inference-seed robustness, paired bootstrap intervals, human preferences, and inference overhead.

Critical strengths
  1. The architecture adds opacity while reusing the pretrained RGB latent space and decoder. The fixed-latent preservation property is stated precisely and follows directly from the one-way feature dependencies. The stop-gradient and coupled-decoder ablations support letting alpha supervision update the generator and keeping alpha features out of the RGB path.

  2. The experiments cover several backbones and both generation and extraction. Three inference seeds, bootstrap intervals that account for shared source images, and blinded human comparisons give more support to the reported gains. The profiling also shows that the alpha branch adds less than 0.5% to end-to-end inference time under the measured settings.

Critical weaknesses
  1. The experiments do not isolate the loss components, and the evidence for preserving RGB appearance during generation is limited. In Appendix B.5, the alpha-L1 baseline removes the Fisher coordinate, Fenchel-Young gap, endpoint-group weighting, and root aggregation together, and directly regresses a linear alpha output. Table 5 shows that the complete objective beats this baseline; it does not show which parts are needed or rule out simpler losses that support exact endpoints and reweight errors. Controlling the loss scale would also make the comparison easier to interpret. Proposition 1 correctly establishes equal RGB decoding for a fixed latent, but LoRA training changes the sampled latent, as Appendix A.1 acknowledges. VAE reconstruction results and the single generation example in Figure 6 do not quantify retention of the original generator's foreground detail or appearance distribution across prompts. Thus, an exact decoder property supports only part of the practical motivation for RGB preservation (Sections 3.2-3.3 and 4.2-4.4; Appendix B.5).

  2. The benchmarks leave uncertainty about alpha accuracy on real semitransparent content and about the source of the gains over other systems. RGBA-Image's reference layers are generated by an image generation/editing model and manually screened. Appendix B.3 explicitly notes that they need not reconstruct the source exactly. Agreement with those layers therefore does not establish accuracy against independently measured opacity. Text-to-RGBA metrics mainly assess appearance and alignment on black and white composites, with limited direct assessment of opacity or foreground-color decomposition. Human preferences support visual quality, but do not establish reference accuracy. The comparisons also vary in backbone and training data. FLUX.1 versus AlphaVAE does match a backbone; further comparisons using the same backbone and data, including simple alpha heads/losses and RGB-plus-matting pipelines, would help attribute the gains to the proposed training and loss design (Tables 1-3; Appendices B.2-B.4 and C.6).

Questions for authors
  1. Which parts of EA Loss account for the improvement? Please separately test Fisher versus physical-alpha coordinates, Fenchel-Young versus simpler supervision that supports exact endpoints, foreground/background grouping, and root versus mean aggregation. Tune the loss weights or control gradient scales, and report results separately for endpoints, boundaries, and genuinely partial-opacity regions.

  2. How well does the trained model retain the original RGB generator's appearance and detail over a substantial set of matched prompts and noise samples, allowing for the intended removal of backgrounds? Proposition 1 holds at a fixed latent, while LoRA changes the generator's latents. Please distinguish the architectural guarantee from the properties demonstrated by generated outputs.

  3. Can you test alpha and foreground-color accuracy on independent real images with trusted mattes, including glass, smoke, fine boundaries, and other semitransparent cases? For RGBA-Image, please quantify reference reliability beyond manual filtering, for example through source-reconstruction audits, independent matte annotation, and compositing on additional backgrounds.

  4. Beyond the FLUX.1/AlphaVAE backbone match, how do the methods compare when trained with the same backbone, RGBA data, update budget, and inference settings? Useful controls would include a simple attached alpha head with tuned standard losses and the original RGB generator followed by a strong matting/extraction model, using comparable image or text conditioning.

8 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer eZNu (Q1)

We sincerely thank the reviewer for their time and effort in reviewing our work, as well as for their constructive and valuable feedback.

EA Loss is motivated by the specific requirements of opacity learning. Alpha values of 0 and 1 represent complete transparency and opacity, while intermediate values describe continuous partial transparency. Localized alpha errors, such as residual opacity or small holes, can visibly degrade an asset despite affecting only a small fraction of its pixels. These requirements guide both how individual opacity predictions are supervised and how their errors contribute to the overall training objective. Our motivation is to make opacity supervision reflect how alpha controls compositing. Each alpha value determines both opacity α\alpha and complementary transparency 1−α1−\alpha. We therefore introduce the opacity representation (α,1−α)(\alpha, 1−\alpha), which expresses these two quantities jointly. This opacity representation lies on a Bernoulli simplex, with exact transparency and opacity at its two vertices and continuous partial opacity between them. We construct both pixelwise supervision and spatial aggregation from this opacity representation, so that the same opacity representation guides how prediction errors are measured and how they contribute to the complete objective.

The effectiveness of the complete objective is evaluated on the final generated outputs in Table 5, and the results establish the effectiveness of EA Loss, complementing the mathematical characterization. Targeted component ablations require several additional full generative training runs and are currently in progress. The experiments take a considerable amount of time to run, and we will add the results as soon as they are complete.

AuthorsOpenReview ↗

Rebuttal to Reviewer eZNu (Q2)

RGB preservation. In our paper, RGB preservation means that adding opacity does not introduce RGB reconstruction degradation through changes to the VAE. The pretrained RGB VAE already provides high-quality reconstruction, while the comparisons in Section 4.2 show that RGBA VAE retraining can affect fine details and color fidelity. Retaining the original RGB decoding function therefore removes an avoidable source of quality loss; Proposition 1 formalizes this protection for every fixed latent. LoRA training learns RGBA generation and can change the sampled latents, so we additionally evaluate the generated outputs with the trained LoRA enabled.

RGB generation performance. We evaluate the original RGB models and trained AlphaAnything checkpoints on GenEval, together with CLIPScore, AES, and ImageReward. Prompts, initial noise, and sampling settings are matched within each backbone. AlphaAnything’s RGBA outputs are evaluated on black and white backgrounds using the predicted alpha. No prompt enhancement is used.

Table A. Evaluation on GenEval

BackboneModelGenEval ↑CLIPScore ↑AES ↑ImageReward ↑
FLUX.1Original0.66080.80025.66561.0234
AlphaAnything0.66020.78835.48990.9258
FLUX.2Original0.83050.82615.19411.1555
AlphaAnything0.85760.82305.38471.3105
LongCatOriginal0.74460.78775.31881.0774
AlphaAnything0.77090.81115.43531.1589
Qwen ImageOriginal0.75010.82015.42101.1986
AlphaAnything0.73950.80635.44761.0457

Overall, GenEval performance remains comparable to that of the original backbones, and appearance and preference scores also remain broadly stable. These results indicate that extending the pretrained models to RGBA generation does not substantially compromise their general image generation quality or semantic understanding.

AuthorsOpenReview ↗

Rebuttal to Reviewer eZNu (Q3)

Clarifying reference construction and quality control. RGBA-Image evaluates instruction-guided extraction of semantic RGBA layers across graphic-design posters, stylized scenes, and photorealistic scenes. As detailed in Appendix B.3, we construct candidate reference layers through image generation and editing, followed by manual review to assess the requested semantic content, opacity, and color fidelity.

Each candidate layer is inspected alongside its source image and alpha matte. Layers with any inaccurate opacity, mismatches with the description or extraction instruction, or any visible color deviations are rejected. Annotation review also corrects ambiguous target descriptions and incomplete semantic grouping instructions. This process checks both the visual quality of the reference layer and whether it represents the intended extraction target.

Generating candidate layers followed by human review has precedents in transparent-image dataset construction. LayerDiffuse (Transparent Image Layer Diffusion using Latent Transparency, SIGGRAPH 2024) uses iterative generation and human selection during dataset development, while PrismLayers (PrismLayers: Open Data for High-Quality Multi-Layer Transparent Image Generative Models, arXiv 2025) combines generation and matting with automatic filtering and human selection. These workflows provide a practical basis for constructing diverse RGBA assets with explicit quality control.

Additional evaluation on independent real images. To directly assess alpha accuracy using external annotations, we evaluate AlphaAnything and baselines on RefMatte-RW100 (Referring Image Matting, CVPR 2023) under same inference setting as Table 3 in our paper, using its official alpha annotations. This dataset includes semitransparent regions and fine object boundaries, allowing evaluation of continuous opacity on real images.

Table B. I2rgba Evaluation on RefMatte-RW100

MethodRGB L1 ↓Soft IoU ↑SAD ↓LPIPS ↓Alpha Grad ↓Alpha Conn ↓
Qwen-Image-2.10.10710.6265179.60070.164242.0059175.3418
OmniAlpha0.07590.7802121.69000.113127.3980121.3204
Ours (FLUX.2)0.14490.6469158.35930.149742.1427157.3270
Ours (Qwen Image Edit)0.06820.863266.60900.072027.559465.2704

Our Qwen Image Edit variant achieves the best Soft IoU, SAD, and Alpha Conn among the evaluated methods, with scores of 0.8632, 66.6090, and 65.2704, respectively. These results provide independent evidence of accurate alpha prediction on real images beyond our constructed RGBA-Image benchmark.

Reviewer_bGy7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Traditional image editing models could only process RGB images, and this paper makes some attempts to extend a pretrained RGB generation model to process RGBA images. To this end, this paper introduces AlphaAnything, which trains a LoRA module in the generative transformer and attaches an alpha branch to the pretrained RGB VAE decoder. Experiments show that the proposed method achieves competitive results compared to models that specifically trained an RGBA VAE and DiT.

Critical strengths
  1. This paper shows that without training an RGBA VAE and DiT, a lightweight LoRA and alpha decoder with alpha supervision could achieve comparable results to SOTA RGBA-based models.
  2. The endpoint-aware alpha loss is new and effective, as shown in the experiments.
Critical weaknesses
  1. It would be interesting to see if the performance on RGB image generation/editing tasks drops by evaluating the trained RGBA model on RGB tasks and comparing it with the pretrained model. This assessment is important for deploying the model for real-world applications.
Questions for authors

I suggest the authors evaluate the trained AlphaAnything model on general-purpose RGB generation tasks on standard benchmarks to see if it retains the generation ability of the original model.

4 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer bGy7

We sincerely thank the reviewer for their time and effort in reviewing our work, as well as for their constructive and valuable feedback.

Deployment and applications. AlphaAnything is trained specifically for RGBA asset generation, so transparent backgrounds are an intended part of its output. RGBA generation uses the trained LoRA and alpha branch, while for RGB generation without an alpha channel, the LoRA can be disabled, allowing the pretrained model to generate images as usual. This allows both RGB and RGBA generation to coexist within the same deployment, sharing the frozen pretrained backbone and RGB VAE.

Generation ability after RGBA training. We also evaluate whether the pretrained RGB models’ visual quality and text–image alignment are retained in the trained AlphaAnything checkpoints. We use GenEval, together with CLIPScore, AES, and ImageReward, with matched prompts, initial noise, and sampling settings within each backbone. The original models are evaluated on their native RGB outputs.

BackboneModelGenEval ↑CLIPScore ↑AES ↑ImageReward ↑
FLUX.1Original0.66080.80025.66561.0234
AlphaAnything0.66020.78835.48990.9258
FLUX.2Original0.83050.82615.19411.1555
AlphaAnything0.85760.82305.38471.3105
LongCatOriginal0.74460.78775.31881.0774
AlphaAnything0.77090.81115.43531.1589
Qwen ImageOriginal0.75010.82015.42101.1986
AlphaAnything0.73950.80635.44761.0457

Overall, GenEval performance remains comparable to that of the original backbones, and appearance and preference scores also remain broadly stable. These results show that AlphaAnything largely retains the pretrained models’ visual quality and text–image alignment when producing RGBA assets with explicit transparency.

We hope our responses have addressed your concerns. If you have any remaining questions or concerns, we would be glad to provide further clarification and address them.

Dear Reviewer bGy7,

We sincerely thank you for your time and effort in reviewing our work and for your valuable feedback.

As the discussion period draws to a close, we would like to briefly summarize how our rebuttal addresses your main concern:

  1. Following your suggestion, we evaluated the trained AlphaAnything models on GenEval across four backbones, and found broadly comparable GenEval performance after RGBA training.
  2. We clarified that standard RGB generation remains available by disabling the LoRA and using the original RGB output, allowing RGB and RGBA generation to coexist within the same deployment.

We would be glad to provide further clarification and answer any remaining questions. If you feel our responses have addressed your concerns, we would appreciate it if you could consider raising your score.

Dear authors,

Thank you for submitting the rebuttal with the RGB generalization study, which has addressed my concerns. I will remain with the Weak acceptance rating.

Best, Reviewer bGy7

Dear Reviewer bGy7,

Thank you very much for your time and careful consideration of our rebuttal and additional experiment.

We are very glad that our response and additional result have addressed your main concerns. We also sincerely appreciate your positive feedback and recognition of our work.

Thank you again for your thoughtful review, encouraging feedback, and support.

Best regards, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Kind Follow-up Before the Discussion Closes

Dear Reviewers,

Thank you again for your feedback and for taking the time to review our work. We would like to briefly follow up and ask whether our rebuttal and additional results have addressed your main concerns, as the discussion period draws to a close. We are really eager to hear your further feedback. If any concern remains, we would be glad to provide more clarification or address any remaining questions.

Thank you again for your time and consideration.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.