PaperlayerView exact revision on arXiv

Source-linked AI summary

AlphaAnything: Extending Pretrained RGB Generators to RGBA

Qijie Wang, Zhixin Xu, Zhe Li, Haowei Zhu, Honghao Qiao, Bo Peng, Dawei Li, Bin Wang

arXiv:2610.32413v1cs.CV

TL;DR

Existing RGBA methods commonly rely on reconstruction-trained transparency representations, while AlphaAnything extends pretrained RGB generators directly in their original latent space. It adds a unidirectional alpha pathway and endpoint-aware opacity supervision, and reports competitive performance across text-to-RGBA and image-to-RGBA tasks. The authors note that scarce high-quality RGBA data and accepted evaluation benchmarks limit how broadly the method can be evaluated.

  • Problem

    Existing RGBA methods commonly learn a transparency-specific representation through reconstruction, adding a training stage and potentially changing the pretrained RGB decoding function.

    Source: §1 · p. 1

  • Method

    AlphaAnything adds a unidirectional alpha branch to a frozen RGB VAE decoder, letting opacity supervision shape the original RGB latent, and uses EA Loss based on complementary compositing weights.

    Source: §1 · pp. 1, 3

  • Results

    Experiments across multiple backbones and text-to-RGBA and image-to-RGBA tasks show competitive RGBA performance, with improvements in aesthetic quality and text-image alignment.

    Source: §1 · p. 3, §5 · p. 10

  • Takeaways & Limitations

    Opacity is learned as part of generation in the original RGB latent space rather than estimated post hoc from a completed RGB image.

    Source: §3.2 · p. 4

    High-quality RGBA data and widely accepted evaluation benchmarks and metrics remain scarce, limiting how broadly the method can be evaluated.

    Source: §5 · p. 10

Abstract

from arXiv · show

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.

1 INTRODUCTION

Existing RGBA methods commonly add a reconstruction-trained transparency representation, while AlphaAnything learns opacity directly in a pretrained RGB generator’s latent space. Its unidirectional alpha pathway preserves fixed-latent RGB decoding, and EA Loss supervises opacity according to its compositing role.

1 INTRODUCTION

  • Existing RGBA methods commonly add a reconstruction-trained transparency representation, adding a training stage and potentially changing the pretrained RGB decoding function.Zippo reuses a pretrained RGB VAE but still models appearance and opacity with separate latent states.Source: §1 · p. 1
  • AlphaAnything inserts LoRA modules into the generative transformer and attaches a trainable alpha branch to the frozen RGB VAE decoder.Flow-matching supervision updates transformer LoRA while opacity supervision shapes the same predicted latent.Source: §1 · p. 1
  • The alpha branch receives RGB decoder features but never feeds information back, so fixed-latent RGB output remains identical to the pretrained decoder’s output.RGB and alpha come from one latent and one decoder traversal, without an RGBA autoencoder or separate alpha latent.Source: §1 · p. 3
  • EA Loss derives opacity supervision from complementary background–foreground compositing weights, covering exact transparency, exact opacity, and partial transparency.Its endpoint-compatible formulation retains corrective gradients when a prediction saturates at the wrong endpoint.Source: §1 · p. 3
  • EA Loss separates background- and foreground-associated errors before root aggregation, increasing the relative contribution of localized defects such as thin structures, small holes, and background leakage.The paper characterizes these properties theoretically and evaluates the complete objective against standard alpha L1 supervision.Source: §1 · p. 3
  • The paper evaluates shared-latent RGBA generation across multiple pretrained backbones and text-to-RGBA and image-to-RGBA tasks, reporting competitive generation quality.It presents shared-latent extension and endpoint-aware opacity supervision as its contributions.Source: §1 · p. 3

2 RELATED WORK

Pretrained RGB generators provide text- and image-conditioned visual priors, but their outputs remain tied to RGB backgrounds. Prior RGBA methods use different transparency representations, whereas AlphaAnything learns in the original RGB latent space while retaining the pretrained autoencoder.

2 RELATED WORK

  • Pretrained diffusion and flow-based generators offer strong appearance synthesis and editing, but produce conventional three-channel images tied to an RGB background.LoRA offers a parameter-efficient way to introduce capabilities while largely reusing a pretrained backbone.Source: §2 · p. 3
  • Matting estimates opacity from observed RGB evidence, unlike RGBA generation, which jointly synthesizes appearance and opacity from semantic instructions.Native RGBA methods include reconstruction-trained transparency representations and replicated-alpha channels with separate color and matte latent states; AlphaAnything instead uses the original RGB latent space.Source: §2 · p. 3

3 ALPHAANYTHING

AlphaAnything extends a pretrained RGB generator with a shared-latent alpha pathway, preserving fixed-latent RGB decoding while learning opacity. Its endpoint-aware loss handles exact and partial opacity and separates foreground- and background-associated errors.

3.2 EXTENDING RGB GENERATION WITH OPACITY

  • A lightweight alpha branch reuses frozen-decoder RGB features, while transformer LoRA and alpha supervision shape RGBA generation from the original RGB latent.RGB and opacity come from the same sampled latent; opacity gradients pass through the frozen decoder to the LoRA parameters.Source: §3 · p. 3, §3.2 · pp. 4–5

3.1 PRELIMINARIES

  • Compositing combines premultiplied foreground color with background color weighted by 1 − opacity; color-padding supplies a dense RGB target in transparent regions.The padding propagates foreground color from coarse to fine before the frozen RGB encoder defines the target latent.Source: §3.1 · p. 4

3.2 EXTENDING RGB GENERATION WITH OPACITY

  • For any fixed latent, the augmented decoder preserves the pretrained RGB output because RGB operations do not depend on the alpha state.The decoder remains differentiable, so alpha-loss gradients can still supervise the generator.Source: §3.2 · p. 5

3.3 ENDPOINT-AWARE ALPHA LOSS

  • EA Loss maps opacity through Fisher–Rao coordinates, represents exact endpoints with finite scores, and retains corrective gradients for predictions saturated at the wrong endpoint.Partial-opacity targets also have finite exact solutions.Source: §3.3 · p. 6
  • Pair-weighted square-root aggregation assigns errors to target background and foreground mass, with partial-opacity pixels contributing to both terms.For endpoint-group defects over a fraction ρ of pixels, the root contribution scales as √ρ times the mean gap, rather than ρ times the mean gap under direct averaging.Source: §3.3 · p. 6

3.4 JOINT GENERATIVE TRAINING AND INFERENCE

  • Training jointly updates transformer LoRA and the alpha branch while freezing the pretrained transformer and RGB VAE; inference decodes the final clean latent directly.Alpha supervision is reduced at high noise levels, where the one-step clean-latent estimate is less reliable.Source: §3.4 · pp. 6–7

4 EXPERIMENTS

Experiments evaluate AlphaAnything across text-to-RGBA and image-to-RGBA tasks, then test RGB-detail preservation and two design choices: propagating alpha supervision to the DiT and using EA Loss. The reported comparisons favor AlphaAnything on visual-quality and text–image-alignment metrics, while ablations support the shared-latent training design and EA Loss.

4 EXPERIMENTS

  • The evaluation uses 8,963 curated text-to-RGBA training samples and tests multiple pretrained backbones across text- and image-conditioned tasks.Models are trained for 20K optimizer updates; test sets include RGBA-Text, RGBA-Image, and AlphaTest.Source: §4 · p. 7, §4.1 · p. 7

4.1 MAIN RESULTS

  • ALPHAANYTHING improves visual-quality and text–image-alignment metrics in text-to-RGBA evaluations on RGBA-Text and AlphaTest.The AlphaTest comparison covers quantitative results, while the paper also reports qualitative comparisons.Source: §4.1 · p. 7, Table 2 · p. 8

4.2 PRESERVING RGB DETAILS

  • LPIPS rises from 0.03807 to 0.05197 and ∆E00 from 1.917 to 2.078 for OmniAlpha’s RGBA VAE versus the original Qwen-Image-Edit RGB VAE.The reconstruction comparison covers 295 RGBA-Image source images; Figure 5 reports that original RGB VAEs preserve small lettering and colors more faithfully.Source: §4.2 · p. 8, Figure 5 · p. 9
  • Foreground RGB L1 increases from 0.1155 to 0.1349 on FLUX.2 and from 0.0746 to 0.2435 on Qwen-Image-Edit with alpha-to-RGB feedback.The coupled-decoder ablation indicates feedback could degrade foreground RGB fidelity even when pretrained RGB parameters remain frozen.Source: §4.2 · p. 9, Table 4 · p. 10

4.3 EFFECT OF PROPAGATING ALPHA SUPERVISION TO THE DIT

  • Propagating alpha-loss gradients to the DiT improves opacity prediction, alpha boundaries and connectivity, and composite fidelity on both backbones.The reported metrics are higher Soft IoU and lower SAD, LPIPS, Alpha Grad, and Alpha Conn than when the clean-latent estimate is detached.Source: §4.3 · p. 9, Figure 7 · p. 9, Table 5 · p. 10

4.4 EFFECTIVENESS OF THE ENDPOINT-AWARE ALPHA LOSS

  • EA Loss achieves higher Soft IoU and lower SAD, LPIPS, Alpha Grad, and Alpha Conn than direct alpha L1 regression on FLUX.2 and Qwen-Image-Edit.The reported gains are larger on Qwen-Image-Edit.Source: §4.4 · p. 10, Table 6 · p. 10

5 CONCLUSIONS, LIMITATIONS AND FUTURE WORK

The conclusion presents AlphaAnything as a reconstruction-free RGBA extension that keeps RGB decoding unchanged for any fixed latent. The appendix analyzes RGB preservation and gradient flow, opacity geometry, and endpoint-compatible supervision; evaluation is limited by scarce RGBA data and benchmarks.

5 CONCLUSIONS, LIMITATIONS AND FUTURE WORK

  • AlphaAnything extends pretrained RGB generators to RGBA without RGBA reconstruction pretraining or a separate alpha latent.Experiments across text-to-RGBA and image-to-RGBA tasks and multiple backbones show competitive performance, with improvements in aesthetic quality and text-image alignment.Source: §5 · p. 10
  • Scarce high-quality RGBA data, widely accepted benchmarks, and metrics limit how broadly the method can be evaluated.The authors call for broader datasets and metrics covering appearance, opacity, and compositing quality.Source: §5 · p. 10

A.1 RGB PRESERVATION AND GRADIENT FLOW

  • For a fixed latent, the augmented decoder produces the original RGB output for every alpha-branch parameter setting.The RGB output has zero derivative with respect to alpha-branch parameters wherever the derivative exists; the latent itself can change during LoRA training.Source: §A.1 · p. 14
  • The retained gradient path propagates opacity supervision through the predicted velocity to transformer LoRA parameters.Clipping gives unit derivative for unclipped latent coordinates and zero derivative for clipped coordinates away from clipping boundaries.Source: §A.1 · p. 14

A.2 FISHER–RAO GEOMETRY OF OPACITY

  • The Fisher–Rao opacity coordinate extends continuously across [0, 1], with endpoint distance dFR(0, 1) = π.Its inverse is a continuous bijection on [0, 1], and squared coordinate distance is first-order near opacity endpoints but locally quadratic for interior targets.Source: §A.2 · p. 14

A.3 EXACT ENDPOINTS AND CORRECTIVE GRADIENTS

  • EA Loss represents exact transparent and opaque targets at finite scores while retaining corrective gradients for incorrect saturated predictions.Every partial-opacity target has an exact finite fit at s = 2u −1.Source: §A.3 · p. 15

A.4 ROOT AGGREGATION AND ERROR BOUNDS

The appendix analyzes root aggregation and compositing-error bounds, then specifies the experimental protocol and implementation details. For localized binary-target errors, root aggregation scales with the square root of the affected area fraction rather than the fraction itself.

A.4 ROOT AGGREGATION AND ERROR BOUNDS

  • For localized binary-target errors covering fraction ρ of pixels, the root contribution is √ρ¯g, while direct averaging gives ρ¯g.This comparison holds for a fixed mean gap ¯g within one endpoint group.Source: §A.4 · p. 16
  • The compositing-error bound separates foreground RGB error from the opacity contribution controlled by EA Loss and holds for every background in the stated range.The stated range is F, bF, and B in [0, 1]^(N×3) in the common linear color space.Source: §A.4 · p. 16

B EXPERIMENTAL PROTOCOL

  • The appendix identifies the implementation, training, evaluation, and ablation settings used for Section 4.It notes that additional numerical results and visual comparisons appear in Appendices C and D.Source: B EXPERIMENTAL PROTOCOL · p. 16

B.1 IMPLEMENTATION DETAILS

  • The RGB-to-alpha projections use low-rank channel reduction and spatial mixing, with a learned stage gate controlling each injection’s contribution.The five alpha stages use widths (256, 256, 192, 128, 96), and all projections use rank 64; alpha residual outputs and bottleneck cross-attention output are zero-initialized.Source: §B.1 · pp. 16–17

B.2 TRAINING AND INFERENCE SETTINGS

The protocol evaluates text-to-RGBA and image-to-RGBA across several pretrained backbones, with their native interfaces retained. It specifies training data, optimization, conditioning, and inference settings for these comparisons.

B.2 TRAINING AND INFERENCE SETTINGS

  • Text-to-RGBA evaluation uses RGBA-Text, AIM-500, and AlphaTest, while image-to-RGBA evaluation uses RGBA-Image.The backbone comparison includes four text-to-RGBA models and two image-to-RGBA models, retaining each backbone’s native conditioning interface and latent layout.Source: §B.2 · p. 17
  • Text-to-RGBA training uses 8,963 curated assets, while image-to-RGBA training uses ∼180K source-image-target-layer pairs.The training examples in these runs occupy the 1024 × 1024 bucket.Source: §B.2 · p. 17
  • All models train for 20,000 updates, with effective global batch size 16 in the main experiments and learning rate 10−4 for both trainable components.The original transformer, RGB VAE, and text encoder remain frozen; flow and alpha terms each have coefficient 1.Source: §B.2 · p. 17
  • Inference uses 28 steps for FLUX.1 and 50 for FLUX.2, Qwen, and LongCat, with guidance 3.5 or 4 respectively.Controlled comparisons keep sampler, guidance, seed, conditioning inputs, and sampling-step count fixed.Source: §B.2 · p. 17

B.3 CONSTRUCTION OF RGBA-TEXT AND RGBA-IMAGE

The appendix constructs RGBA-Text and RGBA-Image as evaluation sets and describes their curation. RGBA-Text contains prompts without paired reference images, while RGBA-Image contains manually screened layers linked to source images.

B.3 CONSTRUCTION OF RGBA-TEXT AND RGBA-IMAGE

  • RGBA-Text and RGBA-Image are constructed for evaluation only, and neither set is used to train AlphaAnything.RGBA-Text prompts are reported with reference-free scores because they lack paired reference images.Source: §B.3 · pp. 17–18
  • RGBA-Text contains 294 English prompts across nine categories, with descriptions ranging from 10 to 95 words and a median of 53 words.The prompts are curated for coverage, clarity, and distinctiveness, and checked against training captions for overlap.Source: §B.3 · p. 18
  • RGBA-Image contains 735 image-conditioned examples from 295 source images, each with a source image, target RGBA layer, description, and extraction instruction.The examples combine selected poster layers and layers from generated scenes.Source: §B.3 · pp. 18–19
  • The final RGBA-Image set contains 257 poster layers from 110 posters and 478 generated-scene layers from 185 images.Layers are manually screened for opacity accuracy, alignment with descriptions and instructions, and color fidelity; generated extraction does not guarantee exact pixelwise reconstruction.Source: §B.3 · p. 19

B.4.1 TEXT-TO-RGBA EVALUATION

The evaluation compares AlphaAnything with existing text-to-RGBA methods on RGBA-Text and AlphaTest using appearance and text–image alignment metrics.

B.4.1 TEXT-TO-RGBA EVALUATION

  • Comparisons include AlphaVAE, LayerDiffuse, OmniAlpha, and Qwen-Image-2.1 across RGBA-Text and AlphaTest.The evaluation reports FID, CLIPScore, Aesthetic Score, and ImageReward.Source: §B.4.1 · p. 19
  • Appearance metrics are averaged across lossless black- and white-background composites, while FID is computed separately for each dataset and background before averaging.Sample-level scores are averaged uniformly.Source: §B.4.1 · p. 19
  • Table 1 evaluates all 294 RGBA-Text prompts with CLIPScore, Aesthetic Score, and ImageReward; the set has no reference images for FID.AlphaTest uses the benchmark prompts supplied by AlphaVAE.Source: §B.4.1 · p. 19

B.4.2 IMAGE-TO-RGBA EVALUATION

The evaluation specifies image-to-RGBA metrics and protocols, then reports RGB reconstruction fidelity, prompt-template robustness, and results on AIM500. Ablations isolate feature sharing, alpha supervision of the generator, and the alpha loss.

B.4.2 IMAGE-TO-RGBA EVALUATION

  • The 735-example RGBA-Image comparison uses shared prompts and paired target layers, evaluating RGB L1, alpha agreement, alpha structure, and composite perceptual fidelity.Soft IoU and SAD use normalized alpha over the whole image; LPIPS averages black- and white-background composites.Source: §B.4.2 · pp. 19–20

B.4.3 RGB RECONSTRUCTION EVALUATION

  • RGB reconstruction is evaluated on 295 source images using deterministic encode–decode with posterior-mode latents and fully opaque alpha for RGBA encoders.LPIPS and ∆E00 are computed over complete images and averaged equally across inputs.Source: §B.4.3 · p. 20

B.4.4 RGB AND RGBA GENERATION EXAMPLE

  • The generation example compares the original FLUX.1-dev RGB model, AlphaVAE, and AlphaAnything with FLUX.1 using one shared prompt and seed 69.All outputs use 28 steps, guidance 3.5, and 1024 × 1024 resolution; RGBA outputs are displayed on white.Source: §B.4.4 · p. 20

B.5 ABLATION SETTINGS

  • Ablations compare coupled alpha-to-RGB feedback, detaching alpha-loss gradients from the DiT, and replacing EA Loss with direct alpha L1 under matched training settings.The shared settings use 20K optimizer updates on the same 735 target layers.Source: §B.5 · pp. 20–21

B.6 EFFICIENCY ACCOUNTING

  • Trainable-parameter accounting includes only transformer LoRA and alpha-branch parameters; inference adds the alpha branch and its RGB-to-alpha feature injections.The parameter counts are reported by backbone.Source: Table 9 · p. 21, §B.6 · p. 21

C.1 TEXT-TO-RGBA RESULTS ON AIM500

  • All four AlphaAnything backbones obtain higher Aesthetic Score and ImageReward than compared baselines on 500 AIM500 prompts; FLUX.1 leads Aesthetic Score and FLUX.2 leads ImageReward.FID and CLIPScore use the same black/white-composite protocol as the AlphaTest comparison.Source: Table 10 · p. 22, §C.1 · p. 22

C.2 RGB RECONSTRUCTION FIDELITY

  • Both RGBA VAEs have higher LPIPS and CIEDE2000 errors than their corresponding original RGB VAEs: Qwen-Image-Edit rises 0.03807→0.05197 (LPIPS) and 1.917→2.078 (∆E00), while FLUX.1-dev rises 0.01216→0.04762 and 2.109→3.240, respectively.Source: §C.2 · p. 22

C.3 ROBUSTNESS TO PROMPT TEMPLATES

  • Across tested prompt templates, overlap and perceptual fidelity remain robust on both backbones, while absolute alpha error varies more; background-removal instructions lead four metrics for Qwen Image Edit.The tested alternatives also improve Qwen Image Edit’s RGB L1 and Soft IoU over its standard template.Source: Table 12 · p. 23, §C.3 · p. 23

C.4 ROBUSTNESS ACROSS INFERENCE SEEDS

Across three inference seeds, the reported text- and image-to-RGBA findings persist. Paired bootstrap intervals quantify input-sampling uncertainty for comparisons with strong baselines.

C.4 ROBUSTNESS ACROSS INFERENCE SEEDS

  • Across all three seeds, all four AlphaAnything text-to-RGBA backbones outperform compared baselines in ImageReward on both text benchmarks.On RGBA-Image, Qwen Image Edit ranks first on all six metrics in each run, with mean Soft IoU 0.9236 ± 0.0023 and LPIPS 0.0132 ± 0.0009.Source: Table 13 · p. 24, §C.4 · p. 24

C.5 BOOTSTRAP CONFIDENCE INTERVALS

  • The 95% BCa intervals quantify input-sampling uncertainty for means across saved runs, using paired prompt resamples or source-image clusters as appropriate.The comparisons use each metric’s best-observed-mean baseline.Source: §C.5 · p. 25
  • LongCat’s ImageReward gains over OmniAlpha are 0.2771 [0.2282, 0.3296] on RGBA-Text and 0.3060 [0.2715, 0.3435] on AlphaTest; Qwen Image Edit’s RGBA-Image Soft IoU gain is 0.2350 [0.2115, 0.2605].All six Qwen Image Edit paired differences on RGBA-Image favor AlphaAnything.Source: Table 14 · pp. 25–26, §C.5 · p. 25

C.6 HUMAN EVALUATION

AlphaAnything adds less than 0.5% end-to-end runtime overhead across four evaluated backbones. The human evaluation protocol compares anonymized I2RGBA predictions, but the supplied passages do not state its outcomes.

C.6 HUMAN EVALUATION

  • Three experienced annotators judged randomized, anonymized I2RGBA prediction pairs for instruction consistency, visual quality, and natural transparency and object boundaries.Each AlphaAnything backbone was compared with Qwen-Image-2.1 and OmniAlpha, without comparisons between AlphaAnything backbones.Source: §C.6 · p. 26, Table 15 · p. 26
  • The human evaluation reports win rate excluding ties and tie-adjusted win rate for each backbone–baseline pair.The supplied table caption identifies both rates as percentages but provides no outcome values.Source: §C.6 · p. 26, Table 15 · p. 26

C.7 INFERENCE OVERHEAD

  • Inference adds computation only from the alpha branch and its RGB-to-alpha feature injections, with profiling at 1024 × 1024 resolution and 50 sampling steps.The DiT LoRA weights can be merged into pretrained weights at inference.Source: §C.7 · p. 27
  • Across all four evaluated backbones, AlphaAnything adds less than 0.5% runtime overhead to end-to-end generation.The overhead estimate uses additional VAE decoding time relative to matched-control generation time.Source: Table 16 · p. 27, §C.7 · p. 27

D ADDITIONAL QUALITATIVE RESULTS

The appendix adds matched qualitative comparisons for text-to-RGBA and image-to-RGBA generation, component ablations, and an exploratory RGBA-editing study. Its ablations show examples of color shifts and opacity artifacts in variants, while paired editing inputs yield visually similar outputs.

D.1 TEXT-TO-RGBA GENERATION

  • Eleven additional RGBA-Text prompts compare four baselines with four AlphaAnything settings on complete canvases displayed on a common checkerboard.The comparisons follow the evaluation settings in Appendix B.Source: §D.1 · p. 27, Figure 8 · p. 28, Figure 9 · p. 31

D.2 IMAGE-TO-RGBA EXTRACTION

  • Fourteen additional RGBA-Image examples compare four extraction methods using complete canvases and, where informative, matched crops of RGBA details or alpha boundaries.These visual comparisons complement the full-dataset scores in Table 3.Source: §D.2 · p. 33, Figure 10 · p. 34, Figure 11 · p. 36

D.3 UNIDIRECTIONAL FEATURE SHARING

  • On the shown RGBA-Image examples, the coupled decoder introduces foreground color shifts that AlphaAnything avoids.The comparison uses matched checkpoints and inference settings.Source: §D.3 · p. 38, Figure 12 · p. 38

D.4 PROPAGATING ALPHA SUPERVISION TO THE DIT

  • On Qwen-Image-Edit examples, Stop-grad creates opacity artifacts, drops a lettering line, and fills wheel-spoke gaps, while the full model retains more accurate opacity structure.The compared cases include a price badge and wine bottle.Source: §D.4 · p. 39, Figure 13 · p. 39

D.5 ENDPOINT-AWARE ALPHA LOSS

  • On Qwen-Image-Edit examples, alpha L1 produces spurious opacity and transparent holes in opaque text panels, while AlphaAnything preserves target opacity more closely.The examples include a chocolate mold and craftsperson.Source: §D.5 · p. 40, Figure 14 · p. 40

D.6 EXPLORING RGBA UNDERSTANDING THROUGH EDITING

  • In six paired Qwen-Image-2.1 editing cases, RGBA inputs and white-background composites with all-one alpha yield visually similar outputs in overall appearance and opacity structure.The pairs cover color modification, rotation, object removal and addition, material modification, and style transfer; the authors identify the distinctive value of RGBA understanding as future work.Source: §D.6 · p. 41, Figure 15 · p. 41

Related papers

Loading 2610.32413v1…