We introduce the first method, to the best of our knowledge, for adapting image-to-video models to layer-aware text (glyph) animation, a capability critical for practical dynamic visual design. Existing approaches predominantly handle the transparency-encoding (alpha channel) as an extra latent dimension appended to the RGB space, necessitating the reconstruction of the underlying RGB-centric variational autoencoder (VAE). However, given the scarcity of high-quality transparent glyph data, retraining the VAE is computationally expensive and may erode the robust semantic priors learned from massive RGB corpora, potentially leading to latent pattern mixing. To mitigate these limitations, we propose TransText, a framework based on a novel Alpha-as-RGB paradigm to jointly model appearance and transparency without modifying the pre-trained generative manifold. TransText embeds the alpha channel as an RGB-compatible visual signal through latent spatial concatenation, explicitly ensuring strict cross-modal (RGB-and-Alpha) consistency while preventing feature entanglement. Our experiments demonstrate that TransText significantly outperforms baselines, generating coherent, high-fidelity transparent animations with diverse, fine-grained effects.
Figures & tables
Figure 1: Visual frame-wise results of our image-to-video (I2V) RGBA glyph animation model. TransText generates diverse and complex glyph animations with accurate, clean transparency while strictly preserving the style of the reference image. See the Supplementary Materials for additional video results.
Figure 2: Overview of the TransText pipeline . We obtain input latents by encoding the RGB video and the RGB-projected Alpha video ( Alpha-as-RGB ) via the VAE. These latents are spatially concatenated for alignment. Additionally, the reference image and its derived trimap serve as structural conditions to guide the joint generation of both RGB textures and α mattes. During training, in addition to the standard velocity prediction loss Lmse , we introduce an α -oriented reconstruction term Lrec . Reconstruction loss performs one-step denoising using the predicted velocity to reconstruct the clean latent state, and explicitly aligns the reconstructed α with the ground-truth matte, thereby significantly improving fine-grained transparency generation.
Figure 3: Comparison with various I2V generalist models on our text animation dataset. Each example shows the high-fidelity, motion-consistent RGB video (top row) and the corresponding transparent video (bottom row) generated by TransText . Zoom in for better details.
Method
Generation Metric
Alignment & Quality Score
Time Efficiency
FVD ↓
Soft α -mIoU ↑
RGBA Alignment (%) ↑
Motion Quality † ↑
Training ↓
Inference ↓
Wan * + ConText Zhang et al. (2025a)
480.92
57.24
13.51
29.55
53.44
7.92
Transpixeler * Wang et al. (2025)
673.11
43.76
83.01
64.52
72.31
11.25
TransText
123.48
79.90
87.31
75.94
69.72
10.11
Table 1: Comparison with existing RGBA-based methods on our validation set. TransText outperforms all baselines in generation quality with promising efficiency. * Fine-tuned on our training data. † Scores derived from a user study. Bold indicates the best performance, and underlined indicates the second best.
Figure 5
Figure 5: Qualitative results on background compositing. Visual effects include: Letter Collection , Flickering Green , Snow Falls , and Small Squares .
Figure 6: Qualitative results on background compositing. Visual effects include: Dark Fire , From the Sand , Lightning Strikes , and Blue Lights .
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
ID
Effect
Train
Val
1
From the sand
2,515
137
2
Flickering Green
2,471
127
3
Letters Collected
2,318
117
4
Small Squares
2,080
109
5
Lightning Strikes
1,992
99
6
Snow Falls
1,561
86
Appendix
Table 1: Distribution of 8 visual effects in the dataset. Below shows the sample counts for training (Train) and validation (Val) splits.
Figure 1: Visualization of original samples from our glyph animation dataset, with one example shown for each visual effect. The reference image (the column marked in blue) corresponds to the middle frame of each video clip. Please zoom in for a better view.
RGBA video generation is essential for creating compositable visual assets in filmmaking, game development, and digital content production. However, generating high-quality RGB appearance alongside temporally coherent transparency remains a significant challenge. Existing text-to-RGBA video methods often entangle RGB and alpha distributions and offer limited control over the precise appearance of a target asset. In addition, incorporating image conditioning typically suffers from the emergence-disappearance effects, which fail existing first-frame or first-and-last-frame conditions due to their fully transparent boundary frames. To address these issues, we present Wan-Alpha++, a unified framework that brings controllable RGBA video generation to both text- and image-to-video diffusion backbones through a shared RGBA latent space, extending our conference version Wan-Alpha from the text-conditioned setting to image-conditioned generation with an interior reference frame. Our method first learns shiftable RGB-alpha distributions while preserving the pretrained RGB prior. Specifically, we introduce a transparency-aware bidirectional diffusion objective to produce distinguishable RGBA latent representations, and a transparency-guided mean sampler that shifts the noise distribution to improve transparency generation and enable spatial controllability. Building upon this RGBA generation prior, we introduce temporal phase reparameterization for image-conditioned effect generation. By cyclically transforming a zero-positive-zero alpha pattern into a positive-zero-positive pattern, it relocates visible content to both temporal boundaries, facilitating effective appearance conditioning with limited RGBA data.
Haotian Dong, Xin Wang, Wenjing Wang +7
College of Intelligence and Computing, Tianjin University, China. · Laboratory for Artificial Intelligence in Design Limited (AiDLab), The Hong Kong Polytechnic University, Hong Kong. · independent researcher. +3
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.
Qijie Wang, Zhixin Xu, Zhe Li +5
Tsinghua University · ByteDance · Peking University
RGBA videos combine RGB appearance with an alpha channel, enabling animated assets to be applied across arbitrary backgrounds, which are heavily used in gaming industry. However, generating high-quality RGBA animations for games remains challenging for two reasons. First, most existing RGBA video datasets are dominated by photorealistic content, with limited coverage of game assets. Second, the traditional generate-then-matte pipelines estimate alpha only after RGB synthesis, so semi-transparent regions are often blurred by background, resulting in unstable matting outputs. More recently, many methods have begun to model RGB and alpha jointly, but existing approaches are mostly text-conditioned, and still have unresolved issues in efficiency and quality. To address these challenges, we introduce GameAlpha-2.4K, a 2.4K-clip game-style RGBA video dataset built with matte-friendly synthesis, multi-hypothesis alpha recovery, and compositing-based quality gates. Using this dataset, we train a reference-conditioned RGBA video generator that jointly produces RGB frames and alpha mattes in a single pass. To improve efficiency, we propose a visibility router that identifies transparent tokens in an early stage and bypasses their later DiT updates, while x_0-lock guides them along the original flow-matching schedule toward self-predicted endpoints. Our model obtains lower FVD than traditional two-stage pipelines, and the visibility router skips 35% of token evaluations in the final two DiT denoising steps, providing a 1.2x backbone speedup with negligible quality degradation compared to dense inference.
Zhe Li, Honghao Qiao, Zhixin Xu +3
Peking University · ByteDance · Tsinghua University