MoCha: Towards Movie-Grade Talking Character Generation

Dear reviewer qDos,

We sincerely appreciate the time and effort you have dedicated to reviewing our work and providing valuable feedback. Below, we are pleased to share our detailed responses.

W1&L1 Use the test images from the MoCha are likely out-of-domain for the baselines. Q2 & Q3 &L2: Recommended to compare MoCha directly with other t2v approaches and the base model

Answer: We thank the reviewer for raising this important point! Our work is dedicated to the novel task: talking characters. The experiments were designed to compare with existing methods that can only do talking head generation. To follow their inference settings, we used the first frame from MoCha as input. While we cropped the frames to match baselines' training distribution and avoid blurry or artifacts, we acknowledge that potential distribution mismatch may still exist. We conducted further experiments using three diverse groups of inputs to test the robustness of baselines.

Group 1: 3–5 real images from the Internet matching the prompts (e.g., “woman driving a car”), with clear head poses and scene alignment.

Group 2: 5 synthetic images generated by CogView4, GPT-Image-1, and FLUX.1-dev.

Group 3: MoCha-generated first frame.

Key Observations: (1)Baseline performances are robust across groups, showing MoCha-generated frames are not out-of-domain for them. (2)MoCha significantly outperforms all baselines, confirming its superiority is not due to biased inputs.

In response to Q2 & Q3 &L2. We thank the reviewer for the suggestions! and compared MoCha with other t2v models as required. We observed that since mocha is trained on talking human videos, it achieves the best facial expression realism. It can handle complex prompts like "speaking while gently sipping water." while other models fail to produce talking behavior in this case. This comparison further supports our claim: MoCha bridges the gap between Movie Generation and Talking-Head Synthesis, making cinematic storytelling possible.

W2 & Q1 & L3: Lack of Discussion on the uniqueness of audio conditioning and multi-stage training strategy

We provide discussions below, followed by supporting results.

1. Localized audio attention Most talking head generation models (EMO, Hallo) employ 2D Unet model and align audio and video features on a per-frame basis. DiT utilizes 3D VAEs with temporal compression of video frames, creating a resolution mismatch between audio and video features. Notably, no existing work (Hallo3, omnihuman) thoroughly studies how to align audio and video within a DiT framework effectively. In Appendix Section A.1, we are among the first to conduct a comprehensive ablation on this, evaluating (1) Naive Audio Cross Attn (2) Naive Cross-Attn with positional embeddings (RoPE, learnable, sinusoidal) (3) Our proposed localized audio attention. Our method consistently outperforms all variants. This study offers valuable insight for future research. Furthermore, we compare it to Hallo3’s alignment strategy. Hallo3 temporally compresses audio features twice via stacked convolutions to match the video tokens. We find that this compression degrades fine-grained audio details, making it challenging to render characters accurately pronouncing phonemes like /iː/, /ai/, /tʃ/, and /θ/. In contrast, our method avoids temporal compression. We constrain each video token to only attend to a local window of audio tokens, which preserves the fine-grained details of the audio signal.

We provide a direct comparison of both methods on a DiT-T2V-4B backbone. Our approach results in superior audio-visual synchronization.

2. Training Strategy Our training strategy contains two parts Shot-Type-Based Curriculum + Mixed-Modal Sampling.

Existing works(hallo, hallo3, omnihuman) do use multi-stage training. However, they use it to add auxiliary signals, gradually introducing them (face image → audio → pose) to the model. At the audio stage, they simply feed all audio-annotated data into the model. This works fine as these models' tasks are talking heads generation, and the data are mostly cropped human face videos with similar face sizes. But for MoCha, our task and data are much more complex, involving talking faces→half-body gesture→full-body movements→camera control→multi-character, multi-clip. As the visual scope increases, the mouth region becomes smaller and harder to supervise. Training on all data types simultaneously to learn audio control is clearly inefficient, especially since MoCha doesn't use strong control signals like reference frames. To tackle this, our Shot-Type-Based Curriculum is not used to add new signals like existing work, but instead structures the learning process based on shot complexity. This method significantly improves convergence, as shown in Appendix Table 5.

SoTA methods(including hallo3, omnihuman) struggle with generating characters that go beyond static poses and fixed camera views. They fail to generalize to realistic cinematic scenarios, such as characters talking while expressing emotions, naturally performing actions, interacting with objects, or moving alongside the camera in response to text conditions. We identify two reasons (1) They overlook the richness of text conditions, often reducing prompts to oversimplified inputs like “a person is talking.” (2)They rely exclusively on human data, sampled primarily from audio/pose-driven datasets. To directly address this, we propose Mixed-Modal Sampling: We always provided high-quality and multi-grained captions to the model (as shown in Figure 2); During training we consistently sample 20/% of the data from a diverse text2video dataset.

This method enables MoCha to achieve superior generalization in scenarios involving complex lighting, human-animal interactions, and unlock the universal controllability via prompts eventually unlock the multi-character turned based converstation ablity.

We provide the following experiment against the strongest baseline Hallo3. The base models of Hallo3 and MoCha-4b have similar size and on-pair performance initially. After learning the audio condition, MoCha-4b outperforms Hallo3 on all axes. Hallo3 suffers from worse generalization, as evidenced by its degraded performance in text alignment and action naturalness compared to its base model. This underscores the uniqueness of our contribution.

W3: The method uses a large, strong base video generation model differs from the others

As shown in the tables above, our proposed methods consistently outperform baselines, even when using base models of similar size and capability. This demonstrates the effectiveness of our approach, not limited to large models. Our research goal—as stated in the claims—is to bridge the gap between generic video generation and constrained talking-head synthesis. We also want to deliver the message that movie-grade storytelling can be achieved without relying on auxiliary prior signals. However, as shown in the table above, base models with 4B/5B parameters still fail to consistently generate videos with a score of 3, meaning they cannot reliably produce mostly artifact-free and prompt-following results. Therefore, to support our goal, we intentionally build on top of a large base model to ensure satisfactory visual fidelity in the results.

L3: The method’s contributions are unclear.

We would like to clarify and highlight the key contributions of our work: MoCha is among the first to bridge the gap between video generation and talking-head synthesis, making cinematic storytelling possible. We introduced a new task —talking characters, along with a corresponding benchmark and a promising model for this task. MoCha is also among the first to demonstrate that movie-grade storytelling can be achieved without relying on auxiliary prior signals. Finally, it is the first to support multi-character, turn-based dialogue clips generation. A capability not achieved by any existing method to date.

Dear Reviewer qDos,

We are encouraged that you find our response has resolved most of your concerns, and we greatly appreciate the opportunity to continue this discussion.

On Multi-stage Training:

Thank you for your agreement with our proposed strategy — “Given the impressive results, I agree that this is a fair approach.” To further clarify its necessity. We analyzed Hallo3’s training data captions. Among the 101,543 captions used in Hallo3, 33,909 (33.39%) exactly match the phrase "a person is talking." In contrast, Mocha’s training strategy provides detailed captions including emotions, body language, actions, and camera movements and multiclip descriptions. We also consistently sample 20% of the data from a diverse text-to-video dataset. This helps Mocha unlock the ability to control the person beyond a frozen pose or a fixed camera-facing view — enabling movement and natural interaction with the environment and objects that are not present in audio-annotated data.

On the Localized Audio Attention:

We would like to clarify the motivation and novelty of our design.

The core challenge we address is the mismatch in resolution between audio and video tokens in DiT-based models. Suppose the video frames $f_1,f_2,f_3,f_4$ correspond to audio features $\alpha_1,\alpha_2,\alpha_3,\alpha_4$ respectively:

$f_1↔\alpha_1$; $f_2↔\alpha_2$; $f_3↔\alpha_3$; $f_4↔\alpha_4$

For EMO, which is built on a 2D U-Net and does not compress frames, there is no mismatch issue — each frame aligns directly with its corresponding audio feature. Previous U-Net-based approaches studied whether to augment $\alpha_i$ with adjacent features —just as the reviewer mentioned, “each frame attends to audio tokens near its timestamp”:

$f_1$ ↔ {$\alpha_1,\alpha_2$}

$f_2$ ↔ {$\alpha_1,\alpha_2,\alpha_3$}

However, in DiT-based models, due to 3d vae, 4 or 8 adjacent frames are compressed into a single video token: $f_1,f_2,f_3,f_4→x_1$

This breaks fine-grained alignment. Our study focuses on how to recover the fine-grained alignment, e.g., how $f_1 ↔ \alpha_1$ can still be preserved under this compression.

Essentially, the question becomes: How do we align {$x_1,x_2, …$} to {$α_1,α_2,α_3,α_4,α_5,α_6,α_7,α_8, …$}?

This challenge is different from the one addressed in EMO, which explores whether $f_i$ should be augmented with features from $f_{i+1}$.

This is a critical question for dialogue-driven video generation using DiT, as most modern video generation models are based on DiT. However, it has not been thoroughly explored in existing work

We raise the following research questions:

1. Should we compress the audio features to enforce one-to-one alignment (as in Hallo3)?

   {$α_1,α_2,α_3,α_4$} → $α_1*$ via convolution.

   then $x_1$ and $α_1*$ are directly aligned i.e. $x_1↔α_1*$

2. Should we keep audio features in their original resolution? (Naive Audio Attention in Appendix)

   {$x_1,x_2, …$} ↔ {$α_1,α_2,α_3,α_4,α_5,α_6,α_7,α_8,…$}

3. If we keep the audio features in original resolution:

   • Does positional embedding help alignment? (Audio Attention with Positional Embedding in Appendix)

     Ideally, as in LLMs, tokens can learn where to attend. Can the $x$ tokens similarly learn to use positional information to extract relevant audio tokens? Which type of positional embedding is more effective — absolute (sinusoidal), learnable, or relative (RoPE)?

     {$x_1$} ↔ learns to attend to {$α_1,α_2,α_3,α_4$}

     {$x_2$} ↔ learns to attend to {$α_5,α_6,α_7,α_8$}

   • Does a strong inductive bias like windowed attention help alignment? (Our proposed method)

     {$x_1$} ↔ attends only to {$α_1, α_2, α_3, α_4$}

     {$x_2$} ↔ attends only to {$α_5, α_6, α_7, α_8$}

Our findings show that the proposed window-based cross-attention achieves the best results among all variants. This focus is valuable for future research on building dialogue-driven video generation models in the DiT framework. Our study carefully and systematically explores many previously unaddressed directions and provides important insights and lessons for future work.

Dear Reviewer qDos,

Thank you for your suggestion and thoughtful discussion! In the final version, we will revise the abstract, introduction and method sections to focus the main contribution on the multi-clip and multi-turn generation according to the suggestions.

On the Novelty of MoCha:

We would also like to highlight the novelty of the MoCha framework itself and also draw attention to Section 3.3: Multi-character Turn-based Conversation.

1. MoCha framework is novel as it targets the new task: text+audio2video generation. Distinct from all existing work relying on separate stages or auxiliary inputs, MoCha's promising results demonstrate that vivid talking characters can be generated from scratch directly using text to define a scene and audio to control facial movements. We don't need a pre-generated first frame or hand pose. We don't need any facial masks during training. This flexibility provides new insight to the community.

2. In Section 3.3: Multi-character Turn-based Conversation. We introduce a novel method that employs character-tagged prompts and self-attention across video tokens to generate multi-clip in parallel and enables, for the first time multi-speaker interaction. This is a critical component for generating complex, movie-style scenes, as turn-based conversations between multiple characters are essential in cinematic storytelling— a novel capability not achievable with prior stage-based or mono-speaker approaches.

In addition to

3. MoCha's strong audio-video synchronization results from our careful and unique study on aligning video and audio tokens within the DiT framework. Our novel window-based audio attention mechanism makes this synchronization possible.

4. MoCha achieves universal controllability via natural languages and realistic character movements through a novel designed training strategy.

Dear reviewer gs1v,

We sincerely appreciate the time and effort you have dedicated to reviewing our work and providing valuable feedback. Below, we are pleased to share our detailed responses to your comments.

W1 The method has not been open-sourced, nor are there plans for open-sourcing it.

As authors, we are fully committed to open-source research. We have implemented MoCha on top of the open-source backbone Wan2.1 and will include a link to the code in the final revision. We have already included technical details such as model parameters, the full data processing pipeline, and example captions in the submission. We will further enrich the final version with as many details as possible to benefit the community.

W2 The five steps mentioned in Curriculum Training lack ablation studies for each stage, making it seem unnecessary to design such a complex process.

Answer: Thank you for the suggestions. Below, we provide both explanation and supporting results to justify the necessity of our curriculum design.

Traditional talking-head generation tasks typically use cropped face videos with limited variability. In contrast, MoCha tackles a significantly more complex task, involving: talking faces→half-body gesture→full-body movements→camera control→multi-character, multi-clip. As the visual scope increases, the mouth region becomes smaller and harder to supervise. Training on all data types simultaneously is highly inefficient for learning audio-driven control—especially since MoCha does not rely on strong control signals like reference frames. To tackle this, our Shot-Type-Based Curriculum is used to structures the learning process based on shot complexity. This method significantly improves convergence.

We provide an ablation study of different stages using the 4B version, trained for the same number of steps (20K) and evaluated on MoChaBench. Stage 0 corresponds to T2V pretraining, and Stage 1 introduces the audio signals, which are indispensable for the task. Therefore, our ablation focuses on the contributions of Stage 2 and Stage 3

Analysis: Medium close-up data introduces information such as co-speech gesture and the character's overall motion. Joint training on close-up and medium close-up data shows significantly degraded results in lip-sync metrics compared with curriculm training, as the facial region in medium close-up data is relatively smaller, providing limited supervision for the audio-conditioning module and leading to slower learning.

Analysis: Medium shot data focuses on full-body actions and camera control. Joint training with all shot types at once further degrades lip-sync performance and text alignment. In contrast, curriculum training gradually builds the model’s ability to handle increasing complexity, resulting in balanced performance across metrics. While lip-sync scores are slightly lower than the previous stage, gains in action naturalness and text alignment are substantial. These ablation studies collectively demonstrate the necessity and effectiveness of each stage in our proposed curriculum.

W3: Consider freezing most of the network modules during training. L1: The experimental overhead is too high

Answer: We thank the reviewer for raising this important question.

To address concerns about high resource consumption, we provide the following experiment to show that our proposed training strategy and localized audio attention methods work for small models. The base models of Hallo3 and MoCha-4b have similar sizes and comparable initial performance. After learning the audio condition, MoCha-4b outperforms Hallo3 across all axes. Hallo3 suffers from worse generalization, as evidenced by its degraded performance in text alignment and action naturalness compared to its base model.

This demonstrates that the effectiveness of our approach is not limited to large models requiring high resource consumption.

Our research goal—as stated in the claims—is to bridge the gap between generic video generation and constrained talking-head synthesis, which requires a promising talking character model that can sync well with speech and perform actions naturally with overall good visual quality.

However, as shown in the table above, base models with 4B/5B parameters still fail to consistently generate videos with a score of 3, meaning they cannot reliably produce mostly artifact-free and prompt-following results. Therefore, to support our goal, we conduct the research on top of a large base model.

We have also conducted experiments during research where we only trained the newly introduced audio attention module while freezing the remaining backbone. We provide the results as follows:

We observed a dramatic degradation in model performance across all axes when training only the audio cross-attention module, compared with full training (MoCha) for the same number of steps. This observation differs from talking-head tasks, where training only the newly added module is often more effective than full fine-tuning. We hypothesize that since talking-head tasks always include a first-frame/pose condition—a strong conditioning signal—they converge more easily. In contrast, talking character generation is a more complex task. Thus, we consider full training of the model to be more effective in our setting.

The resulting model, MoCha, is among the first to demonstrate that movie-grade storytelling can be achieved without relying on auxiliary prior signals, and the first to support multi-character, turn-based dialogue clip generation, which bridges the gap between generic video generation and constrained talking-head synthesis. Given the promising results of MoCha, we believe it paves the way for future research on smaller yet stronger models (possibly distilled from large models) to extend talking character generation to broader applications.

Dear reviewer oyFP,

We sincerely appreciate the time and effort you have dedicated to reviewing our work and providing valuable feedback.

We are encouraged that you find our paper to be "well written" and our proposed methods to be significant and novel."

Below, we are pleased to share our detailed responses to your comments.

W1 Please include some sections from the appendix to the main work in the final version

Thank you for the suggestion. In the final version, we will move the Section A ablations from the appendix to the main paper to ensure clarity and completeness

W2 No mention of intention to open source the code, data and checkpoints. Though this is not compulsory but recommended.

As authors, we are fully committed to open-source research. We have implemented MoCha on top of the open-source backbone Wan2.1 and will include a link to the code in the final revision. We have already included technical details such as model parameters, the full data processing pipeline, and example captions in the submission. We will further enrich the final version with as many details as possible to benefit the community.

Q1 The authors have presented three good limitations in the appendix. What are some possible ways to resolve these problems and further improve the method?

For the limitation: "If the caption is too vague and doesn't describe the facial attribute or shot type, the model may generate a wide shot where the character is far from the camera, making lip sync difficult to observe." We can fine-tune the model via SFT on a curated dataset. We could collect a dataset of (caption, preferred shot framing) pairs where vague captions are matched with preferred video compositions (e.g., close-up shots for talking scenes). Use this to fine-tune the model to better infer appropriate shot types even when explicit cues are missing. For instance, given "A man playing skateboard at a skatepark" with accompanying speech, the model could learn to prioritize closer shots for improved lip sync alignment. We could also deploy annotators to rate generated videos on lip sync clarity and framing quality. Use these ratings to train a reward model and apply reinforcement learning or preference-based learning to optimize generation toward human-preferred outputs. This would teach the model to implicitly prioritize facial visibility when speech is present, even from vague captions.

For the limitation : "When multiple characters appear in the same scene, sometimes the lip-sync quality is degraded when the character is far from the camera, potentially due to limited training data." When multiple characters are present, the model may struggle to identify which character the speech corresponds to, based solely on the text. To mitigate this, we could incorporate character embeddings alongside speech embeddings to provide clearer grounding. For example, a set of learnable embeddings could be added to the speech embeddings to indicate the target character. This may help the model better associate the audio with the correct character and improve lip sync quality, particularly in multi-character scenes.

For the limitation :"When increasing the speech CFG from the default value of 7.5 to 12, the model tends to generate overly expressive characters. For example: 'A tracking shot circling around the man as he ties a tie over his blue suit. He speaks to the camera while adjusting the knot, maintaining eye contact throughout.'" To mitigate this, we could rescale the norm of the guided prediction at each denoising step to match the original conditional signal. This helps prevent exaggerated behaviors caused by excessively high CFG values.

Q2 What is the total number of training hours and what kind of GPUs were used for training?

Our experiments were conducted on H100 GPUs, totaling 49,152 GPU hours.

Dear reviewer Mvje,

We sincerely appreciate the time and effort you have dedicated to reviewing our work and providing valuable feedback. Below, we are pleased to share our detailed responses to your comments.

W1 Localized Audio Attention seems has been explored in Hallo3

To address the reviewer’s concern, we present a design comparison between Hallo3 and MoCha. Hallo3 applies stacked convolutions twice to temporally compress audio features, aligning them with the 3D VAE’s compression ratio used for video frames. However, we find that such compression degrades fine-grained audio details. In real-life, dialogue-driven movie generation scenarios, mouth movements can be extremely rapid, and accurately reproducing phonemes such as /iː/, /ai/, /tʃ/, and /θ/ in quick succession requires high-resolution audio features. To preserve these fine-grained details, MoCha avoids temporal compression of audio features. Instead, we propose to constrain each video token to attend to a local window of high-resolution audio tokens, enabling precise audio-visual alignment while retaining rich temporal fidelity in the audio signal.

We provide a direct comparison of both methods on a DiT-T2V-4B backbone:

Key Observation: Our novel localized audio attention approach results in superior audio-visual synchronization. Additionally, as shown in visual demos, Hallo3 often exhibits articulation errors and lip-sync artifacts when generating challenging phonemes like /iː/, /ai/, /tʃ/, and /θ/. In contrast, MoCha captures fine-grained articulation accurately and surpasses all baselines by a large margin.

Furthermore, we acknowledge that Hallo3(Section 3.2) studies 3 audio-injection strategies: (1) self-attention, (2) adaptive-norm, and (3) cross-attention, with cross-attention proving to be the most effective. This is indeed a valuable insight! However no existing work thoroughly studies how to effectively align audio features with temporally compressed video tokens within the cross-attention framework. Our work addresses this gap directly. As detailed in Appendix Section A.1, we conduct a comprehensive ablation study comparing:(1) Naive Audio Cross Attention (2) Naive Cross-Attn with various positional embeddings (RoPE, learnable, sinusoidal) (3) Our proposed localized audio attention. Our method consistently outperforms all variants. This study provides valuable useful for future research on audio-visual alignment.

W2 Address the distinctions between proposed curriculum-based multimodal training strategy and that of OmniHuman

We provide an explanation below, followed by supporting results.

Our training strategy contains two parts Shot-Type-Based Curriculum + Mixed-Modal Sampling.

We agree that existing works(omnihuman) do use multi-stage training. However, they use it to add auxiliary signals, gradually introducing them (face image → audio → pose) to the model in different stages (omnihuman Section 4.1). At the audio stage, they simply feed all audio-annotated data into the model. This works fine as these models' tasks are talking heads generation, and the data are mostly cropped human face videos with similar face sizes. But for MoCha, our task and data are much more complex, involving talking faces→half-body gesture→full-body movements→camera control→multi-character, multi-clip. As the visual scope increases, the mouth region becomes smaller and harder to supervise. Training on all data types simultaneously to learn audio control is clearly inefficient, especially since MoCha doesn't use strong control signals like reference frames. To tackle this, our Shot-Type-Based Curriculum is not used to add new signals like existing work but instead structures the learning process based on shot complexity. This method significantly improves convergence, as shown in Appendix Table 5.

SoTA methods(including hallo3, omnihuman) struggle with generating talking characters that go beyond static poses and fixed camera views. They fail to generalize to realistic cinematic scenarios, such as characters talking while expressing emotions, naturally performing actions, interacting with objects, or moving alongside the camera in response to text conditions. We identify two reasons (1) They overlook the richness of text conditions, often reducing prompts to oversimplified inputs like “a person is talking.” (2)They rely exclusively on human data, sampled primarily from audio/pose-driven datasets. To directly address this, we propose Mixed-Modal Sampling: We always provided high-quality and multi-grained captions to the model (as shown in Figure 2); During training we consistently sample 20/% of the data from a diverse text2video dataset. This method enables MoCha to achieve superior generalization in scenarios involving complex lighting, human-animal interactions, and unlock the universal controllability via prompts, eventually unlock the multiple-character turned based converstation-a capability not supported by any existing method.

We provide the following experiment against the strongest baseline Hallo3. The base models of Hallo3 and MoCha-4b have similar size and on-pair performance initially. After learning the audio condition, MoCha-4b outperforms Hallo3 on all axes. Hallo3 suffers from worse generalization, as evidenced by its degraded performance in text alignment and action naturalness compared to its base model. This highlights the effectiveness of our proposed training strategy and underscores the uniqueness of our contribution.

W3 It is unclear how the reference images are selected for these baselines in the comparison

Answer: We would like to emphasize that our work is dedicated to the novel task: talking characters. The experiments were designed to compare with existing methods that can only do talking head generation, which are all I2V models. To follow their inference settings, we used the first frame of the video generated from MoCha as input. We manually inspect each frame and avoid blurry or artifacts. The frames also contain clear facial details, making them suitable as reference images. We then crop and resize the frame to match each baseline's training distribution and ensure the head region is centred. Despite these efforts, we acknowledge that the first frame from MoCha may still fall outside the baselines’ training distributions. To address this, we conducted further experiments using three groups of inputs:

Group 1: 3–5 real images from the Internet matching the prompts (e.g., “woman driving a car”), with clear head poses and scene alignment.

Group 2: 5 synthetic images generated by CogView4, GPT-Image-1, and FLUX.1-dev.

Group 3: Original MoCha-generated frames.

These groups test the robustness of baselines across diverse inputs.

Key Observations: (1)Baseline performances are robust across groups, showing MoCha-generated frames are not out-of-domain for the baselines. (2)MoCha significantly outperforms all baselines, confirming its superiority is not due to biased inputs.

W4 How the proposed method could be adapted to reference-conditioned settings.

Reference-conditioned generation is a relatively well-studied area, with many existing methods achieving strong results. MoCha can be extended using approaches like FullDiT—encoding the reference image via the VAE and concatenating it with video tokens for 3D self-attention processing, or following the Movie Gen's Personalization method, concatenating the reference image embedding to the text embedding and further being processed through cross-attention. Meanwhile, dialogue-driven video generation area was limited to talking-head synthesis, the main research focus of this paper is to bridge the gap between generic video generation and talking-head synthesis, making cinematic storytelling possible. We introduced a new task called talking characters, along with a corresponding benchmark and a promising model. MoCha is among the first to demonstrate that movie-grade storytelling can be achieved without relying on auxiliary prior signals, and the first to support multi-character, turn-based dialogue clip generation. Given the promising results of MoCha, we believe it paves the way for future research on unified models that generate talking characters guided by reference images.

Dear Reviewer Mvje,

We sincerely appreciate the time and effort you have dedicated to reviewing our work, as well as your constructive comments.

We are encouraged that you found our paper fills a clear gap in automating cinematic storytelling and content creation.

We have submitted point-by-point responses to your comments. As the discussion phase is nearing its end. We hope our responses have addressed your concerns, and we would be happy to clarify anything further if needed.

Best regards,

The Authors