EchoShot: Multi-Shot Portrait Video Generation

We are so grateful that all the reviewers give our work positive initial ratings (4454), which is a tremendous encouragement for us. We particularly thank you for your insightful feedback and for highlighting the innovation of our work. For each of the questions, we provide a detailed clarification below.

---

Q1: Video Results

All videos in our supplementary materials were generated under the MT2V setting, which represents the core contribution of our work for multi-shot generation from textual descriptions. These videos are representative samples from our 100-prompt evaluation set, which is the primary basis for the quantitative and qualitative analysis in our paper. The prompts, featuring 2 to 4 shots, were generated by LLM (Gemini-2.5 Pro) to ensure diversity and reduce bias.

We would like to provide further details about the inference of two extensions. Throughout the inference process, the resolution is 480p and the total length is 8 seconds, the same as MT2V. The PMT2V and InfT2V models are tuned based on the MT2V model. For PMT2V inference of Figure 6, given an input image, we first instruct Gemini-2.5 Pro to describe the ethnicity, gender, age group of the person and generate multi-shot prompts based on the attributes. We then randomly decide a segmentation of shot lengths and crop the face region with a face detection model. Finally, we input the face crop, the shot-length segmentation and multi-shot prompts to PMT2V model to generate the personalized multi-shot video. For InfT2V inference of Figure 7, we instruct Gemini-2.5 Pro to generate 10 prompts ($p_1$-$p_{10}$) featuring the same person and sample 10 random noises $z_1$-$z_{10}$. $p_1, z_1$ is treated as the reference. We first input $z_1, z_2$ and $p_1, p_2$ to generate shots $s_1$ and $s_2$, using the InfT2V model. Then, we iteratively input $z_1, z_k$ and $p_1, p_k$ ($k\in$ 3, 10 $$) to generate the other eight shots.

Unfortunately, the rebuttal policy prevents us from supplying additional video files of Figures 6 and 7. Instead, we provide partial prompts used for Figures 6 and 7 (please see postscript). Besides, to enhance reproducibility and contribute to the community, we promise to publicly release our full evaluation set, the detailed evaluation methodology, and the visual results from EchoShot and all baseline models. We have established a clear open-sourcing timeline. In fact, to facilitate early access, we have already publicly released codes and weights. We believe this will significantly strengthen our work and be a valuable resource for future research.

Q2: Extension to Multi-Character

This is an excellent question that we have also discussed ourselves. Rethinking EchoShot, it is inherently adaptable to general multi-shot scenarios including single character, multiple characters and even objects. The consistency across shots can be captured by introducing specific semantic patterns as indicators. Then EchoShot can automatically learn the correspondence between these semantics and the generated visual content. For your consideration, here we propose one potential solution for the multi-character scenario: annotate each character in text prompts with a unique identifier. For instance:

[Shot 1] (PERSON 1: a man with black hair) is running in a park …

[Shot 2] (PERSON 1: a man with black hair) is walking on the beach with (PERSON 2: a young boy wearing a blue shirt) ...

[Shot 3] (PERSON 2: a young boy wearing a blue shirt) is talking to (PERSON 3: a woman in a white dress) ...

Then EchoShot could be trained directly on a dataset structured this way. We consider this a very promising direction for future work.

Q3: Typos

We will perform a thorough proofread of the entire manuscript to correct these and any other errors.

---

Thanks again for your efforts. We believe the rebuttal stage is highly meaningful, and we will add the rebuttal discussion to the new version of our paper as much as possible. We would be greatly grateful if you would consider raising your score in light of these responses.

---

Postscript 1: Prompts of Figure 6

Morgan Freeman

[Shot 1] This segment shows a black, male, elderly person. The character's hairstyle is short, grey hair. The character is wearing a knitted cardigan, paired with a shirt and a tie. The character's expression is serious, with focused eyes, as if they are listening attentively to the phone. The corners of the character's mouth slightly downward, revealing a contemplative emotion. The character's movements are sitting on a beige patterned sofa, holding a white vintage telephone receiver in their right hand and holding an open notebook in their left hand. The video scene is a luxuriously decorated room with dark brown wooden walls as the background. There is a table on the left with a white tablecloth. The video is lit in a low-key warm tone, with light mainly focused on the character's face and a relatively dark background, creating a mysterious and oppressive atmosphere.

[Shot 2] This segment shows a black, male, elderly person. The character's hairstyle is short, grey hair. The character is wearing a knitted cardigan, paired with a shirt and a tie. The character’s expression is relaxed and cheerful. The character's movement is sitting on a rocking chair, with their left hand gently resting on the armrest. The video scene is a sunlit porch with lush green grass and distant mountains in the background, creating a beautiful scenery. The video is illuminated with bright light, making the facial details and background of the characters clearly visible.

[Shot 3] This segment shows a black, male, elderly person. The character's hairstyle is short, grey hair. The character is wearing a knitted cardigan, paired with a shirt and a tie. The character’s expression is serene and calm, with a kind gaze and a gentle smile at the corner of their mouth, conveying a sense of tranquility and wisdom. The character's action is resting on a bench in the park, with his eyes closed and listening to the sounds around them. The video scene is a vibrant park with dense trees in the background. The video is illuminated with bright natural light, evenly illuminating the characters and surrounding environment, creating a relaxed and pleasant atmosphere.

Emma Watson

[Shot 1] This segment shows a Caucasian, female, young person. The character's hairstyle is shoulder length short hair, with a light golden color and a black thin headband. The hair is fluffy, slightly messy, and has a slight inward button at the end. The character is wearing a light blue shirt. The character's expression is happy and joyful. The character's movements are standing still, raising the head, looking up at the top of the screen and speaking. The video scene is a dark forest. There are tall trees in the background with mist and trees. The ground is covered with fallen leaves and some green plants. The overall lighting is dim, creating a gloomy and mysterious atmosphere. The light mainly comes from above, illuminating the face of the character.

[Shot 2] This segment shows a Caucasian, female, young person. The character's hairstyle is shoulder length short hair, with a light golden color and a black thin headband. The hair is fluffy, slightly messy, and has a slight inward button at the end. The character is wearing a deep red short sleeved dress and gold wire glasses. The collar of the dress is circular, and the sleeves reach the middle of the upper arm. The character's expression is slightly worried and uneasy. The character's movements are standing still and gazing downward. The video scene is a desolate place shrouded in mist. In the background, you can see some blurry mountain and forest. On the right side of the character, there is a tall stone pillar with a rough surface and some blurry patterns engraved on it. The lighting in the video is low-key, with overall dim lighting, creating a oppressive and mysterious atmosphere. The light mainly comes from above and in front, but the light is soft and there are no obvious shadows.

[Shot 3] This segment shows a Caucasian, female, young person. The character's hairstyle is shoulder length short hair, with a light golden color and a black thin headband. The hair is fluffy, slightly messy, and has a slight inward button at the end. The character is wearing a black V-neck dress. The character's expression is extremely sad and painful, with tears in their eyes, red eyes, and slightly parted lips. Their overall mood is low. The character's movements are standing still, with their head slightly raised, their eyes looking up, their lips slightly trembling. The video scene is indoors, with a light brown wall adorned with abstract tree patterns. The overall scene is simple and artistic. The video is illuminated with soft light, which is uniform and bright.

Postscript 2: Prompts of Figure 7

[Shot 1] This segment shows an Asian, female, and young person. The character's hairstyle is dark brown, with split hair and two thin braids tied at the back of the head. The ends of the braids are secured with black rubber bands. The character is a white high-neck shirt layered under a beige trench coat. The character’s expression is gentle and pleasant. The character is sitting and is in the process of drinking from a white ceramic coffee cup. The video scene is in a cafe, with a softly blurred background showing wooden furniture and green plants. The lighting is bright and soft, illuminating the subject evenly and creating a warm, inviting atmosphere.

[Shot 2] This segment shows an Asian, female, and young person. The character's hairstyle is dark brown, with split hair and two thin braids tied at the back of the head. The ends of the braids are secured with black rubber bands. The character is wearing a white t-shirt and a silver chain necklace, layered under a blue denim jacket ...

（Omitted due to text length limitations）

We are so grateful that all the reviewers give our work positive initial ratings (4454), which is a tremendous encouragement for us. We particularly thank you for your recognition and for highlighting the overall contributions of our work. For each of the questions, we provide a detailed clarification below.

---

Q1: Inference Runtime

The inference time of EchoShot is identical to the original Wan-1.3B model, because our proposed RoPE design introduces no additional computational overhead. To specify, generating a multi-shot video of 8 seconds in total length at 480p resolution requires 24GB of VRAM and takes approximately 360 seconds on a single NVIDIA A100 GPU.

Q2: Faces for PMT2V Training

We assume you are asking about the construction process of our PMT2V dataset. For each video clip, we employ the following multi-stage process to obtain the face:

• We use a face detection model to extract face crops from every frame.
• These crops undergo a rigorous filtering process based on multiple criteria, including face orientation, sharpness, and the relative size of the face within the frame. The filter is driven by in-house analysis tools.
• We randomly sample at most 5 face images from the filtered crops, ensuring a minimum temporal gap of 1 second between any 2 face images.

Q3: Face Consistency

Regarding the Emma Watson example in Figure 6, the perceived imperfect match between the generation and the input can be attributed to both objective and subjective factors.

• Objective Factor: The Inherent Trade-off between Identity and Motion. There is a common challenge in personalization tasks involving the trade-off between preserving identity and generating dynamic motion. A single reference image input provides limited information, making it difficult to perfectly model the full spectrum of a person's facial dynamics. Our model, trained on a diverse dataset, is designed to excel at generating a wide range of dynamic features, such as varied expressions and head poses. In the Emma Watson example, EchoShot generates complex facial dynamics (e.g., laughing, crying) while baselines fail to achieve. This strong emphasis on motion can lead to subtle deviations from the static input image, which is perceived as an imperfect identity match.
• Subjective Factor: Perceptual Sensitivity to Celebrity Faces. Human perception is more sensitive to inconsistencies in the faces of well-known celebrities. This is because we have strong prior knowledge of their appearance from various angles and in diverse contexts. We once conducted a small-scale comparative user study, where we provided videos generated by EchoShot based on celebrity faces and non-celebrity random faces. We found that participants generally perceived a higher degree of identity consistency when seeing generations of non-celebrity faces.

We believe this challenge can be effectively mitigated by incorporating multiple reference images to provide a more comprehensive identity representation. We consider this a promising direction for future work.

Q4: Video Result

We included video examples of EchoShot in the supplementary materials (zip file). We believe these results effectively demonstrate the strong temporal consistency and excellent realism achieved by our model. Besides, we promise to publicly release our full evaluation set and the visual results from EchoShot and all baseline models. Our goal is to establish the first comprehensive benchmark for multi-shot portrait video generation, which we believe will significantly strengthen our work and be a valuable resource for future research.

---

Thanks again for your efforts. We believe the rebuttal stage is highly meaningful, and we will add the rebuttal discussion to the new version of our paper as much as possible.

We are so grateful that all the reviewers give our work positive initial ratings (4454), which is a tremendous encouragement for us. We particularly thank you for the meticulous feedback and for highlighting the significance of our work. For each of the questions, we provide a detailed clarification below.

---

Q1: Generalization to Non-Realistic Styles

Though our training dataset predominantly consists of videos in a realistic style, we have observed that EchoShot exhibits an emergent capability to generate consistent shots in certain artistic styles, such as 3D animation. We are sorry that due to rebuttal policy, we cannot include the corresponding video examples. To explain this phenomenon, we hypothesize that EchoShot effectively inherits the versatile generative capabilities of the base mode (i.e., Wan-1.3B), benefiting from our proposed RoPE designs. However, due to the lack of explicit training on these artistic domains, the visual quality of these stylized generations is not yet on par with that of the realistic style. Furthermore, for styles that are beyond the capability of the base model itself, such as stop-motion animation, EchoShot is consequently unable to generalize.

Q2: Camera Control

Our dataset annotation and model generation capabilities focus on common camera movements, including static, handheld, pan left/right, tilt up/down, zoom in/out, limited-angle orbiting shot. Consequently, the model currently does not support more complex camera motions, such as 360-degree orbiting shots or fully customized camera paths. We acknowledge this as a current limitation and a valuable direction for future research.

Q3: Extension to Multi-Character

This is an excellent question that we have also discussed ourselves. Rethinking EchoShot, it is inherently adaptable to general multi-shot scenarios including single character, multiple characters and even objects. The consistency across shots can be captured by introducing specific semantic patterns as indicators. Then EchoShot can automatically learn the correspondence between these semantics and the generated visual content. For your consideration, here we propose one potential solution for the multi-character scenario: annotate each character in text prompts with a unique identifier. For instance:

[Shot 1] (PERSON 1: a man with black hair) is running in a park …

[Shot 2] (PERSON 1: a man with black hair) is walking on the beach with (PERSON 2: a young boy wearing a blue shirt) ...

[Shot 3] (PERSON 2: a young boy wearing a blue shirt) is talking to (PERSON 3: a woman in a white dress) ...

Then EchoShot could be trained directly on a dataset structured this way. We consider this a very promising direction for future work.

Q4: Video Quality

Artifacts do occasionally occur during generation, which we attribute to the limited parameters of the base model. Wan-1.3B is a very small model compared to peer models. Our internal tests confirm that original Wan-1.3B exhibits inferior performance to original Wan-14B in terms of dynamic quality and face details. Considering the computation cost, we decided to conduct this study on Wan-1.3B. Though, the method we proposed can be seamlessly applied to larger models. For future work, we plan to explore parameter-efficient fine-tuning methods (e.g., LoRA) to adapt EchoShot to the 14B model at a lower cost. Based on the current results, we are confident that migrating to a more powerful base model would achieve even higher quality.

Q5: Explanation of the Bottleneck Effect

We use the term the bottleneck effect to describe a phenomenon in keyframe-to-video synthesis pipelines. These cascaded pipelines typically consist of a consistent text-to-image (CT2I) model followed by an image-to-video (I2V) model, where the output of the former serves as the input for the latter. In this architecture, the quality of the final video generated by the I2V model is heavily constrained by the quality of the initial keyframes produced by the CT2I model. Thus, the performance of the CT2I model acts as a "bottleneck”, limiting the overall system's potential, regardless of the I2V model's inherent capabilities. This effect is empirically validated in Figure 5 of our paper. Both "StoryDiffusion+Wan" and "IC-LoRA+Wan" use the same I2V model (Wan-14B-I2V) and different CT2I models (StoryDiffusion and IC-LoRA), yet their resulting visual quality differs dramatically. This disparity arises because IC-LoRA, as a CT2I model, exhibits superior overall performance compared to StoryDiffusion. In this case, the capacity of Wan-14B-I2V is underutilized.

Q6: Ablation Study

Thanks for your advice. We will revise the paper to add critical ablation studies into the main part to strengthen our experimental analysis.

Q7: User Study Videos

The supplementary material (zip file) contains video examples generated by EchoShot for user study. We apologize that due to the policy, we are unable to provide videos generated by baseline methods in this rebuttal. To enhance reproducibility and contribute to the community, we promise to publicly release our full evaluation set, the detailed evaluation methodology, and the visual results from EchoShot and all baseline models. Our goal is to establish the first comprehensive benchmark for multi-shot portrait video generation, which we believe will significantly strengthen our work and be a valuable resource for future research.

Q8: Typos

We will perform a thorough proofread of the entire manuscript to correct these and any other errors.

---

Thanks again for your efforts. We believe the rebuttal stage is highly meaningful, and we will add the rebuttal discussion to the new version of our paper as much as possible. We would be greatly grateful if you would consider raising your score in light of these responses.

We are so grateful that all the reviewers give our work positive initial ratings (4454), which is a tremendous encouragement for us. We particularly thank you for the insightful feedback and for highlighting the strengths of our work. Our primary motivation is to contribute to the video generation community, so we are willing to take any action that enhances the openness and impact of our work. For each of the questions, we provide a detailed clarification below.

---

Q1: Evaluation Set

Thank you for highlighting the importance of an open evaluation set. We agree that this is crucial for enabling future work to compare against and build upon our method. Thus, we promise to release our full evaluation set, video results and evaluation pipeline.

To provide immediate clarity on our evaluation process, we would like to share some key details here. The Vision-Language Model (VLM) used for our automated scoring is Gemini-2.5 Pro. For the Prompt Controllability metrics, the VLM scores the generated videos across nine distinct aspects: ethnicity, gender, age group, hairstyle, apparel, expression, action, background, and lighting. The final scores are calculated as follows: App. score: the average of scores for ethnicity, gender, age group, hairstyle, and apparel. Mot. score: the average of scores for expression and action. Bg. score: the average of scores for background and lighting. For the Visual Quality metrics, the Sta. score and Dyn. score are directly scored by the VLM.

Due to the format limitations of the rebuttal, we can only provide a prompt example from our evaluation set and the scoring instructions input to Gemini-2.5 Pro (please see the postscript). The full benchmark will be released according to the timeline below.

Q2: Timeline for Open-Sourcing

We have established a clear open-sourcing timeline. In fact, to facilitate early access, we have already finished certain steps:

• Before June 15th, 2025: Launch project page and generation gallery. (Finished)
• Before July 15th, 2025: Publicly release codes. (Finished) 
• Before July 15th, 2025: Publicly release the model weight. (Finished)
• Before August 15th, 2025: Publicly release the full test set, evaluation details, and baseline benchmarks.
• Before September 30th, 2025: Publicly release the PortraitGala dataset.

We are committed to this schedule and believe it directly addresses the concerns regarding reproducibility.

Q3: Evaluation on Public Benchmarks

The primary reason for not including evaluation on public benchmarks was the lack of a public a large-scale dataset with many-to-one annotations (i.e., multiple video shots of the same person paired with consistent text descriptions), to the best of our knowledge. This very gap is exactly the reason why we construct and commit to releasing the PortraitGala dataset. Regarding your specific suggestion of WebVid, it consists of diverse scenes and is designed for general-purpose video generation. It lacks sufficient videos and annotations of the same person in different contexts. As a result, we are unable to conduct training or evaluation on WebVid in the multi-shot portrait generation setting. For this issue, we promise to release PortraitGala dataset, detailed evaluation methodology, and the videos generated by EchoShot and all baseline models. By doing so, we will establish the first public benchmark specifically for the multi-shot portrait generation task. We believe this contribution will markedly strengthen our work, provide a lasting resource for the community, and address your concerns about comparative evaluation.

---

We believe the rebuttal stage is highly meaningful, and we will add the rebuttal discussion to the new version of our paper as much as possible. We would be very grateful if you would consider raising your score in light of these clarifications.

---

Postscript 1: An Example of 100-prompt Evaluation Set

[Shot 1] Camera pans left, medium shot. This segment shows a Caucasian, male, middle-aged person. The character's hair is dark brown, medium length, voluminous, slightly fluffy, with natural waves, and some hair falls over the forehead and sides of the face. The character is wearing a light blue long-sleeved shirt, with the collar open, and a dark blue striped tie. The character's expression is serious, with a focused gaze, corners of the mouth slightly downturned, seemingly thinking or observing. The character's action is standing, body slightly angled to the left, head turned to the right, seemingly observing or listening, with hands naturally hanging down and no obvious body movements. The video scene is an office, with a light green background wall. On the left side, there is a whiteboard covered with names and school names, and on the right wall, a white bulletin board is hanging. The video lighting is generally dim, with light coming from the left, casting shadows on the character's face, creating a serious and oppressive atmosphere.

[Shot 2] Still shot, close-up. This segment shows a Caucasian, male, middle-aged person. The character's hair is dark brown, slightly long, layered, naturally combed back, with some hair falling on both sides of the forehead. The character is wearing a light blue and white striped shirt, with the collar open, revealing the collarbone. The character is wearing a dark blue tie with white polka dots. The character's expression is calm, slightly serious. The character's action is sitting on a sofa, body slightly leaning forward. The character occasionally makes gestures with both hands. The video scene is indoors, with a light green wall in the background, which has subtle textures. The scene is simply arranged, without excessive decoration. The video lighting is soft natural light, with light coming from the front, illuminating the character's face and upper body. The light is evenly distributed, with no obvious shadows. The overall lighting conditions are relatively bright.

[Shot 3] Still shot, close-up. This segment shows a Caucasian, male, middle-aged person. The character's hair is dark brown, slightly fluffy, naturally combed back, with slightly longer sideburns. The character is wearing a light blue Hawaiian shirt, with white and dark blue patterns printed on it, and the collar slightly open. The character's expression is relaxed and slightly smiling, with a focused gaze, seemingly listening or thinking. Occasionally, the corners of the mouth subtly turn up, showing a sense of calmness and confidence. The character's action is slight head movements, subtle lip movements, seemingly speaking, with the body remaining relatively still, shoulders relaxed, and hands possibly placed on a table or in front of the body. The video scene is an indoor office, with a blurred background. Some stacked files and office supplies are visible. The video lighting is soft natural light, with light coming from the front, illuminating the character's face. Shadows are not obvious, and the overall light is even, creating a comfortable and relaxed atmosphere.

Postscript 2: Rating Instruction for Prompt Controllability Metrics

Please play the role of a video generation model evaluator. I will give you a prompt and a corresponding generated video clip. Please judge whether the content of the video clip meets the description of the prompt in terms of ethnicity, gender, age group, hairstyle, apparel, expression, action, background, and lighting. If it fully meets the description, output 100, if it does not meet the description, output 0, and if it partially meets the description, give a score between 0 and 100 according to the degree. Each aspect should be judged independently, and the entire content of the prompt must be strictly considered. Only give the results, written in the form of a python dictionary, with the key being the evaluation aspect and the value being the score.

Postscript 3: Rating Instruction for Visual Quality Metrics

Please play the role of a video generation model evaluator. I will give you a prompt and the corresponding generated video clip. Please rate the video, which is divided into two scores, static and dynamic, each from 0 to 100. The static dimension (single-frame picture quality) is considered from the aspects of clarity and sharpness, color performance, detail restoration ability, noise and artifact control, composition and aesthetics, resolution consistency, etc. The dynamic dimension (continuous picture performance) is considered from the aspects of motion smoothness, motion blur processing, timing consistency, dynamic range and brightness changes, action realism, scene transition and occlusion processing, dynamic detail retention, etc. Only give the results, written in the form of a python dictionary, with the key being the evaluation aspect and the value being the score.