DreamFactory: Pioneering Multi-Scene Long Video Generation with a Multi-Agent Framework

• Long multi-scene videos remain challenging because character portrayal, style, and backgrounds must stay consistent while production requires more than assembling short generated clips.Source: §1 · p. 1
• DreamFactory simulates a film-production team whose LLM agents collaborate on scripts, storyboards, character design, keyframes, and video synthesis.Source: §1 · p. 2
• The authors evaluated generation using UTF-101 and HMDB51, compared framework videos with outputs from the original tools, and also used Inception Score and human judges on internet videos.Source: §1 · p. 2
• DreamFactory videos surpassed the average quality of manually produced videos, while the framework also outperformed native models on the reported evaluation mechanisms.Source: §1 · p. 2

• Multi-agent LLM work includes iterative methods such as CoT (Wei et al., 2022), ToT (Yao et al., 2024), and ReACT (Yao et al., 2022), alongside systems such as MetaGPT (Hong et al., 2023) and ChatDev (Qian et al., 2023).Source: §2 · p. 2
• Video synthesis has shifted from GAN-based methods (Brooks et al., 2022) toward diffusion processes and latent video diffusion models.Source: §2 · pp. 2–3
• The paper identifies Pika, Stable Video, Runway, and Sora as four leading video-generation models.Source: §2 · p. 3

3.1 Role Definition

• Role cards assign production roles—including director, screenwriter, filmmaker, and reviewer—whose tasks include choosing style, writing scripts, and drawing.Prompts specify each role’s job, task, and requirements, and discussions are bounded by a configurable number of rounds.Source: §3.1 · p. 3

3.2 DreamFactory Framework pipeline

• The pipeline organizes work into phases and conversations, with phase outputs retained as local memory for later stages.A phase processes textual or pictorial input through roleplay and collaboration; later phases can reuse saved conclusions such as styles and scripts.Source: §3.2 · pp. 3–5
• DreamFactory decomposes long-video creation into stages, from task, style, story, and script work to keyframe design.Agents use designated roles, and each phase shares a prompt defining its roles, tasks, desired conclusions, discussion format, and requirements.Source: §3.2 · pp. 3–4

3.3 Keyframe Iteration Design

• Long-term consistency across scene transitions requires preserving drawing style, character continuity, and narrative flow despite LLM randomness and drift.The Keyframe Iteration Design Method turns this challenge into successive short-term generation steps; the first frame establishes style, technique, characters, and background as the Base.Source: §3.3 · p. 5
• A Monitor reviews each generated frame, while later keyframes use the prior scene description and Base Description to maintain adjacent-scene continuity and long-distance visual consistency.The Base Description records features such as style, background, and character traits for preservation; the authors report that modified prompts produced a consistent, practically valuable image series.Source: §3.3 · pp. 5–6

4.1 Traditional Video Quality Evaluati1on

• FID, IS, and CLIP Score assess generated-image quality or text alignment, while FVD and KVD compare real and synthesized video feature distributions.The paper describes FVD as extending FID to videos and KVD as using a kernel function.Source: §4.1 · p. 6
• The Regular phase used 70 COCO-derived prompts, while the Script phase used 70 dataset tasks to assess script filling and the Animation Department.The DreamFactory label denotes framework-produced keyframe images corresponding to the scripts.Source: §4.1 · p. 6
• The framework's image quality stands out in Text2Image, and Table 2 reports a similar enhancement for multi-scene long-video generation.The paper also reports that script-generated images are more refined on average than images generated from everyday prompt words.Source: §4.1 · p. 6, Table 1 · p. 6, Table 2 · p. 6

4.2 Multi-scene Videos Evaluation Scores

• The Cross-Scene Face Distance Score targets facial-feature consistency across scenes, averaging face similarities between each keyframe and subsequent frames.The scoring experiment located 68 facial landmarks, encoded face regions with OpenAI CLIP's ViT, and used vector dot products; reported scores were predominantly above 0.5.Source: §4.2 · pp. 7–8
• The Cross-Scene Style Consistency Score is the share of keyframes assigned to the most frequent style category by GPT-4V.In the example, three of four scenes were classified as realistic and one as anime, yielding a 75% score.Source: §4.2 · p. 7
• Average Key-Frames CLIP Score averages each keyframe's CLIP score against its scene prompt to measure alignment with the script.The authors motivate this measure by noting that added consistency information could lead generation away from the text.Source: §4.2 · pp. 7–8
• The cross-scene comparison used 70 character-centric dataset entries generated by DreamFactory, GPT-4, and DALL-E 3, against a DALL-E 3 script-input baseline.The evaluation covered face distance, style consistency, and average CLIP Score; the supplied passage does not report the table's values.Source: §4.2 · p. 8

• DreamFactory combines multi-agent video generation with keyframe iteration to align style, characters, and scenes across frames.The framework proposes metrics for cross-scene face and style consistency and text-to-visual alignment.Source: §5 · p. 8

• Prompt-controlled agents are not capable of highly creative plotting without specific datasets for fine-tuning.The authors also note that speech-centered segment editing can make the result appear assembled from clips, and video generation consumes substantial resources.Source: §6 · p. 9

• The authors propose watermarking, fact-checking collaboration, transparency, ethical-use guidelines, and attention to diversity in training data as responsible-use measures.They also call for community engagement to establish ethical standards for AI-generated video.Source: §7 · p. 9

A.2 DreamFactory Responsibility allocation

• LLMs take roles such as CEO, Director, and Creator, then collaborate through social interaction to create a video from a story.The framework is described as a simulated animation company.Source: §A.1 · p. 12
• Small video-generation models act as tools that let the LLM team accomplish a larger video-generation task.Source: §A.1 · p. 12
• The collaboration can produce a series of consistent, stable multi-scene videos as the plot progresses.Source: §A.1 · p. 12

A.2 User Study

• Participants preferred DreamFactory videos in the human evaluation across role consistency, scene consistency, plot quality, storyboard fluency, and overall quality.Source: §A.2 · p. 12
• In the 2AFC protocol, participants saw same-story pairs containing one web-collected short video and one DreamFactory video, then selected the superior video.Source: §A.2 · p. 12
• The study collected 1,320 human scores, comparing DreamFactory videos with 150 multi-scene short AI videos collected from the internet.Source: §A.2 · p. 12

A.3 Case Study

• The user-facing Short generation mode typically produces around ten scenes, with the exact count depending on the task input.Video lengths from random prompts are shown in Figure 10.Source: §A.3 · p. 12, Figure 10 · p. 13
• The responsibility chart shows each employee’s role and portrait, participation stages and roles, and essential prompt parts.Source: Figure 8 · p. 13
• The human-evaluation comparison presents DreamFactory videos alongside internet AI videos.Source: Figure 9 · p. 13