CNCC 2026 Forum: Video Generation at a Crossroads—More Realistic or More Moving?

Right now, the competitive focus of video generation is shifting from "does it look real" to "can it be used."

Metrics such as image quality, resolution, and single-clip duration are being rapidly closed by different research teams. As the gaps in the imagery itself converge, the remaining gap concentrates in places that are harder to quantify—such as whether the narrative is coherent, whether the emotion lands, and what happens once AI content enters real dissemination pipelines.

This is precisely the question that the CNCC 2026 conference's themed forum "The Intelligence and Emotion of Video Generation" attempted to discuss on the afternoon of October 23.

Yang Jufeng, a professor at Nankai University, is the convener of this themed forum. According to him, this topic actually has an earlier origin.

In 2024, ten domestic scholars including him jointly proposed "Research on Digital Humans and Robots with Both Emotion and Intelligence." This topic was selected as the first of the year's Top Ten Frontier Scientific Questions of the China Association for Science and Technology.

And this year, this research direction has converged onto the more specific track of "video generation."

Yang Jufeng likens "intelligence" to the human skeleton—this part of the technology is relatively mature, and what remains more unresolved is the "emotion" part. In his opening, he left a question to provoke everyone's thinking: if we had to choose one of the two, would we want the generated video to be more realistic, or more moving?

As one of the media outlets reporting on the CNCC 2026 conference, Leiphone.com brings you the exciting findings of this themed forum from three dimensions: engineering delivery, underlying technical routes, and social impact.

01

In video generation's development to date, the imagery itself is no longer difficult to produce, but when it actually needs to be used for work, more granular challenges arise: how to keep the plot coherent? How to get the emotion right?

A single foundation model clearly cannot support delivery on its own. How to achieve better delivery has become a mandatory question facing practitioners.

Hu Baotian, a professor at Harbin Institute of Technology (Shenzhen), in his report "Multi-Agent Collaboration for Long-Sequence Video Generation: From Rule Controllability to Narrative Empathy," splits long video generation into two lines: "quality" and "emotion."

Quality is the cornerstone of realism, requiring conformity to physical laws and spatial relationships; emotion is the source of appeal, encompassing narrative aesthetics and empathic expression.

According to him, at present, a single foundation model can stably generate lengths of 15 to 20 seconds; beyond that, cross-shot character drift, cross-modal audio-visual misalignment, and long-cycle narrative breakdown begin to appear en masse.

He gave a plain example: generating a scene between characters like Liu Neng and Zhao Si from "Country Love" may be harder than generating a sci-fi blockbuster—the difficulty is not in the spectacle, but in the continuity of character relationships.

In response, his team has mainly done three pieces of work:

One is making the workflow controllable. The team built a generation workflow benchmark for ComfyUI, using node optimization plus reinforcement learning to shore up shortcomings in long-range planning; a 14B-scale model can achieve relatively high accuracy, and this system has been integrated by ComfyUI as a native plugin.

Another is long-range narrative. Pixel Video uses story templates to coordinate the generation process and maintain character and timeline states; the related open-source project has gained 30,000 stars on GitHub. The later Video Vlog weakens rigid rules and shifts to a multi-agent director framework that can generate long videos of the serial-drama type.

A third is editing. Cut Master centers on a shared editing state and can handle hour-level videos, with editing quality scores improving by up to 17 points compared with the average level.

He set an eight-character goal for this endeavor: controllable logic, perceptible emotion.

Qian Fan, a researcher under the Steed Program at Inner Mongolia University, picks up this goal from another angle in her report "Physical Perception and Video Emotional Expression in Video Generation."

She divides the problem into three parts.

In the physical perception part, early approaches used external constraints such as trajectories, cameras, and motion fields to control dynamic changes—controllable, but not guaranteeing physical correctness such as collisions and forces.

In recent years, industry progress has shifted toward introducing physical laws; some research has set up 2,000 kinds of physics experts, first extracting laws from text and then refining them into local visual features.

Her observation is that the goal of this direction has changed: in the past it pursued a good-looking next frame; now it emphasizes more whether states can persist and whether responses can be made to environmental changes.

Emotional expression is closest to "emotion."

She believes that controlling emotion requires satisfying three conditions: clarifying the semantics of the event currently happening, defining continuous changes along emotional dimensions, and setting up a temporal structure to organize the process of emotional change.

One easily overlooked point is perspective. Past research has mostly focused on the emotions of characters within the video, but to generate videos that meet creative goals, what needs attention is the emotional feedback from the audience's perspective.

In the music part, the same footage with different music can make the audience feel completely different emotions. The common shortcoming of existing solutions is that emotional intensity cannot be adjusted; the model can only generate "happy" or "tense" according to instructions.

Qian Fan's team's approach combines scalable low-rank adaptation with reinforcement learning to achieve continuous intensity adjustment of style, emotion, and timbre.

02

Will vision be reduced to a "renderer"?

Engineering-wise, multi-agent systems and workflows are more like patching on top of foundation models. To solve long-video drift and physical errors, everyone has built "scaffolding" all around the model, and this approach raises a question:

If the "intelligence" in an intelligent system mainly comes from language models, how much independent value does visual intelligence still have? Will visual intelligence be reduced to a renderer?

Yuan Li, a researcher at Peking University Shenzhen Graduate School, mentioned in his report "How Visual Generation and Understanding Can Move Toward Collaborative Unity" that with the emergence of models like GPT-6 Astra, the positioning of visual intelligence is facing a new shock.

In the face of powerful language models, the path of understanding the physical world through vision is proving difficult. If vision is only responsible for drawing out results that others have already computed, where does vision's independent intelligence lie?

In Yuan Li's view, the root cause of vision degenerating into a "drawing-tool worker" lies in the long-term split between understanding and generation at the underlying logical level.

From the perspective of information theory, understanding is the compression and encoding of redundant physical signals, while generation is the decoding and reconstruction of abstract semantics. The two are inherently inverse processes, yet in reality they each go their own way.

On the one hand, the knowledge a model acquires in language understanding is hard to transfer freely to the generation stage; on the other hand, existing visual understanding habitually compresses images and projects them into language space, and video understanding also requires frequent frame extraction, so continuous physical laws and spatial relationships are almost entirely lost in compression.

To bridge this gap, the field has gradually split into two solutions.

One is the currently mainstream serial unification, which is equivalent to letting understanding act as the "left brain" and generation as the "right brain," aligned through language space in between. Although this approach can directly reuse mature open-source large models, the narrow intermediate channel inevitably brings information loss.

The other is parallel unification, in which understanding and generation share a backbone network. Although this route is still trapped in the bidirectional conflict of "understanding needs to grasp semantics, generation needs to capture details," Yuan Li believes this is the inevitable direction for deep fusion.

Regarding the relationship between engineering and models, he has a clear judgment: foundation models determine the upper limit, and agents hold up the lower limit.

As for the future of vision, Yuan Li classifies himself as a "video faction." Video inherently carries spatiotemporal continuity and belongs to the 2.5D modality; its data cost is far lower than pure 3D, and it is entirely possible to move directly toward 4D dynamic world models.

The ultimate independent value of visual intelligence is not to compete with language models in abstract reasoning, but to serve as the system's "eyes," providing physical verification for decisions by observing real states and generating intermediate deductions.

03

When video is realistic enough, the risk is no longer in the picture

Once generated content is plentiful and realistic enough, it will flow into social networks. At that point the nature of the problem changes, and "what happens when content enters society" becomes the more pressing issue.

Liu Wu, a distinguished professor at the University of Science and Technology of China, placed the starting point on agents themselves in his report "Emotionally and Intellectually Capable Multi-Agent Content Generation and Social Simulation." His entry point is the composition of human intelligence: problem-solving ability corresponds to IQ, and group collaboration ability corresponds to EQ.

His research work is divided into two parts: one is cultivating agents with diverse empathetic cognition and large-scale social behavior capabilities, and the other is building a social network simulation environment that truly maps reality.

Here there is an unavoidable obstacle:

Using large models to simulate social groups, first, makes it difficult to display diverse characteristics, and second, as the number of interacting people increases, cognitive homogenization easily occurs, with everyone's cognition and emotions tending toward consistency, which makes the simulation itself meaningless.

His team's response is to introduce an intelligent foundational information theory framework and design a dedicated "cognitive fingerprint" for each agent, letting it act separately across the entire process of perceptual filtering, memory updating, and content generation.

On this basis, the team can use real network data to build a social simulator, throw this group of agents into it, and let them chat, interact, and cooperate like real people, "training" their EQ and empathy in specific scenarios.

This simulator inputs the dynamics and target events of real networks, lets agents simulate real social interaction inside it, and deduces the possible development trends and final outcomes of events.

She Dongyu, an assistant researcher at Zhongguancun Laboratory, discussed the other side of this issue in her report "From Video Generation to Cognitive Influence: Multimodal Content Computing for Cognitive Security."

She categorizes the social risks of AIGC video into five types: tampering with historical facts, spreading false disaster information, misleading through contextual clipping, cognitive manipulation regarding key figures and events, and dissemination of harmful value orientations.

Here there is a key distinction. Content security focuses on the compliance, authenticity, and health of the content itself; cognitive security takes people as the subject and protects correct cognition at the individual, group, and even societal levels.

Among these, cognitive risk can be divided into three layers: the factual layer, referring to whether the events, figures, and claims described in the video are supported by sources; the narrative layer, referring to tampering with the original meaning through clipping and reorganizing materials, re-dubbing, and implying false causality; and the value orientation layer, referring to content that glorifies violence or rationalizes violations.

Why is cognition easily manipulated? She offered a three-layer explanation. After receiving a video, audiences must go through three stages of processing: visual transcription, logical linking of events, and interpretation combined with experience, with emotion also participating.

Experiments have shown that information paired with images greatly increases credibility.

In addition, the dissemination chain also amplifies risk. She gave a cross-border case: a domestically generated video about the demolition of a temple in Henan was moved overseas and rewritten as the demolition of a temple in Tibet, and then transmitted back to China.

Traditional technology is mainly rule-based judgment and can only identify explicit violations such as images and text, but cannot identify implicit influences hidden in narrative logic and material associations.

From this, She Dongyu summarized three difficulties: looking at any single frame alone, nothing violates the rules, and the risk is hidden in the associations; technology iterates quickly, and test samples cannot cover new forms; and the same statement carries completely different attitudes in different scenarios.

Zhe Dongyu's team's exploration focuses on two areas. One is emotion understanding, which can predict different audiences' differentiated feelings based on the same image. The other is a fact-checking agent that integrates three types of behavior—evidence gathering, questioning, and summarization—so the model can autonomously plan investigation paths while combining knowledge search.

In her view, the reasoning ability of large models themselves is certainly important, but for this kind of content computing, the ability to organize and use tools is equally important.

Yuan Li's judgment on this is that after unification, creativity will not be suppressed but instead stimulated.

He gave an example of painting.

People usually assume that creation requires understanding first and planning first—figuring out clearly what to paint before starting. But in Yuan Li's view, the painting itself is the creative process; the real sensitive points and creative points often burst out during the act of putting brush to paper and cannot be written out in advance.

Playing Go is the same: some moves only become clear once you reach that step.

▎Question 2: Is emotion modeled or performed?

When people express emotion, is the emotion modeled, or does it flow naturally from the heart? Applied to AI video production, can all the emotion in short videos be planned out through programming—at what point in time, which character, what emotion to express—or will some emotion emerge on its own after generation?

In Qi Fan's view, when applied to models, "performing" refers to whether the model can possess a whole performance process that is commonsense-based, consistent, and understands "why one is sad"; "modeling" refers to having the video generation model model the facial muscle control movements of sadness.

To make videos emotional, the focus should be on understanding the meaning of sadness and understanding how sadness arises, which helps more with the naturalness of the generated result.

Spontaneously feeling emotion and not expressing emotion according to a plan—this direction is exactly what the next generation of models should pursue.

In Liu Wu's view, at the current stage the emotion of generative models is modeled; whether agents or large models will later give rise to genuine emotion cannot be absolutely determined.

He called this possibility a rather frightening thing: once it happens, it involves ethics—does an agent count as a person or as a machine?

It depends on the purpose of the video. If it is for learning knowledge, providing training data, or modeling real events, then physical truth should be followed and emotional interference minimized. If it is artistic content such as short dramas or films that needs to express strong emotion—especially science fiction dramas—then violating some physical truth is not necessarily unacceptable.

Priority is determined by the ultimate need and can be dynamically adjusted according to that need.

▎Question 4: What if it is very real and very moving, but the orientation is wrong?

Zhe Dongyu believes that if a video is very real and very moving but has a problematic orientation, then it is problematic.

In cognitive security, value orientation is one of the key and independent links. At present there is not much technology that can ensure the detection of problematic value orientations.

So what we need to do is push intervention as far forward as possible. For example, give prompts at the generation stage to tell users where the problem lies in this thing, which can block some unintentional actions.

The parts that cannot be blocked must rely on later links. She gave an example: a video popularizing floods shows no problem at generation time, but with malicious intent and a misleading title, it will produce harmful social impact. In her view, generation, dissemination, and the post-viewing links after audiences watch all need corresponding governance measures.

▎Question 5: In the coming year, what do you most want to accomplish?

Zhe Dongyu values self-evolving agents. As the saying goes, the devil is one foot higher than the dao—security technology must continuously improve according to existing technology, letting security agents iterate themselves in an autoregressive way rather than through inefficient static deployment.

Liu Wu hopes to make emotion and intelligence more real and harder to distinguish.

Qi Fan hopes to generate some videos with more cinematic quality.

Yuan Li hopes that in the next three to five years or even longer, foundation models can simultaneously generate and understand the physical world; if speaking of the coming year, he hopes the model will first do a good job of understanding the physical world before generating this world.

05

Conclusion

Returning to the question left at the opening: should generated videos be more real, or more moving?

This forum did not give an either-or answer. Following the presentations of the several speakers, the route on the side of intelligence is relatively clear—moving toward world models and physical laws; the side of emotion is still looking for measurable handles, and the emotional intensity of soundtracks and feedback from the audience's perspective are still at an early stage.

One observation worth retaining is that when image quality and duration are no longer thresholds, the next watershed for video generation probably lies in what happens after content enters the dissemination chain. Technology is responsible for producing the images, and the right of evaluation has been handed to society.

The answers to the above questions can be left for the next forum to continue exploring. Leiphone Leiphone Leiphone

CNCC has ended. During the conference, AI Technology Review learned about many new technological advances and also gained a lot from exchanges with peers. Researchers, students, and readers who follow frontier progress in the computer field are welcome to join our CNCC exchange community.

Here, you can meet peers, exchange research, share admissions information, scientific research cooperation, dinner invitations, submissions, team formation for attending the conference, and more, and look for potential partners.

During the conference, we continuously brought on-site report updates, interpretations of guests' speeches, summaries of frontier directions, and forum highlights, and we will continue to share related content afterward.

See you in the group.

? Group entry portal: Scan the QR code to join the group or add WeChat Qvv0909777, with the note: CNCC + unit/school + name + direction.