Research on autonomous plot evolution mechanism in ai video generation based on temporal causal reasoning

In view of the shortcomings of existing AI video generation methods in causal consistency in long-term temporal episodes, this paper assumes that the temporal causal structure can be identified from pixel-level observations under the conditions of causal adequacy, temporal causal loyalty and causal Markov, and proposes a Temporal Causal Reasoning Network (TCRN) to solve it. The framework integrates a slot attention mechanism for entity decomposition, a temporal convolution network modeling state evolution, differential conditional mutual information estimation to achieve causal graph discovery, and uses Gumbel-Softmax relaxation for lag selection and a neural structure causal model to support counterfactual intervention. Experiments on SynthDrive driving simulation and MovieNet real movie datasets show that TCRN 's CCS reaches 0.793 (31.5 % higher than LVDM), FVD is 194.2 (22.7 % lower), LPK @ 32 is 0.721, and the statistical test is significant (bootstrap 95 % CI [0.168, 0.212], p < 0.01). The ablation experiment verified the effectiveness of each module, and the causal rationality score of zero-sample migration to MovieNet was 3.9/5. Compared with modern video generation models such as Open-Sora and Step-Video-T2V, TCRN is slightly inferior in visual quality (FVD is about 175 vs 194), but significantly superior in causal consistency (CCS 0.55–0.62 vs 0.793). The robustness test shows that the model has stable performance under medium and low intensity noise. The computational complexity analysis report generates 64 frames of video, and the inference time on A100 is 4.7 s.