AI Video Generation Crisis: Prompt Drift in WAN 2.2 I2V + SVI Workflows Exposed

AI Video Generation Crisis: Prompt Drift in WAN 2.2 I2V + SVI Workflows Exposed

Stable Diffusion video generation communities are sounding the alarm over a systemic flaw in WAN 2.2’s Image-to-Video (I2V) and Spatial Video Interpolation (SVI) pipelines. According to a detailed Reddit thread from a seasoned AI artist, users are experiencing severe prompt drift—where subsequent video generations ignore new textual directives and instead perpetuate motion patterns from prior frames. This phenomenon, described as ‘momentum carryover,’ undermines creative control and has sparked urgent discussions about model architecture, LoRA integration, and workflow design in generative video systems.

The issue primarily manifests in multi-stage workflows where users generate video segments sequentially. The initial frame generation, triggered by a precise prompt, often produces accurate motion and dynamics. However, when the output is fed into a second stage—typically using the WanImageToVideoSVIPro node—the model appears to lose contextual alignment with the new prompt. Instead of adapting to requested changes such as ‘softer motion,’ ‘shallower impact,’ or ‘slower tempo,’ the system defaults to replicating the prior frame’s motion vector field, sometimes even amplifying speed or intensity contrary to user input. This drift is exacerbated when using multiple LoRAs simultaneously, including the Lightx2v Rank 128 Wan2.1, Lightx2v 1030 Wan2.2 (on high), and Lightx2v 1022 (on low), alongside Noise-Aware Guidance (NAG) systems.

Users report that body dynamics, such as impact forces, facial expressions, and limb acceleration, become unnaturally muted or flattened. In ‘spicy’ or highly stylized contexts—where precise motion control is paramount—this degradation is particularly detrimental. One user noted that attempts to transition from a high-energy dance sequence to a slow, sensual sway resulted in the model retaining 70–80% of the original motion, with only superficial adjustments. The result is a video that feels mechanically repetitive rather than narratively intentional.

Interestingly, the same user discovered that switching to the WanImageMotion node from the IAMCCS-nodes repository dramatically improved prompt fidelity. This alternative node, designed specifically for motion extension, appears to better isolate and recontextualize latent representations between stages, effectively resetting motion momentum. While the exact technical mechanism remains undocumented, early adopters suggest it employs a more robust latent conditioning layer or applies dynamic prompt reweighting during interpolation.

Experts in generative AI video suggest this issue stems from how SVI models handle temporal continuity. Unlike frame-by-frame diffusion, SVI relies on latent space interpolation across time steps, which can cause the model to prioritize smoothness over fidelity to new prompts. When multiple LoRAs are stacked, their gradients may conflict or reinforce unintended motion biases, creating a ‘path dependency’ effect where each generation becomes a prisoner of its predecessor. The use of NAG, while beneficial for noise reduction, may inadvertently suppress the model’s ability to respond to subtle prompt variations by over-smoothing latent transitions.

As the demand for AI-generated video grows in content creation, advertising, and entertainment, such flaws pose serious ethical and creative risks. Without prompt adherence, AI video tools risk becoming instruments of uncontrolled repetition rather than expressive mediums. Developers of WAN 2.2 and associated node ecosystems have yet to issue an official statement. However, the community is increasingly calling for transparent documentation of SVI’s temporal conditioning architecture and open-source contributions to mitigate prompt drift. Until then, users are advised to treat multi-stage video generation not as a seamless pipeline, but as a series of high-stakes creative interventions requiring constant manual correction.