I was working on a new experiment for the scroll-driven version of my website and needed a video model to bring it to life. The job sounds simple: animate the same source image, move the camera forward, and keep the penguin intact. In practice, each model interpreted that direction differently.
This is the source photo I used for all four tests:
I ran WAN 2.2, LTX 2.3, and HunyuanVideo 1.5 locally on an RTX Pro 6000 Blackwell. My setup uses ComfyUI, so I could run each model through a node-based workflow while keeping the source image and direction consistent. I then finished each result at 1920×1080. This is a practical test for one website, not a general benchmark.
WAN 2.2 Link to heading
WAN 2.2 is Alibaba’s open video model. Its weights were released in July 2025, making it roughly a year old when I ran this experiment.
WAN gave me expressive motion and strong camera travel. The tradeoff was softer character detail, which matters when the same mascot has to survive several scenes.
LTX 2.3 Link to heading
LTX 2.3 generates synchronized audio alongside the video, so the result comes with sound rather than requiring a separate audio-generation step.
LTX rendered this test in about 31 seconds. It kept the mascot stable and gave me predictable motion, so I used it for the full ScrollWorld homepage.
HunyuanVideo 1.5 Link to heading
HunyuanVideo 1.5 is Tencent’s 8.3-billion-parameter video model, with support for both text-to-video and image-to-video generation.
Hunyuan took 911 seconds for the same shot. The result had overexposed details, face drift, and vertical ring artifacts. On this workflow, the extra render time did not buy better output.
Seedance Link to heading
For comparison, I also ran the same idea through Seedance, ByteDance’s closed-source video model, using Jimeng. This one did not run locally or through my ComfyUI setup.
Seedance produced an eight-second 720p clip with sound, smooth cinematic camera movement, and strong environmental detail. It kept the penguin recognizable, although the character eventually moved out of frame. The polished result makes it a useful reference point, but it does not offer the privacy, control, or repeatable local workflow I wanted from a self-hosted model.
My pick Link to heading
Among the self-hosted models, I chose LTX 2.3. WAN was more expressive in places, but LTX gave me the character stability and camera control I needed for a scroll-scrubbed site. Hunyuan was too slow for the result it produced. Seedance looked polished, but relying on a closed service would defeat the purpose of this local workflow.