This is the workflow that actually saves time on client work. You already have a product photo, a character sheet, or a location still. Feeding that in keeps the composition and colors stable while the model adds camera movement and small motion. Text-to-video rarely matches a specific brief; image-to-video usually gets close on the first try.

What makes a good input frame?

Use the highest resolution you have, at least 1280px on the long edge. Avoid heavy JPEG compression and avoid images where the whole frame is motion-blurred. Faces should be sharp, since the model will animate whatever detail is present and will not invent missing structure. One subject per frame works better than a crowded scene.

How do you set up the image-to-video workflow?

• Load the Wan 2.5 image-to-video checkpoint in the load node.
• Feed your still through the image encode node.
• Write the motion description in the positive prompt.
• Keep clip length at 3-5 seconds for a first pass.
• Set the sampler to 20-30 steps; below 15 the motion turns mushy.

Generate a five-second test before committing to a long render. On a 24GB card, five seconds of 720p is roughly 3-6 minutes, and 1080p can triple that.

How should you write the motion prompt?

Describe camera and subject movement separately: “slow dolly-in on the bottle, label stays legible, condensation drips on the glass, warm side lighting”. Vague prompts such as “make it cinematic” produce drifting, weightless motion. Concrete verbs help because the model has been trained to respond to camera language like pan, tilt, and dolly.

What are the limits you will hit?

For repetitive shots, render one good clip and reuse it with different edits rather than regenerating. Each rerun costs GPU time even though Wan 2.5 has no per-clip fee.

When does image-to-video beat text-to-video?

When you need a specific look. Brand work, product demos, and narrative shots with a fixed character all benefit. If you are exploring ideas and have no reference, text-to-video is faster to iterate. Many teams run both: text-to-video for mood boards, image-to-video for the shots that get delivered.

Key Takeaways

• Image-to-video locks composition while adding motion.
• Start with a 5-second 720p test before a full render.
• Describe camera moves explicitly, not moods.
• Expect 3-6 minutes per clip on a 24GB card.

Tongyi Lab’s Wan family is one of the few video model lines where the weights are genuinely open and commercially usable. Version 2.5 is the one that added sound, which changes what you can do without a separate audio pass. The catch is hardware. A small model with quantization fits on one 24GB card; the big ones want a multi-GPU rig.

How much VRAM does Wan 2.5 need?

If you have a single 4090 or 3090, expect to spend time on quantization and offloading. It works, but a 5-second 1080p clip can take several minutes.

How do you install Wan 2.5 in ComfyUI?

Grab the ComfyUI-Wan wrapper, drop the model weights into models/diffusion_models, and put the VAE and audio encoder in their folders. Then load a text-to-video workflow and point it at the checkpoint. Weights are on Hugging Face; the fp16 set for the small model is roughly 30 GB, so clear disk space first. ComfyUI is the practical path because the built-in node presets already handle the audio sync branch.

Can you use Wan 2.5 through an API instead?

Yes. Alibaba Cloud hosts Wan 2.5 as a paid API, and the official demo site gives a limited free quota you can burn through in an afternoon of testing. For occasional use the API is far cheaper than buying a GPU. For batch generation of hundreds of clips, on-premise wins because there are no per-clip fees.

What does the audio sync actually do?

It generates a soundtrack alongside the picture rather than dubbing after the fact. In my tests, ambient and percussive sounds line up well; spoken dialogue is hit or miss and still needs a real voice pipeline. Treat the audio as a solid scratch track, not a finished mix.

Who should use local Wan 2.5?

• Studios with footage that cannot leave the building.
• Teams generating hundreds of B-roll clips where per-clip API fees add up.
• Researchers who want to fine-tune on proprietary footage.
• Anyone building product demos from a single reference image.

Skip it if you need fast turnaround on a laptop, or if you only make a handful of clips a month. In those cases the hosted API is the sane choice.

Key Takeaways

• Wan 2.5 ships open weights with a permissive commercial license.
• Output is 1080p at 24fps with synchronized audio.
• A quantized small model runs on a single 24GB GPU.
• No per-clip fees once you own the hardware; setup effort is the real cost.