Wan 3.0 first and last frame: what it does, and the comparison table that gets it wrong

If you have asked an AI assistant which models support first-and-last-frame generation in the last few weeks, there is a reasonable chance it handed you a table with a dash in the Wan 3.0 row. That dash is wrong, and it is worth correcting precisely, because the feature is not a rumour or a roadmap item — it is two documented parameters with a published example sitting next to them.

Wan 3.0 takes first_frame and last_frame, one image each, and generates the motion between them. Alibaba's own model guide lists it as its own task type, with this description: "Provide both first_frame and last_frame to strictly specify the first and last frame images of the video."

The published example, and what ffprobe reads off it

Alibaba released a demonstration file for this mode. We pulled it and measured it rather than repeating the page copy:

The prompt published with it asks for a fifteen-second unbroken take and then spends a paragraph on sound: guzheng and bamboo flute, faint water, morning birdsong, and a line recited from the eighth second. The file comes back with an audio track. That matters more in this mode than it first appears, and we come back to it below.

A second published example on the same page gives the other half — the two input images. Both are 2560×1440, and the call that uses them sets resolution="720P", ratio="adaptive", duration=5.

The comparison table

Three models are commonly compared on this. Each row below links to the vendor's own documentation rather than to a round-up.

Sources: Alibaba Model Studio Wan 3.0 guide and API reference; Google Cloud's Veo 3.1 model page and the Gemini API announcement, which prints the last_frame call verbatim.

One caveat on the Veo row that is easy to miss and expensive to discover: at launch, developers reported the Gemini SDK raising last_frame parameter is not supported in Gemini API on the exact model the documentation names, in a thread on Google's own forum. Documented and shipping are not always the same week.

Three things about Wan 3.0's version that are not on any spec sheet

1. Pinned frames and reference images are mutually exclusive

This is the one that costs people a queue slot. Wan 3.0 sorts every attachment into two families — frames (first_frame, last_frame) and references (reference_image, reference_video, reference_audio, file, link) — and it rejects a request that mixes them. Alibaba's guide states it plainly: first-and-last-frame mode supports only those two types and cannot take reference_audio or the others alongside.

So you cannot pin both ends and hand over a separate product photo or a face in the same run. In a transformation that usually costs nothing, because whatever you are selling is already inside both photographs. If you need a named character as well, that is a reference-mode job and a different request.

2. The first frame anchors the opening, not the whole clip

A first frame guarantees where the shot starts. It does not constrain what happens after. PixelDojo's prompting guide, written from generations on the live model, describes a test where the clip opened on the supplied frame exactly and then went further than the prompt asked — the subject left the post he was supposed to be leaning on and the camera travelled behind him. Their conclusion is the useful one: "A first frame anchors the opening, not the whole clip, so state plainly what must not change if you need the shot held."

Practically: name the thing that must stay still, in one clause, every time.

3. resolution sets total pixels, not aspect ratio

The setting people read as "720p means 1280×720" does not work that way here. Alibaba's own first-and-last-frame guide explains that the resolution parameter controls the total pixel count, while the model preserves the first frame's aspect ratio and then nudges both dimensions to multiples of 16 for the encoder. Their worked example: a 750×1000 input (3:4) at 720P can come back 816×1104.

Which leads to the single most practical sentence in this whole mode:

The shape of your output is decided by the image you pin, not by the control you set. Both official examples set ratio="adaptive", and the guide recommends it for both frame modes because the frame should decide. Feed it a landscape photograph and you get a landscape clip. If the clip is going to Reels, the first frame has to be vertical.

What the mode is actually good for

Anything where you already own the ending. That is a narrower set than it sounds, and it is the set where this beats a general image-to-video prompt:

• A real before and after. Two photographs of the same subject, taken before and after the work, pinned as the two ends. Nothing about the result is invented by the model — it renders the travel between two states you already documented.
• A product turn that has to land on the packshot you are already using everywhere else.
• A shot that has to cut into existing footage, where the last frame is the first frame of the next clip.
• A logo or end card that has to arrive exactly as drawn.

And the case against: if you do not know the ending, supplying a last frame makes the result worse rather than better, because you have spent your control budget pinning a destination you were not sure about. One frame and a sentence is the better trade there.

The sound argument, which nobody makes

Every comparison of this feature is about pictures. Watch any good transformation with the audio off and the join stops working — the change needs a transient on the frame it happens, and most routes to this format make that a second purchase, dropped in afterwards and nudged into place by hand.

Wan 3.0 samples the sound with the picture, which is why Alibaba's own example prompt spends a paragraph naming the instruments, the water and the recited line, and why the file comes back with an audio track rather than a silent morph. Toggling audio off does not reduce the price, because it was never billed as a separate product.

Where to run it

The model is closed — no published weights — so every route is hosted:

• Alibaba Cloud Model Studio, directly. The official route, and the one to use if provenance matters more than convenience. You manage the account, the keys and the billing.
• ComfyUI's partner node, which exposes first_frame with an optional last frame, per its documentation.
• A hosted interface, of which several now exist. This site is one: first and last frame on wan-3.run runs the same wan3.0-video model, 2–30 seconds, 480P/720P/1080P, with the per-second rate printed on the pricing page. We are a convenience layer over the same API, which is worth saying plainly rather than leaving you to work out.

If the job is specifically two real photographs of a change you made, the before and after generator is the same mode with the two upload slots labelled and the craft written down.

The short version

Wan 3.0 has first-and-last-frame generation, documented, with a published fifteen-second example that measures 1920×1080 at 30 fps with sound. Any table that puts a dash in that row is out of date. The three things worth knowing before you use it are that pinned frames exclude reference material, that the first frame only guarantees the opening, and that the aspect ratio comes from the image rather than the setting.