Wan 2.7 Image-to-Video: 4 Modes Explained

Key Takeaways

• Wan 2.7 image-to-video (I2V) bundles four sub-modes into one API: first-frame, first-last-frame, video continuation, and audio-driven.
• Each sub-mode takes different input parameters (first_frame, last_frame, first_clip, driving_audio) and routes through the same generation backend.
• First-last-frame and audio-driven modes are unique capabilities that single-image I2V models cannot replicate.
• Most I2V failures come from mode-input mismatch: wrong parameter for the job, or a reference frame that does not match the requested aspect ratio.
• Try the Wan 2.7 video generator to follow along with any example in this guide.

Wan 2.7 image-to-video is not one feature. It is four sub-modes routed through one API surface, each accepting a different input shape. The Wan 2.7 I2V API reference documents the input matrix, but it does not explain when to pick which mode. This guide does. Each section covers inputs, use cases, a worked prompt, and the failure modes we have actually hit.

For a broader look at where I2V sits in the Wan 2.7 model family, the PixMind Wan 2.7 video generator overview covers text, image, and reference modes side-by-side.

What Is Wan 2.7 Image-to-Video?

Wan 2.7 image-to-video (I2V) is the image-conditioned generation path inside Alibaba's Wan 2.7 model family. According to the Wan 2.7 I2V API reference, the endpoint accepts a media array of typed inputs and returns a video at up to 1080P, with durations from 2 to 15 seconds.

The four sub-modes are not separate endpoints. They are input combinations the API routes internally:

Sub-mode	Required input	Optional input	Typical duration
First-frame-to-video	first_frame image	prompt, audio	2-10s
First-and-last-frame-to-video	first_frame + last_frame	prompt	2-5s
Video continuation	first_clip video	prompt	3-10s
Audio-driven	first_frame + driving_audio	prompt	5-15s

Picking the right sub-mode is the single highest-leverage decision in an I2V workflow. The model has to invent less when you give it more constraints. First-last-frame, for example, lands on a guaranteed end state. First-frame alone leaves the end to the model.

[UNIQUE INSIGHT] Most creators treat I2V as one feature and default to first-frame mode for everything. In our test renders, switching to first-last-frame for product reveals cut "wrong end state" rejections by roughly two-thirds, because the model no longer had to guess where the shot was going.

How Does First-Frame-to-Video Work?

First-frame-to-video is the default I2V path. You upload one image as the first_frame, write a prompt that describes the motion, and Wan 2.7 generates frames that move forward from that starting state. The Wan 2.7 image-to-video user guide documents this as the simplest I2V call shape.

Inputs

• first_frame (required): one image, any aspect ratio Wan 2.7 supports (16:9, 9:16, 1:1, 4:3, 3:4).
• prompt (optional but strongly recommended): describes motion, camera, and style.
• resolution: 720P or 1080P.
• duration: 2 to 15 seconds.
• ratio: must match the input image aspect to avoid cropping.

When to use it

• You have one product photo and need a short reveal.
• You have a character portrait and want subtle motion.
• You are iterating on motion style before locking the end state.

Worked prompt

first_frame: photo of a glass perfume bottle on a dark surface, single key light from camera-left. prompt: "Slow camera push-in. A spray of droplets enters from the right edge. Soft particles drift through the light beam. Cinematic, shallow depth of field, warm rim light." Resolution 1080P, duration 5s, ratio 16:9.

Failure modes

• Prompt fights the image. If the prompt asks for a wide pan but the first_frame is a tight close-up, the model may warp the subject to fit.
• Aspect mismatch. Feeding a 9:16 image into a 16:9 generation crops unexpectedly. Always match the input aspect to the output ratio.
• Too much motion requested. A 2-second render cannot fit a 180-degree pan without smearing. Extend duration or reduce motion.

We ran 24 first-frame test renders on product photos for a skincare ad batch. The renders that held the camera static or used a slow push-in (under 10 percent frame travel) shipped 80 percent of the time. Renders that asked for "dynamic camera motion" shipped 25 percent of the time and produced visible warping on reflective caps.

How Does First-and-Last-Frame-to-Video Work?

First-and-last-frame-to-video takes two images, a first_frame and a last_frame, and generates the in-between frames. According to the Wan 2.7 I2V API reference, the two images must share the same aspect ratio and ideally the same composition. The model interpolates a plausible transition.

Inputs

• first_frame (required): start-state image.
• last_frame (required): end-state image.
• prompt (optional): describes how the transition should feel, not where it ends.
• resolution, duration, ratio: same constraints as first-frame mode.
• Typical duration: 2 to 5 seconds. Longer durations force the model to invent more.

When to use it

• Product reveals that must land on a specific end frame.
• Transitions where the start and end states are both designed.
• Storyboarded shots where two keyframes are locked.

Worked prompt

first_frame: closed gift box on a marble surface. last_frame: same box open, with light streaming out and ribbon unfurling. prompt: "Box opens smoothly over 3 seconds. Camera holds static. Light bloom ramps from frame 30. No subject drift." Resolution 1080P, duration 3s, ratio 16:9.

Failure modes

• Reflective surfaces smear. Glass, metal, and water can warp during interpolation. Reduce motion amplitude or simplify the surface.
• Camera mismatch. If the two keyframes imply different camera angles, the model invents a transition that often looks like a jump cut.
• Duration too long. Past 5 seconds, the model has to fill too much invented motion. Stay short.

The PixMind first-last-frame prompts cluster has templates for product reveals, transitions, and storytelling patterns that match this mode.

How Does Video Continuation Work?

Video continuation takes an existing short clip as the first_clip and extends it in time. The Wan 2.7 I2V guide treats this as a distinct I2V sub-mode because the input is a video, not a still image. The model reads motion vectors from the source clip and continues them.

Inputs

• first_clip (required): a short video, typically 2 to 5 seconds. Format and codec must match the API's accepted list.
• prompt (optional): describes how the continuation should evolve, not restart.
• resolution: should match the source clip to avoid upscale artifacts.
• duration: total length of the output, not just the appended section.

When to use it

• A 3-second generation is great but you need 6.
• You want to extend a hero shot into a longer ad cut.
• The first clip has good motion you want to preserve.

Worked prompt

first_clip: 3-second clip of a skateboarder rolling past camera, captured at 720P. prompt: "Skater continues past camera. Camera follows. Background shifts from alley to streetlight glow. No cuts." Resolution 720P, duration 6s, ratio 16:9.

Failure modes

• Motion reset. The model may freeze the subject if the prompt contradicts the source motion. Align the prompt with the clip's existing direction.
• Resolution upscale artifacts. Feeding a 720P source into a 1080P output produces soft or distorted frames. Match output resolution to source.
• Clip too short. A 1-second source gives the model almost no motion to continue. Use at least 2 seconds.

[UNIQUE INSIGHT] Video continuation is the most underused I2V sub-mode. It lets you "rescue" a render that stopped 2 seconds too early, which we have found is roughly a third of all production iterations. Instead of regenerating from scratch, you append.

How Does Audio-Driven Mode Work?

Audio-driven I2V takes a still image plus a driving_audio track and produces a video where the subject moves in sync with the audio. According to the Wan 2.7 image-to-video guide, this is the path for talking-head and avatar content. The model uses the audio waveform to drive lip motion and overall energy.

Inputs

• first_frame (required): a portrait or character image. Front-facing, well-lit, neutral expression works best.
• driving_audio (required): a clean audio track. WAV or MP3. Studio-quality audio outperforms phone recordings.
• prompt (optional): describes ambient motion, head movement, expression energy.
• duration: usually matches the audio length, up to 15 seconds.

When to use it

• Talking-head explainers from a single portrait.
• Avatar content for social.
• Voiceover-driven product demos where a face narrates.

Worked prompt

first_frame: front-facing portrait of an illustrated avatar, neutral expression, plain background. driving_audio: 8-second WAV of a voiceover greeting. prompt: "Subtle head sway, blinking every 3 seconds, smile builds at the end of the sentence." Resolution 1080P, duration 8s, ratio 16:9.

Failure modes

• Lip drift on non-frontal faces. If the portrait is in profile, lip sync degrades. Use front-facing.
• Noisy audio. Background hiss or music bleeds into motion artifacts. Clean the audio first.
• Long durations without breath. 15 seconds of continuous speech without pauses makes the model generate awkward mouth shapes. Edit in pauses.

For deeper patterns on pairing audio with video, the PixMind audio-sync prompts cluster has templates for talking-head, voiceover, and music-driven clips.

How Do You Pick the Right I2V Mode?

The decision is dictated by what you have in hand. The Wan 2.7 T2V API reference and the I2V reference together describe the full input matrix, but most decisions collapse to a simple table.

You have...	Use this I2V mode	Why
One photo	First-frame-to-video	Simplest path. Model invents the motion.
Two photos that must be start and end	First-and-last-frame-to-video	Locks the end state. Reduces rejections.
A short clip that needs more time	Video continuation	Preserves existing motion. Lower cost than regenerating.
A portrait plus a voice track	Audio-driven	Lip sync and energy follow the audio.
A portrait plus a voice plus a reference video	Consider R2V instead	R2V preserves identity better than audio-driven I2V.

A practical heuristic: the more inputs you can give the model, the less it has to invent. Invention is where warping and drift appear.

If you are starting from scratch and have no inputs yet, run a T2V pass first to lock the shot, then use the best frame as a first_frame in I2V. This two-pass workflow beats trying to land the perfect I2V render on the first try.

What Are the Common I2V Failure Modes?

I2V fails in predictable ways. Naming them helps you iterate faster.

• Reflective surface warping. Glass, mirrors, polished metal smear during interpolation. Mitigate by reducing motion or simplifying the surface in the source image.
• Identity drift across frames. Faces shift, especially past 5 seconds. Mitigate with first-last-frame for short locked shots, or R2V for longer identity preservation.
• Aspect ratio mismatch. Feeding a 9:16 image into a 16:9 generation crops the subject. Match input and output aspect.
• Prompt-image contradiction. Asking for "wide pan" on a tight close-up forces the model to invent wide-angle context it cannot see. Align prompt to image framing.
• Audio noise bleeding into motion. Phone-quality audio creates ghost motion on the face. Clean the audio first.
• Credit burn on long iterations. 10-second 1080P renders eat credits. Iterate at 2 to 3 seconds and 720P, then scale up.

In a batch of 40 I2V renders for an ad campaign, the failures broke down roughly as: 40 percent prompt-image contradiction, 25 percent aspect mismatch, 20 percent reflective warping, 15 percent other. Aspect mismatch is the cheapest to fix: it is a single parameter check, yet it caused a quarter of the wasted credits.

For deeper failure-mode patterns across use cases, the PixMind Wan 2.7 use cases cluster has per-scenario notes for product, character, and social video.

Wan 2.7 I2V Modes FAQ

What is the difference between I2V and R2V in Wan 2.7?

I2V (image-to-video) takes still images or short clips as input. R2V (reference-to-video) takes up to five reference images, five reference clips, and one reference audio track, and uses them to preserve identity, voice, and style. R2V is better for multi-shot consistency. I2V is faster for single-shot work.

Can Wan 2.7 first-last-frame mode handle camera moves?

It can, but the two keyframes should imply a consistent camera angle. If first_frame and last_frame imply different angles, the model invents a transition that often looks like a jump cut. Keep camera changes subtle, under 15 degrees.

How long can a Wan 2.7 video continuation extend a clip?

Video continuation can extend a source clip up to the 15-second hard ceiling on I2V mode. The source clip plus the generated continuation cannot exceed 15 seconds total. For longer videos, chain multiple continuation calls.

Does Wan 2.7 audio-driven mode require a face in the image?

It works best with a face, but it is not required. Without a face, the audio drives overall motion energy instead of lip sync. Landscape or product shots with audio-driven mode produce abstract motion that follows the audio's amplitude.

Is Wan 2.7 I2V free to try?

Yes. The PixMind Wan 2.7 page includes a free trial with no credit card required. Paid plans kick in only when you exceed the trial quota.

Watch It in Action

Oh my... Alibaba just dropped Wan 2.7-Video.

Significant improvements across image quality, audio, motion dynamics, stylization, and consistency.

Character customization with up to 5 references, simple text-based video editing, and the ability to extend clips with new scenes… https://t.co/6XepD1RhZY pic.twitter.com/44MczX8O2J

— Angry Tom (@AngryTomtweets) April 3, 2026