Experiment · First → Last Frame

First Frame Last Frame AI (Free, Wan 2.2): 3 Runs with Unrelated Images, 3 Different Results

I gave Wan 2.2 a hand-drawn sketch as the first frame and a photo-real apple as the last frame. Same prompt, three runs, three completely different transitions.

Published October 2026 · By Vismantas Šleinota · Rendered locally on RTX 5070, 12GB VRAM

Quick answer: When the first and last frames have nothing in common, Wan 2.2 still connects them, but it picks a different path every run. It may move the drawn character into the end frame's world, stay in the drawing and then cut, or slowly dissolve one world into the other. The drawing style often survives even in a realistic scene. Render a few attempts and keep the best.

The setup

First frame: a pencil and crayon drawing of a spotted, camel-like creature in front of grey mountains.
Last frame: a photo-real apple pierced by an arrow, with a castle and cliffs in the background (made with an image generator).
Mode: First → Last Frame in Loomance (Wan 2.2 I2V, fp8), 57 frames, interpolated to 24 FPS, 2× upscale.
Prompt (same for every run):

The spotted camel-horse creature suddenly lifts its head skyward and lets out a loud, echoing bray, its whole body tensing as the front legs plant firmly into the ground and the tail whips sideways in surprise. Clouds drift rapidly across the sky behind it, sunbeams flickering. Small birds burst out from the bushes at the mountain's base, startled by the sudden noise.

Notice that the prompt says nothing about an apple or an arrow. That was on purpose: I wanted to see how the model bridges the gap on its own.

Run 1 (640×360): the creature moves into the castle world

Run 1: drawn creature placed into the castle landscape, then the apple appears — Frames from start to end. The creature keeps its crayon look but now stands in the end frame's landscape.

Run 1 video (muted, 4.7 s)

The model swapped the background almost immediately and placed the drawn creature into the cliffs-and-castle landscape from the last frame. The creature kept its crayon texture. It turns its head, birds fly out of the bushes, and only in the final moment does the apple appear.

Run 2 (1536×864): stays in the drawing, then a whip cut

Run 2: creature brays in the drawing, then a blurred camera move to the apple — The drawing stays a drawing, then a blurred camera move jumps to the apple scene.

Run 2 video (muted, 4.7 s)

This time the model stayed in the drawing's world. The creature throws its head up and brays, which matches the prompt closely. Then a fast, blurred camera move takes us to the apple scene, which holds for the last couple of seconds.

Run 3 (1536×864, same settings as run 2): a slow dissolve

Run 3: pencil background dissolves into a realistic landscape, creature becomes semi-3D — Same settings as run 2. This time the pencil world gradually turns into the realistic one.

Run 3 video (muted, 4.7 s)

Identical settings to run 2, but a very different result. The pencil mountains slowly dissolve into the real landscape, the creature turns semi-3D and rears up, then the camera drops and the apple fills the frame. Of the three, this is the smoothest transition.

What this tells us

Randomness matters more than resolution here. Runs 2 and 3 used the same settings and still went in opposite directions. Each render starts from a different random seed, and with unrelated anchors there are many possible ways to connect them.
The character usually comes from the first frame, the world from the last. In two of three runs the drawn creature ended up living in the apple's landscape.
Style can survive. The crayon texture stayed visible even in a photo-real scene, which is a fun way to mix styles.
The prompt still steers the motion. The bray, the tensing body and the startled birds showed up across runs, even though the prompt never mentioned the apple.

Bonus: the same drawing in Wild Mode

Wild Mode: the drawing turned into a bold comic-style scene — Wild Mode, no end frame. The drawing becomes a bold comic scene with sunbeams and flying clouds.

Wild Mode video (muted, 4.7 s)

For comparison, I ran the same drawing and prompt in Wild Mode (1280×720, 49 frames), which has no end frame. It reimagined the whole scene as a bold comic panel with a dark blue sky, flying sunbeams and cartoon clouds, while keeping the creature recognisable.

When it breaks: matching frames, mismatched prompt

Two more runs show the opposite problem. Here the first and last frames were almost identical: the last frame of one clip and the first frame of the next, both showing the same room at night. That is how you chain clips into a longer video. But I forgot to change the prompt, so it still asked for a camel-horse braying at mountains.

Same room at start and end, wrong prompt: clouds appear indoors, a burst of smoke, then the frame falls apart into noise.

Same setup: the drawn giraffe leaps out of the monitor into a stormy sky. A striking shot, but the last frames dissolve into noise.

With nothing to bridge between the frames, the model tried to act out a prompt that did not fit the scene. It invented big events (indoor clouds, a giraffe jumping out of the screen) and then could not find its way back to the end frame, so the final frames broke down into noise.

The second run is still useful: the giraffe in the storm became one of the best shots in my music video, after trimming the broken ending.

Rule of thumb: when start and end frames are the same scene, describe only small motion that fits that scene, for example "Static camera. The person keeps working, the globe hologram slowly rotates, city lights twinkle outside." Save big actions for frames that are actually different.

Tips for First → Last Frame

Want control? Use start and end images with similar framing, camera angle and lighting, and describe the transition in the prompt.
Want surprises? Use unrelated images and let the model invent the bridge.
Draft at low resolution. At 640×360 or 832×480 you can try several seeds quickly, then render the winner at high resolution.
Chain renders. Extract the last frame of a good render and use it as the first frame of the next one to build a longer sequence. Update the prompt for each new clip.
Trim the last frames if needed. When a run falls apart at the very end, the rest of it is often still usable.

For render times and memory use at different resolutions, see Wan 2.2 on a 12GB GPU: real render times and RAM use.

Tools and sources

Wan 2.2, the open-source video model by Alibaba's Wan team (Apache 2.0 license)
ComfyUI, the open-source engine Loomance runs on
Loomance, the desktop app used for these renders (free, ad-supported)

VŠ

Vismantas Šleinota builds Loomance, a free local AI animation app, as a solo project in Lithuania. Every render on this page was made and timed on his own PC (RTX 5070 12GB, 64GB RAM). Questions or corrections: info@loomance.app.

Frequently asked questions

What happens if the first and last frames are unrelated?

Wan 2.2 still tries to connect them, but the path changes from run to run. In these tests it either moved the drawn character into the end frame's world, stayed in the drawing and then cut to the end frame, or slowly dissolved one world into the other.

Why do I get a different result with the same settings?

Each render starts from a different random seed. With unrelated anchor images there are many possible ways to bridge them, so the seed has a big effect. Run a few attempts and keep the best one.

Does the drawing style survive in a realistic scene?

Often, yes. In two of three runs the crayon texture of the drawn creature stayed visible even after it moved into a photo-real landscape.

How do I get a smoother transition?

Use start and end images with similar composition, framing and lighting, and describe the transition in the prompt. Unrelated images give more surprising results, but less control.

Try First → Last Frame on your own GPU

Loomance is free to use (ad-supported) and renders locally on your NVIDIA GPU. Pick two images, write a prompt, see what the model invents.

Download free