Phr00t

/

WAN2.2-14B-Rapid-AllInOne

like 1.55k

Hi Phr00t,

I’ve been testing Mega v6 with 2-image input (I2V morphs) and noticed a consistent problem:
when providing a start image and an end image, the model begins the video already showing elements from the final image, instead of starting fully from the first one.

For example, I used:

Image 1: Han Solo standing in a sci-fi corridor

Image 2: Indiana Jones in the jungle

Prompt: “A smooth transformation from Han Solo into Indiana Jones. Starts with Han Solo standing in a sci-fi corridor. Gradually, his outfit, environment, and lighting change into Indiana Jones’ leather jacket, whip, and jungle ruins under warm sunset light. Seamless morph, same camera framing, smooth continuity.”

However, in the result:

The very first frame already includes the Indiana Jones hat and jacket.

The transition doesn’t start from the clean Han Solo image; instead, it mixes both identities from the beginning.

The morph appears too “averaged” instead of progressive.

Expected behavior:

The video should begin fully faithful to the first image,

Then smoothly evolve frame by frame until reaching the second image,

Without introducing features from the second image at the start.

Observed behavior:

The model blends both images immediately.

The hat and other distinctive details from the final image appear in frame 1.

Happens even with balanced prompts emphasizing a “gradual transformation.”

Test setup:

Model: Mega v6

Mode: I2V (2 image input)

Frames: 120

FPS: 24

Sampler: sa_solver

Scheduler: beta

Workflow: standard ComfyUI Mega v6

Both input images are clear, consistent framing, similar size.

Possible cause:
It seems Mega v6 may be giving too much weight to the target image conditioning at frame 0, causing a premature blend instead of starting purely from the first image.

Would it be possible to adjust the I2V conditioning so that the first frames adhere more strictly to the initial image, and the morph to the final one starts later or with smoother temporal weighting?

I tried to replicate the example shown here:
🔗 https://raw.githubusercontent.com/Comfy-Org/example_workflows/refs/heads/main/video/wan/2.2/wan22_14B_flf2v.mp4
from the tutorial:
🔗 https://docs.comfy.org/tutorials/video/wan/wan2_2
However, in my test the transition is not smooth — instead of a gradual morph between images, the video abruptly switches from the first to the second scene almost immediately. prompt (A bearded man with red facial hair wearing a yellow straw hat and dark coat in Van Gogh's self-portrait style, slowly and continuously transforms into a space astronaut. The transformation flows like liquid paint — his beard fades away strand by strand, the yellow hat melts and reforms smoothly into a silver space helmet, dark coat gradually lightens and restructures into a white spacesuit. The background swirling brushstrokes slowly organize and clarify into realistic stars and space, with Earth appearing gradually in the distance. Every change happens in seamless waves, maintaining visual continuity throughout the metamorphosis. Consistent soft lighting throughout, medium close-up maintaining same framing, central composition stays fixed, gentle color temperature shift from warm to cool, gradual contrast increase, smooth style transition from painterly to photorealistic. Static camera with subtle slow zoom, emphasizing the flowing transformation process without abrupt changes.
)

Thanks — Mega v6’s quality and motion smoothness are excellent, but this behavior makes it hard to do clean start→end morphs between two distinct subjects.