The Unreasonable Effectiveness of
Text Embedding Interpolation
for Continuous Image Steering

Supplementary Material

Reve Image

Video Results

Videos with Steering for Wan2.1

Category 01

Cartoon

Edit Type

Cartoon.

Intensity Direction

Left to right: edit intensity increases.

Category 02

Anime

Edit Type

Anime

Intensity Direction

Left to right: edit intensity increases.

Video Results

Videos with Steering for Wan VACE First-Frame Edit

Workflow

For Wan VACE, we first extract the depth map of the input video. We then run inference with VACE using the edited first frame of the video as the first-frame condition.

Category 01

Cyberpunk

Edit Type

Cyberpunk

Intensity Direction

Left to right: edit intensity increases.

Category 02

Summer

Edit Type

Summer

Intensity Direction

Left to right: edit intensity increases.

Image Results

More Results on Qwen Image Edit

Image Results

More Results on Flux

Image Results

Flux2 Results

Method Comparison

Qualitative Results of All Methods

Layout

Each edit type is shown as a subsection. Inside each subsection, every model is displayed as one 5-image row: original plus four edit outputs.