Preparing Source Images with Transparent Backgrounds for Image-to-Video Generation
No image-to-video model preserves alpha channels—transparent backgrounds flatten to a solid fill that bleeds into motion. Here is what to do instead.
What Happens To A Transparent PNG In Image-To-Video Generation
No image-to-video model, as of June 2025, preserves alpha channels. Feed a transparent PNG into Runway Gen-3 Alpha, Kling 1.6, or Pika 2.0 and the system flattens it to a solid fill colour during preprocessing. That colour is black or white. It then bleeds into the generated motion as edge fringing, halo distortion, or background sliding where the model tries to animate the flat fill as if it were scene content.
Composite your cut-out onto a solid background that matches your intended final background before generation. Or use region-masking in Pika 2.0 to animate only the foreground element and composite in post. Here is the exact failure mechanism and the workaround for image-to-video transparent background source images.
Why The Alpha Channel Disappears
Image-to-video models accept PNG input. Runway Gen-3 Alpha and Pika 2.0 both document PNG as a supported format. But they output MP4 with no transparency layer. The pipeline decodes the PNG's RGBA data, reads the RGB channels, and discards the alpha channel during the encoding step before the model ever sees the image. The transparent region is replaced with a matte colour, typically black (0,0,0) or white (255,255,255), depending on the preprocessing code.
That solid fill area is then treated as scene content. If your source image has a foreground element on a transparent background, the model sees that element on a black or white rectangle. When it tries to predict motion, it applies movement to both. The rectangle shifts, warps, and bleeds into the edges, creating visible distortion the model's training never associated with a real background.
The result: the transparency you preserved in Photoshop or GIMP is gone before generation starts. What remains is a flat colour fill that acts like a layer of content the model has to animate.
What The Model Actually Sees
The model's input pipeline performs a colour space conversion from RGBA to RGB, discarding alpha. This is a design constraint of the latent diffusion architecture, not a bug. The model's training data consisted of RGB videos. There is no temporal attention layer trained on transparency data. Even a model that accepted alpha in theory would have no learned behaviour for it.
When you upload a transparent PNG, the preprocessing code fills the alpha region with a default value. That value becomes part of your scene geometry. The model then generates motion across the entire frame, including the fill area. Three failure modes dominate.
Edge Fringing And Halo Distortion
The boundary between your cut-out and the fill colour gets a thin band of transitional pixels. These shift differently from both the foreground and the background. The result is a coloured halo around your element in the generated clip. It is most visible on hair, fur, and fine detail edges.
Background Sliding
The fill colour area drifts independently of the foreground. If your element moves left, the fill area may move right, creating a parallax effect that looks like the foreground is peeling away from a glass pane. The model treats the fill area as a flat coloured object in the scene.
Foreground Drift Into Fill
The model may blend foreground pixels with fill pixels over time. Your element fades or morphs into the flat colour. This is most aggressive when the fill colour is black. The model's training data associates black surfaces with shadow and depth, not with empty canvas.
Which Models Accept PNG Input At All
As of the dossier's stated entity capabilities, Runway Gen-3 Alpha and Pika 2.0 both document PNG as a supported input format. Neither preserves the alpha channel. Kling 1.6 and Luma Dream Machine also accept PNG input per their documentation, with the same alpha-flattening behaviour. Every model's documented input specification confirms PNG support. No model claims alpha preservation. If you need transparency in the output, achieve it through post-generation compositing, not through the model's native output.
| Model | PNG Input Accepted | Alpha Channel Preserved | Output Format |
|---|---|---|---|
| Runway Gen-3 Alpha | Yes | No | MP4 (H.264) |
| Kling 1.6 | Yes | No | MP4 (H.264) |
| Pika 2.0 | Yes | No | MP4 (H.264) |
| Luma Dream Machine | Yes | No | MP4 (H.264) |
| Sora (OpenAI) | Yes | No | MP4 (H.264) |
| Veo (Google DeepMind) | Yes | No | MP4 (H.264) |
| Wan 2.1 | Yes | No | MP4 (H.264) |
| LTX-Video | Yes | No | MP4 (H.264) |
| Stable Video Diffusion | Yes | No | MP4 (H.264) |
The Transparent PNG Image-To-Video Workaround
Since no model preserves alpha, composite your cut-out onto a solid background before generation. This is the transparent PNG image-to-video workaround that compositing professionals use in production. The method depends on your final use case.
When You Know The Final Background
Place your cut-out onto the exact background you intend to use in the final video. That background becomes the model's scene. The foreground sits on real content rather than a flat fill. The model animates both together as a single scene, which is what it was trained to do. Edge fringing disappears because there is no sharp boundary between a cut-out and a fill colour. The model sees a natural transition between foreground and background content.
When You Need A Transparent Final Output
Generate the clip with your element on a green screen or a single-colour background you can key out in post. Use a chroma key green not present in your foreground. After generation, use FFmpeg alpha extraction to pull the key and create an RGBA output. This is the image-to-video compositing source preparation method. It gives you a transparency layer in the final clip, but adds a post-processing step and introduces the risk of keying distortion on edges.
Using Pika 2.0 Region Masks
Pika 2.0's region mask feature lets you select a portion of the source image for animation while keeping the rest static. If your source image already has a solid background, apply a region mask to your foreground, generate the clip, and then composite the animated element onto a transparent background in your video editor. Pika 2.0 does not output alpha, so the compositing step is still required. But the region mask reduces background sliding because the unselected area stays unchanged.
Runway Gen-3 Alpha Transparent Background Support
Runway Gen-3 Alpha accepts PNG input but documents no alpha support in its input specification. The answer to the Runway Gen-3 Alpha transparent background support question is the same as every other model: it does not exist. The documentation specifies JPEG and PNG as image formats, with no mention of RGBA or alpha handling. Transparent regions are filled with black during preprocessing.
Runway's motion brush and camera controls work on the flattened RGB image. If your source image has a transparent background, the motion brush applies motion to both the foreground and the black fill area, producing the distortion described above. The only reliable method in Runway: composite your element onto a real background before upload.
Preparing Source Images Correctly
The preparation step determines the quality of the generated clip more than any other variable. Follow these rules for any image-to-video model.
Never Upload A Transparent PNG
Export your cut-out with a solid background. Black or white cause the worst distortion because they are pure colours with no texture. Use a mid-tone grey or a blurred version of your actual background. The model has less contrast boundary to animate incorrectly.
Match Aspect Ratio To Model Training
Models are trained on specific aspect ratios: 16:9 for landscape, 9:16 for portrait, 1:1 for square. Uploading a different ratio forces the model to pad or crop, which introduces the fill colour problem from the other direction. Crop your source image to the model's native aspect ratio before generation. For Wan 2.1 at 832×480 or LTX-Video at 768×512, resize your source image to those exact dimensions.
Use A Fill Layer That Matches Your Scene
If your foreground is a product on a white backdrop, composite it onto a white background that matches the product photography. If your foreground is a person cut from a location photo, composite them back onto the original location background. The closer the composite is to a natural scene, the fewer artefacts the model introduces.
Post-Generation Compositing With FFmpeg Alpha Extraction
When you need a final clip with transparency, generate your video with a known solid colour background and use FFmpeg to extract the key. The command compares each pixel to your chosen matte colour and creates an alpha channel from the difference.
FFmpeg alpha extraction is the standard method for compositing generated video. It produces a straight alpha output you can import into any video editor that supports RGBA video. The process works best with a bright green background (0,255,0 or #00FF00). Green is furthest from skin tones and common foreground colours in RGB colour space.
The limitation is edge quality. The generated video carries compression distortion from the MP4 encoding process, and that distortion interferes with the keying. A pixel not exactly your matte colour becomes partially transparent in the output. Adjust the tolerance parameter in FFmpeg to reduce this. Higher tolerance increases the risk of keying into your foreground. Test on a short clip before running on your full sequence.
Frequently Asked Questions
Why does my transparent PNG come out as black in the generated video?
The model's preprocessing pipeline discards the alpha channel and fills transparent regions with black as the default matte colour. That black fill is then treated as scene content. The model animates it alongside your foreground element, producing the distortion you see. Composite your cut-out onto a real background before uploading.
Does Runway Gen-3 Alpha support transparent PNG input?
Runway Gen-3 Alpha accepts PNG input per its documented input specification, but it does not preserve the alpha channel. The transparency is flattened to a solid fill colour during preprocessing. The documentation makes no claim of alpha support. Composite onto a background before upload.
Can I use Pika 2.0 region masks to avoid the transparent background problem?
Pika 2.0 region masks let you select which part of the image to animate. If your source image has a solid background, apply a region mask to your foreground and generate the clip. The unselected area stays static. You still need to composite the animated element onto a transparent background in post, but the region mask reduces background sliding during generation.
What colour should I use as the background fill instead of transparency?
A mid-tone grey or a blurred version of your intended final background works better than black or white. Pure colours cause the most distortion because the model has no texture to anchor the animation. Match the fill colour to the actual background your element will appear on in the final video.
How do I extract the alpha channel from a generated video?
Generate the clip with a bright green background (chroma key green, 0,255,0) and use FFmpeg chroma key extraction to create an RGBA video file. The command uses the colorkey filter to remove the green and generate an alpha channel. Expect some edge distortion from the MP4 compression and adjust the tolerance parameter on a test clip.
Does any image-to-video model output video with transparency?
No image-to-video model, as of June 2025, outputs video with an alpha channel. Every model listed in the dossier outputs MP4 with H.264 or H.265 encoding, which does not support transparency. The alpha channel must be achieved through post-generation compositing.
What is the difference between straight alpha and premultiplied alpha in this context?
Straight alpha stores transparency values in a separate channel with no colour modification. Premultiplied alpha multiplies the RGB values by the alpha value before storage. Most image-to-video tools output premultiplied alpha in PNG files but discard it on import. For video compositing, you want straight alpha output from FFmpeg extraction. It gives cleaner keying results in standard video editors.