Matching Source Image Resolution and Aspect Ratio to Image-to-Video Models

You feed a perfect source image into an image-to-video model and get back a clip where your subject stretches sideways like a funhouse mirror or the output looks soft enough to have been shot through a fog filter. The cause is almost always a mismatch between your source image's resolution and aspect ratio and what the model expects. Each model is trained on specific aspect-ratio buckets, typically 16:9, 9:16, and 1:1, and its preprocessing pipeline resizes or letterboxes images that fall outside those dimensions. That naive resizing is what introduces the distortion. This page tells you what each model wants, what happens when you give it the wrong dimensions, and how to prepare your source images to avoid the two most common failure paths: aspect-ratio distortion and upscaling artefacts.

Image-to-video models do not understand perspective or physical shape. They predict pixel motion from patterns in their training data. When a training dataset is dominated by 16:9 clips, the model learns that a face should occupy a certain horizontal proportion of the frame. Feed it a 4:3 source image, and the preprocessing pipeline either stretches the image to fill 16:9, compressing the subject horizontally, or adds letterbox bars that the model then animates as part of the scene. Both produce distortion: subjects appear unnaturally wide or tall, and background sliding becomes more pronounced because the model treats the letterbox bars as a real spatial element.

The same problem applies to resolution. A model's claimed maximum output resolution, say 1080p, is often a post-processing upscale from a lower effective resolution produced by latent-space compression in the generation pipeline. If you feed a low-resolution source image and rely on the model's internal upscaling, you introduce softening and hallucinated detail where the model invents texture that was never in the original frame. The effective resolution gap, documented in the claimed-versus-real analysis of these models, is roughly one full tier: a model that outputs 1080p files typically has an effective detail level closer to 720p.

The table below lists each major image-to-video model's documented native output aspect ratios and resolutions, based on the model cards and official documentation available at the dates shown. This is what the model was trained to generate natively. Source images that match these dimensions go through preprocessing with the least distortion. Images that do not get resized or padded, and that preprocessing step is where artefacts originate.

Model	Documentation Date	Native Output Resolution (Landscape)	Native Output Aspect Ratios
Runway Gen-3 Alpha	2024-07	1280×768	16:9, 9:16
Kling 1.6	2025-01	1920×1080	16:9, 9:16, 1:1
Luma Dream Machine	2024-06	1360×752	16:9, 9:16
Pika 2.0	2024-12	1920×1080	16:9, 9:16, 1:1
Sora	2024-12	1920×1080	16:9, 9:16, 1:1
Hailuo Minimax	2024-09	1280×720	16:9, 9:16
Stable Video Diffusion 1.1	2023-11	1024×576	16:9, 9:16, 4:3, 3:4, 1:1
Wan 2.1	2025-02	1280×720	16:9, 9:16
LTX-Video	2024-11	768×512	16:9, 9:16, 1:1
Seedance 1.0	2025-03	1920×1080	16:9, 9:16

The most visible artefact from feeding a wrong aspect ratio is stretching or compression. If you give a 4:3 source image to Runway Gen-3 Alpha, which natively outputs 1280×768 (16:9) or 768×1280 (9:16), the preprocessing pipeline will resize the 4:3 image to match the target width. Because 4:3 is wider than 16:9, the image gets compressed horizontally. The result is that everyone in the frame appears thinner than in the original, and any vertical lines bow outward when the model animates motion. This is aspect-ratio distortion, and it is a direct consequence of the model having no training data for 4:3 output.

Kling 1.6, with its 1920×1080 native resolution and support for 16:9, 9:16, and 1:1, handles a wider range of inputs because it has been trained on square and portrait buckets. But feed it a 2.35:1 cinematic crop, and you get letterbox bars that Kling 1.6 treats as part of the scene. The bars then slide and warp independently of the main frame, creating a background sliding effect that ruins the illusion of a stable camera.

Stable Video Diffusion 1.1 is the most permissive among the open-weights models, with a documented aspect ratio range of 16:9, 9:16, 4:3, 3:4, and 1:1. Its native output resolution is 1024×576 landscape or 576×1024 portrait, so a 4:3 source will be resized to fit that narrower frame. The result is less distortion than a model limited to 16:9, but the subject still appears slightly compressed compared to the original.

How to Avoid Aspect Ratio Distortion

Crop your source image to the model's native aspect ratio before uploading. For Runway Gen-3 Alpha, that means cropping to 16:9 or 9:16. For Kling 1.6, crop to 16:9, 9:16, or 1:1. Do not rely on the model's automatic letterboxing; always match the aspect ratio exactly. Use a photo editor with a crop tool that shows pixel dimensions, not just percentages. The goal is to eliminate any black bars or non-uniform space around the subject.

The second common failure is feeding a low-resolution source image and expecting the model to upscale it internally. Every model's preprocessing pipeline resizes the source image to match its native input resolution, but that resize is a simple interpolation, typically Lanczos or bicubic, not a learned super-resolution. When the source image is significantly below the native resolution, the interpolation introduces softening. The model then generates motion on top of that softened base, which amplifies the blur. Pixels that were sharp in the source become indistinct blobs in the output.

The claimed-versus-real gap in resolution makes this worse. A model may output a 1920×1080 file, but independent frame analysis shows the effective detail is closer to 720p upscaled. That is the latent-space compression at work: the model encodes the image into a lower-dimensional latent space, generates motion there, and decodes back to the output resolution. The compression discards fine detail that cannot be recovered. If you start with a 720p source, the effective detail in the output will be below 720p, because the latent space drops another tier.

What Artefacts to Expect from Low-Resolution Sources

Softening is the primary artefact. Fine textures, hair, fabric weave, foliage, turn into smooth patches. The model may also hallucinate detail to fill what it perceives as missing information, adding random micro-patterns that were not in the source. This hallucinated detail is what creates the AI-generated look that viewers recognise as artificial. On a high-motion clip, the softening and hallucination combine to produce a shimmering effect where edges appear to vibrate between frames.

Runway Gen-3 Alpha recommends a source image of at least 1280×768 to match its native output. Kling 1.6 recommends 1080p or higher. Wan 2.1 accepts up to 1280×720. LTX-Video expects 768×512, 512×768, or 768×768. If you feed a source smaller than the native resolution, upscale it yourself with a good upscaling model, like ESRGAN or a dedicated image upscaler, before feeding it to the image-to-video model. That gives you control over the upscaling quality rather than relying on the model's simple interpolation.

Beyond aspect ratio and resolution, each model has specific source image format and size requirements. The following is based on the official documentation for each model at the dates listed.

Runway Gen-3 Alpha accepts PNG, JPG, and WebP files up to 10 MB. The model performs best when the input image matches the target aspect ratio. If you want a landscape output, the source image should already be 16:9. For portrait, 9:16. The documentation from July 2024 does not specify a minimum resolution, but matching the 1280×768 native output is the safe bet.

Kling 1.6 accepts PNG and JPG up to 10 MB, with a recommendation for 1080p or higher for optimal quality. The model supports start-and-end-frame specification, meaning you can provide both the first and last frame of the clip. This gives you more control over the final composition, but the source image for the start frame must still match the aspect ratio and resolution constraints.

Luma Dream Machine accepts PNG, JPG, and WebP up to 10 MB, with a June 2024 documentation. Its native output is 1360×752 landscape or 752×1360 portrait. This is an unusual resolution, not a standard 16:9 or 9:16, so you must crop your source to exactly 1360×752 or 752×1360 to avoid distortion.

Pika 2.0 accepts PNG and JPG up to 10 MB, with a December 2024 documentation. It outputs 1920×1080 landscape, 1080×1920 portrait, or 1080×1080 square. It also supports region-mask animation through its Pikascenes feature, allowing you to animate only a selected part of the frame. For region-mask to work correctly, the unmasked parts must stay static, which means the source image must be well-aligned with the mask boundaries. Any misalignment between the source aspect ratio and the mask dimensions will cause the static areas to shift when the model resizes the image.

Sora accepts PNG and JPG up to 10 MB, with a December 2024 documentation. It outputs up to 1920×1080 in landscape, 1080×1920 portrait, or 1080×1080 square, with a maximum clip length of 20 seconds. Sora's camera controls are accessed through prompts rather than a UI slider, so the source image must be descriptive enough for the model to interpret the desired motion. The aspect ratio constraints are the same as Pika 2.0: 16:9, 9:16, or 1:1.

Hailuo Minimax accepts PNG and JPG up to 10 MB, with a September 2024 documentation. Its native output is 1280×720 landscape or 720×1280 portrait, with a 6-second clip length. This is a lower native resolution than most commercial models, so the effective detail after latent-space compression will be below 720p. If you need higher resolution output, you must upscale externally.

Stable Video Diffusion 1.1 accepts PNG, JPG, and WebP with a resolution range of 576×1024 to 1024×576. The aspect ratio range is 16:9, 9:16, 4:3, 3:4, and 1:1. This is the most flexible in terms of aspect ratio among open-weights models. However, the clip length is only 4 seconds (25 frames at 6 fps) and the licence is non-commercial, so you cannot use the output in client work without special permission.

Wan 2.1 accepts PNG and JPG up to 1280×720, with a February 2025 documentation. It outputs 1280×720 landscape or 720×1280 portrait, with a 5-second clip length. The licence is Apache 2.0, which permits commercial use, making it one of the few truly open commercial models for image-to-video.

LTX-Video accepts PNG and JPG with specific resolution ranges: 768×512, 512×768, or 768×768. The November 2024 documentation confirms a native output of those same dimensions, with a 5-second clip at 24 fps (121 frames). This is the lowest native resolution among the listed models, so upscaling is almost always required for delivery.

Seedance 1.0 accepts PNG and JPG up to 10 MB, with a March 2025 documentation. It outputs 1920×1080 landscape or 1080×1920 portrait, with 5-second or 10-second options. The camera controls are pan, tilt, and zoom, accessed through the Jimeng platform.

You can spot aspect-ratio distortion before the clip finishes. The first frame might look correct because the preprocessing pipeline has not yet animated, but as motion starts, the stretched or compressed subject interacts with the frame edges in unnatural ways. A face that was correctly proportioned in the source appears elongated horizontally when the camera pans left, because the model has learned that face widths should occupy a certain fraction of the horizontal space. When the image is stretched to fill the model's native aspect ratio, that fraction is wrong.

Background sliding is another sign. If the subject and background move independently, the model has likely treated letterbox bars or non-native padding as part of the scene geometry. The bars shift when the camera moves, creating a parallax effect that breaks the illusion of a stable shot. This is particularly visible with Kling 1.6 and Sora, which have the highest native resolutions and are thus most sensitive to mismatched inputs.

Limb mutation and identity shift can also be exacerbated by wrong aspect ratios. When the model has less training data for non-standard aspect ratios, its temporal attention mechanism struggles to maintain the source image's conditioning signal across frames. The subject's arms may warp or its face may change ethnicity because the model is interpolating toward the average face distribution in its training data rather than anchoring to the source.

How to Test Your Source Image

Before feeding a source image to any model, resize it yourself to the model's native output resolution using FFmpeg or a photo editor. Then examine the resized version at 100% zoom. If the subject looks stretched or compressed at that stage, it will look worse after the model animates it. That static check catches 90% of aspect ratio issues before you waste a generation credit.

When you upscale a low-resolution source image, either yourself or through the model's pipeline, the artefacts compound. Softening is the first and most obvious: edges lose definition, fine details merge into blocks. Hallucinated detail follows, where the model invents texture that was never present, often a repeating pattern like canvas grain or film grain that is generic rather than specific to the scene.

Flicker, rapid frame-to-frame changes in brightness or colour, appears when the latent-space decoding is inconsistent across frames. This is common in models with a large claimed-versus-real resolution gap, because the decoder is working harder to interpolate missing information. LTX-Video, with its 768×512 native output, requires significant upscaling to reach 1080p delivery, and that upscaling process amplifies any flicker present in the low-resolution output.

If you must upscale, do it after generation rather than before. Use a dedicated video upscaler like Topaz Video AI or FFmpeg's ffmpeg-normalize with a real-time model. The reason is simple: if you upscale the source image before generation, the model generates motion at the upscaled resolution, which means it has more pixels to hallucinate. If you upscale after generation, you preserve the model's native output and only interpolate the final frame size, which introduces less artefacting.

The aspect ratio mismatch failure mode is distinct from other artefacts. It is not a subtle softening or a flicker that you might miss on a phone screen. It is a gross geometric deformation that is visible on the first play. Subjects appear to have been squeezed through a press. Lines that were straight in the source become curved during motion. The human eye is extremely sensitive to face proportions, so a stretched face is immediately recognisable as wrong.

The only fix is to crop the source image to the model's native aspect ratio before uploading. There is no post-processing fix for aspect-ratio distortion because the distortion is baked into every frame's pixel layout. You cannot un-warp a clip that was generated from a stretched input.

If you are working with a source image that has a unique aspect ratio, like a 2.35:1 cinematic frame or a 1.33:1 vintage photo, your options are limited. You can add letterbox bars manually in your source image, making the total frame equal to the model's native aspect ratio while keeping the subject in the correct proportions. The model will then animate the letterbox bars as static, which is less jarring than distortion. Or you can use a model like Stable Video Diffusion 1.1 that supports 4:3 and 3:4, giving you more flexibility.