Preparing Source Images for Image-to-Video Generation

An image that looks sharp on screen will break into flicker, background sliding, or limb mutation the moment a model tries to animate it. This page is the hub for the image-to-video source image preparation guide. Every guide beneath this one starts with a specific source-image flaw, compression artefact, wrong aspect ratio, missing depth cue, and shows what it becomes once the model amplifies it. The human eye forgives a lot in a still. A video model does not. What you must check before you generate is what this diagnostic framework gives you: inspect resolution, aspect ratio, compression artefacts, composition depth, and source type first, because each of these maps directly to a named failure mode in the output.

Start with resolution. A model like Stable Video Diffusion 1.1 has a native resolution of 576×1024 or 1024×576, those exact numbers, from its training data. Feed it a source image that is smaller and it will upscale it, introducing softening and hallucinated detail that was never in the photograph. Feed it a 1080p image and the model downsamples to its latent space, losing sharpness. The rule: match the source image's pixel dimensions to the model's native output resolution as closely as possible. For models that accept 1080p natively, Kling 2.5, Runway Gen-4, Sora, use a 1920×1080 source. For SVD, stick to 576×1024 or the landscape equivalent. Upscaling after generation is a separate step and should never be done to the source image to compensate for a mismatch.

Aspect ratio comes next. Models are trained on specific frame shapes: 16:9, 9:16, 1:1. Forcing a 4:3 or 21:9 image through a model built for 16:9 produces stretching or letterboxing that the model then interprets as part of the scene, causing aspect-ratio distortion in the motion. The subject appears squeezed or the background slides to fill the wrong shape. Use the model's documented native aspect ratio. If your source image is a different shape, crop it to the correct ratio before feeding it in, do not let the preprocessor do it automatically, because automatic resizing is where the distortion starts. The documentation for Veo 3 lists accepted image formats (JPEG, PNG, WebP) but does not list aspect ratio; you match it to the output resolution you request, 1280×720 or 1920×1080, both 16:9.

Compression artefacts are the most common hidden flaw. JPEG blocking from a low-quality export or a heavily compressed social-media download looks like noise to the human eye but becomes flicker in the animated output, the model tries to animate the block boundaries as if they were part of the subject, producing rapid frame-to-frame brightness shifts. The same happens with WebP images that have been re-compressed. The fix is to use a PNG source, lossless at the point of export, and to avoid resaving a JPEG that has already been compressed. If you must use a JPEG, set the quality to 95 or higher in the export dialog. Models do not forgive compression; they hallucinate motion on top of it.

Composition for motion is about headroom and depth cues. A subject placed dead centre with no space to move, no headroom above, no leading lines, no clear foreground-background separation, produces motion collapse. The model has nowhere to send the motion and defaults to a nearly static clip, even with a high motion-strength setting. This is not a model failure; it is a physics failure that the model cannot solve. A person standing against a flat wall with no depth cue will not walk because the model has no way to infer a 3D path. The same image with a shadow on the floor, a receding wall line, or an object in the foreground gives the model something to push against. The depth cue is what tells the latent diffusion architecture where motion should go. Add it in the source image before generating.

If your subject has no room to move, start with the composition for motion guide, it covers how to add headroom and depth cues without reshooting. If your image is the wrong shape, start with the resolution and aspect ratio guide. If you are using an AI-generated still, read the AI source images guide. An AI-made image already sits in the same latent distribution as the model's training data, which means the model compounds its own statistical preferences on top of the source rather than anchoring to an external reality. This produces identity shift, a face changes ethnicity or age over the clip, and limb mutation, because the model interpolates toward its training examples rather than preserving the source. The effect is worse with models that have weak temporal attention, but it happens on all of them. A photograph from a camera does not drift this way. An AI-generated still does.

This page routes, it does not answer. The depth belongs to the guides beneath it. What you get here is the diagnostic framework and the order to apply it: resolution and aspect ratio first (wrong shape), then compression artefacts (hidden noise), then composition (no motion room), then source type (AI vs photographic). Each of these becomes a named failure mode, flicker, aspect-ratio distortion, background sliding, motion collapse, identity shift, limb mutation, in the output. The connection between source flaw and output failure is documented in the model-card documentation and in the dossier's failure-mode analysis. It is not speculative. Every model listed in the fact sheet below shows the pattern: motion coherence degrades as source quality drops, regardless of the provider.

How to Read This Table

The native resolution column is the number that matters first. A model that outputs 720p natively will soften a 1080p source during downsampling, and a model that outputs 1080p natively will add hallucinated detail to a 720p source during upsampling. The claimed resolution and the effective resolution are not the same thing: independent frame analysis shows that many models that output 1080p files have effective detail closer to 720p, because the latent-space compression in the generation pipeline loses data. The native clip length is the second number to check, but treat it with the same scepticism. Many models advertise 10-second generation, but motion coherence degrades visibly after 4-6 seconds, the gap between claimed and usable length is typically 40-60% of the advertised duration, caused by the attention mechanism losing track of the source image over long sequences.

Source Image Requirements for AI Video

Every model in the table accepts JPEG and PNG. Veo 3 and Runway Gen-4 also accept WebP, and Sora accepts all three. The file format matters less than the compression level inside it. A JPEG at quality 60 from a social-media download will fail. A PNG exported losslessly from Photoshop or GIMP will not. For open-weights models like Stable Video Diffusion and AnimateDiff, the source image must be resized to the model's native resolution before inference, the pre-processing pipeline in ComfyUI or Hugging Face Diffusers does not do it for you, and leaving it to the model's own resizer introduces the same artefacts as a bad export.

What Makes a Good Image-to-Video Input

A good source image has three characteristics: it is at the model's native resolution, it has a clear depth cue (foreground, subject, background), and it contains no compression artefacts visible at 100% zoom. A photograph from a camera with a clear focal plane beats an AI-generated still every time because the photographic source has a real, non-synthetic distribution of edges and textures that the model does not try to "correct" toward its own training data. If you are using an AI-generated source image, the risk of identity shift and limb mutation increases proportionally to the amount of synthetic detail in the source. The model amplifies what is already a statistical guess.

Pre-Animation Image Checklist

Run this checklist on every source image before you generate, in this order:

• Is the resolution within one tier of the model's native output? For a 1080p model, use 1920×1080. For SVD, use 576×1024. For AnimateDiff, use 512×512.
• Is the aspect ratio exactly the model's training ratio? 16:9 for almost all commercial models. 1:1 for some open-weights models. Crop, do not letterbox.
• Is the file a lossless format (PNG) or a JPEG at quality 95+? If it came from a website, re-export it.
• Does the subject have headroom and a depth cue? A floor line, a shadow, a receding wall, something that tells the model where 3D space is?
• Is the source image a photograph, not an AI generation? If it is AI-made, expect drift and plan for it.

Image-to-Video Input Failure Prevention

Prevention is a single step: apply the checklist before inference, not after. Once the model has generated a clip with background sliding or motion collapse, the only fix is to regenerate with a corrected source image. Post-processing with FFmpeg, frame extraction, concatenation, format conversion, cannot repair a clip that has already lost subject fidelity because the source image was wrong. The failure modes are structural: they come from the latent diffusion architecture itself, which predicts pixel motion from a single reference frame and has no persistent 3D world model. The model does not understand the scene. It makes a statistical guess at what motion looks like. If the source image gives it bad data, the guess is bad everywhere.

Start with resolution and aspect ratio. If the image is the wrong shape, everything downstream is built on a distorted base. Then check compression artefacts, they are invisible at a glance but destroy temporal consistency. Then check composition depth. Then check source type. Each of these is a separate guide in this section, and each guide diagnoses a specific flaw and tells you what it becomes once animated. This page is the map. The guides beneath it are the detail.

An honest caveat: no amount of source-image preparation will make a short-clip model produce physically consistent motion on a complex scene. The architecture does not have a persistent 3D world model. Limbs will drift, backgrounds will slide, and the subject will drift toward the model's statistical mean over the clip duration. The best source image only reduces the severity of the failure; it does not eliminate it. The models that come closest to physical consistency, Veo 3 and Sora at 1080p, still degrade after 6-8 seconds. For anything longer, you concatenate multiple short clips, and each join is a seam artefact risk. If your project requires physical consistency for more than ten seconds, image-to-video is not the right tool. Start with a 3D render or a real camera, then add generative elements on top. That is where this catalogue ends and another tool set begins.