OpenAI Sora Image-to-Video Generation: Accepted Image Inputs, Output Limits, and Access Conditions

You upload a single still image to OpenAI Sora and get back a moving clip. The OpenAI Sora image-to-video input accepts PNG, JPEG, and WEBP files up to 10 MB, at aspect ratios of 1:1, 16:9, or 9:16, according to OpenAI's help documentation at help.openai.com/en/articles/10472171-sora (checked 26 September 2026). The model treats that image as the first frame and predicts pixel motion from patterns learned during training. It does not understand the scene as a physical space, only as a statistical guess at what comes next. The most common mistake is expecting physically consistent output. Objects morph, limbs drift, and backgrounds warp because Sora lacks a persistent 3D world model.

Accepted File Types And Size Limit

OpenAI Sora accepts three image formats: PNG, JPEG, and WEBP. The maximum file size for any single image input is 10 MB. The documentation specifies three native aspect ratios: 1:1 (square), 16:9 (widescreen), and 9:16 (vertical/portrait). If your source image has a different aspect ratio, the pre-processing pipeline resizes and crops it to one of these three. That cropping introduces distortion: subjects stretch or compress unnaturally.

Output Resolution By Subscription Tier

The model outputs video at up to 1920×1080 (1080p) for Pro plan subscribers, and 720p for Plus plan subscribers (information from the OpenAI support page, checked 26 September 2026). The claimed resolution versus effective resolution gap is roughly one tier. Independent frame analysis suggests the effective detail is closer to 720p upscaled to 1080p, because the latent-space compression in the generation pipeline loses fine detail. Do not expect actual 1080p quality from a 1080p output file.

Sora advertises a native clip length range of 4 seconds to 20 seconds. The maximum native clip length depends on your subscription plan and the output resolution. On the Pro plan, you can generate up to 20 seconds at 1080p. On the Plus plan, the maximum clip length is 10 seconds at 720p (information from OpenAI's help documentation, checked 26 September 2026).

The claimed maximum clip length versus usable clip length gap applies here. Motion coherence degrades visibly after approximately 4 to 6 seconds for most generations. The model's attention mechanism loses track of the source image over longer sequences, causing subject drift, identity shift, and flicker. You can concatenate multiple short clips with FFmpeg, but seam artefacts appear at the join because Sora does not support true end-frame conditioning as of late September 2026. The output frame rate is variable up to 30 fps (information from the system card at openai.com/index/sora-system-card/, published 2024-12-09).

Sora is a cloud-only proprietary model from OpenAI. You access it through the web application at sora.com or through a ChatGPT Plus or Pro subscription. As of late September 2026, there is no API available and no programmatic integration into automated pipelines. The access path distinguishes Sora from Google Veo 2, which is accessed through Vertex AI and VideoFX. Sora is available in most countries where ChatGPT is available, but it is not available in the European Union, the United Kingdom, or Switzerland (information from OpenAI's platform availability page, checked 26 September 2026). The pricing structure is tiered: ChatGPT Plus and Pro subscriptions, with no free tier for generation. Local generation is not available; the model is not open-weights.

Sora is primarily a text-to-video model that also accepts image input. Its prompt adherence on the motion axis is high compared to many competitors, but it has documented failure modes. The model has a text-conditioning bias: when the text prompt conflicts with the source image, the model sometimes generates motion that bears no relationship to either. The Sora image-to-video prompt adherence is strongest when the prompt describes plausible motion from the still image (a person walking forward from a standing pose) and weakest when the prompt asks for physically improbable motion (a person floating upward from a seated position). The system card (published 2024-12-09) documents red-teaming domains including bias, misinformation, and content moderation failures. Prompt-ignored scenarios occur when the model's weak text-conditioning on the motion axis fails to override its default low-motion bias, producing a nearly static clip despite a prompt requesting significant movement. Motion collapse is the specific term for this failure.

Do not use Sora for image-to-video if you need physically consistent output longer than 5 seconds. The motion coherence degrades, the subject drifts, and the face changes identity. The one thing that most often goes wrong here is expecting a 20-second clip to hold together. Instead, generate multiple 4-second clips, select the one with the fewest artefacts, and concatenate them with FFmpeg. Accept that the median generation will have visible AI artefacts. Budget time for selection and post-processing. If you are in the EU, UK, or Switzerland, do not attempt to access Sora from those regions; choose an alternative model that is legally available to you.