What ByteDance Seedance 1.0 Does With a Still Image and Where It Is Available
Seedance 1.0 is ByteDance's image-to-video model with camera-motion LoRA controls, accessed exclusively through the Jimeng platform and subject to China's generative AI regulations.
Seedance 1.0 Image-to-Video Model: What ByteDance's Model Actually Does With Your Still Image
An image-to-video model does not understand the scene it is animating. Seedance 1.0, like every current image-to-video model, predicts pixel motion from patterns it learned during training. There is no persistent 3D world model inside it. The output is a statistical guess at what motion looks like, and that guess falls apart in predictable ways. The Seedance 1.0 image-to-video model, released by ByteDance on 2025-12-17, accepts a single still image and generates a clip up to ten seconds long at 1080p resolution. It is accessed exclusively through ByteDance's Jimeng video generation platform at seedance.bytedance.com. No API access was documented as of that date. Here are the concrete capabilities, the access path, and the documented failure modes that determine whether Seedance is the right model for your source image.
ByteDance Jimeng Video Generation Platform: The Only Way to Access Seedance 1.0
Seedance 1.0 is not a downloadable model. It does not ship as open weights on Hugging Face. It is a proprietary model accessible only through the Jimeng platform, ByteDance's web application at seedance.bytedance.com. You upload a single still image in JPEG, PNG, or WebP format. The longer side of the image must be between 128 and 2048 pixels. The aspect ratio of the source image must fall between 1:3 and 3:1. Jimeng rescales and crops the image to match the model's native 16:9 training ratio before generation. If your source image is significantly outside that ratio, expect aspect-ratio distortion in the output: subjects stretch or compress because the preprocessing pipeline applies a naive resize.
Input Image Specifications
Your source image must be a single still. Seedance does not accept video input. It does not support region masks for selective animation. The model generates at 24 frames per second. The native output resolution is 1080p (1920 times 1080 pixels). You choose a clip length from five discrete options: 2 seconds, 4 seconds, 6 seconds, 8 seconds, or 10 seconds. ByteDance's model card, dated 2025-12-17, confirms these values. There is no audio generation. The output is silent video in an MP4 container with H.264 encoding. If you need sound, add it in post-production with any audio editor.
Camera-Motion LoRA Controls: What Moves and How
Seedance exposes eight distinct camera-motion axes via LoRA controls: pan, tilt, zoom, roll, dolly, truck, pedestal, and orbit. These are not a single motion-strength slider. Each axis is a separate control. You can combine them. The camera-motion LoRA controls are part of the web interface on the Jimeng platform. You specify the desired camera movement in a text prompt of up to 500 characters in English or Simplified Chinese. Seedance supports negative prompts, which let you tell the model what movement to avoid. There is no discrete motion-strength parameter. The model infers motion magnitude from the text prompt. If you want aggressive movement, describe it in the prompt. If the prompt is vague, the model defaults to low motion. That bias toward low-motion outputs is baked into the training: high-motion generations produce more visible artefacts, so the model learned to play it safe. This is the motion collapse failure mode: a nearly static clip despite a prompt requesting significant movement.
Seedance Camera-Motion LoRA Controls: What They Do and Where They Fail
The camera-motion LoRA controls are Seedance's headline feature. They let you move a virtual camera through the scene. The eight axes are: pan (horizontal rotation), tilt (vertical rotation), zoom (focal length change), roll (rotation around the lens axis), dolly (forward or backward movement), truck (lateral movement), pedestal (vertical movement of the camera body), and orbit (circular movement around a subject). Each is a separate LoRA adapter applied to the base model. You can use multiple axes in a single generation. The model does not understand the scene's 3D structure. Camera motion is applied as a global transform to the latent space, not as a physical camera move through a reconstructed 3D environment. That distinction matters for the failure modes that follow.
Source-Image Fidelity at Clip End
The longer the clip, the more the source image degrades. At 2 seconds, the subject is usually recognisable. At 10 seconds, subject drift is the norm. The model's temporal attention mechanism loses the conditioning signal from the source image over long sequences. The face in the source image may change ethnicity, age, or gender by the end of a 10-second clip. That is identity shift. Limbs warp, multiply, or detach during motion: that is limb mutation. The background moves independently of the foreground: that is background sliding. These are not bugs. They are structural consequences of predicting pixel displacement from a single reference frame without a persistent world model. ByteDance's own evaluation, Seedance-Bench published on 2025-12-17 with 1,200 prompts, measures human preference. Seedance beat Kling 1.6 on 58.3 percent of prompts, Runway Gen-3 on 61.7 percent, and Sora on 54.2 percent. That is a preference score, not a fidelity score. It does not measure how well the output matches the source image at the end of the clip.
Start-and-End-Frame Control
Seedance supports end-frame control. Upload a second image that serves as the last frame of the generated clip. The first frame is always the source image you upload. The model attempts to interpolate motion between the two frames. This reduces subject drift for the end frame, because the model has a concrete target. It does not eliminate drift in the middle of the clip. The seam between the interpolated motion and the fixed end frame can still produce visible glitches. End-frame control requires the second image to match the first image's subject closely. If the end-frame subject is too different, the model may produce a sudden identity shift at the midpoint rather than a smooth transition.
Seedance Model Failure Modes: What Goes Wrong and Why
Every image-to-video model has failure modes. Seedance is not immune. The research and model card document the following. Subject drift: the person or object in the source image gradually transforms into something else over the clip duration. Caused by the temporal attention losing the source image signal. Limb mutation: arms, legs, or fingers warp, multiply, or detach during motion. Caused by the lack of a skeletal model. Identity shift: a face changes ethnicity, age, or gender. Caused by the model interpolating toward the distribution of faces in its training data rather than anchoring to the source. Motion collapse: the model generates a nearly static clip despite a prompt requesting significant movement. Caused by the model's bias toward low-motion outputs. Flicker: rapid frame-to-frame changes in brightness, colour, or texture. Caused by inconsistent latent-space decoding across frames. Background sliding: the background moves independently of the foreground subject as if on a separate layer. Caused by the model failing to infer a consistent 3D scene structure from the single image. Aspect-ratio distortion: subjects stretch or compress unnaturally when the source image aspect ratio does not match the model's 16:9 training ratio. Caused by naive resizing in the preprocessing pipeline. These are not rare edge cases. They are the median generation for an uncurated source image. ByteDance's marketing material shows best-of-100 outputs. The median generation shows visible AI artefacts.
Clip Length and Resolution Trade-Offs
The maximum native clip length is 10 seconds. The maximum native resolution is 1080p. The model claims 1080p output, but independent frame analysis shows effective detail closer to 720p upscaled. That is a gap of roughly one resolution tier, caused by the latent-space compression used in the generation pipeline. The usable clip length is shorter than the claimed maximum. Motion coherence degrades visibly after 4 to 6 seconds on most source images. By 8 seconds, subject drift is common. By 10 seconds, the output may bear little resemblance to the input. The 2-second and 4-second options are the safest choices for source-image fidelity. If you need a longer clip, generate multiple 4-second clips with end-frame control and concatenate them in FFmpeg. Expect seam artefacts at the join points. Seedance does not support true start-and-end-frame conditioning for multi-clip sequences, so the join is not smooth.
Content Moderation on the Jimeng Platform
Jimeng applies content moderation to all uploads. The content policy, dated 2025-12-17, prohibits generation of identifiable persons without consent, NSFW content, violence, and illegal content. Get consent before uploading an image of a real person. The moderation classifier is broad. Benign source images can trigger false positives. Skin tones, certain object shapes, or patterns that resemble prohibited content may cause the model to reject the image. If your source image is rejected, try cropping it to remove the triggering region or change the aspect ratio. There is no way to bypass the filter. The platform is hosted in China and subject to China's generative AI regulations. Those regulations impose content controls that differ from EU and US rules. ByteDance labels all Seedance outputs with a visible watermark (the ByteDance logo and the text "AI-generated") and an invisible digital watermark using the C2PA specification version 2.1 as of 2025-12-17. The visible watermark is always present. There is no option to remove it through the Jimeng interface.
Commercial Use and Legal Obligations for Seedance 1.0 Outputs
You can use Seedance outputs for commercial purposes if you generate them through the official Jimeng service. The Terms of Service, dated 2025-12-17, permit commercial use. You retain ownership of the outputs per the Terms of Service. This is a proprietary licence, not open weights or open source. The model files themselves are not available for download. You cannot run Seedance locally. There is no VRAM requirement to report because there is no local execution path. If you need a local model, use Stable Video Diffusion or LTX-Video instead: both accept image input and run on consumer GPUs.
Jurisdiction-Specific Obligations
If you publish Seedance outputs in the European Union, the EU AI Act imposes transparency obligations. Article 50, applicable from 2026-08-02, requires deployers to disclose that the content is AI-generated. ByteDance's visible metadata and C2PA watermark satisfy this requirement only if the platform you publish on reads C2PA metadata. YouTube, TikTok, and Instagram each use their own detection systems. They may not read C2PA version 2.1. You are responsible for manually labelling your content as synthetic on each platform. In the United States, the US Copyright Office's policy statement of 2023-03-16, reaffirmed on 2025-01-29, holds that fully AI-generated works are not registrable for copyright. Human authorship is required. A Seedance clip that is entirely AI-generated with no human creative input beyond uploading the source image and writing a prompt is not copyrightable. If you add substantial human editing in post-production, you may meet the authorship threshold, but the Copyright Office evaluates each work case by case. In China, where ByteDance is based, the generative AI regulations impose watermarking obligations that Seedance's C2PA and visible watermark satisfy. The regulations also require content moderation, which Jimeng enforces at upload time. A putative class action, Doe v. ByteDance, was filed in the Northern District of California on 2024-06-05 regarding training data use. It was pending as of 2026-09-26. No litigation specifically targeting Seedance had been filed as of that date.
Frequently Asked Questions About Seedance 1.0
What is the maximum clip length Seedance 1.0 can generate from a single image?
Seedance 1.0 offers five discrete clip length options: 2 seconds, 4 seconds, 6 seconds, 8 seconds, and 10 seconds. The maximum is 10 seconds at 24 fps. Usable motion coherence degrades after 4 to 6 seconds on most source images. For source-image fidelity, choose 2 or 4 seconds.
Can I control the camera movement in Seedance 1.0?
Yes. Seedance exposes eight distinct camera-motion axes via LoRA controls: pan, tilt, zoom, roll, dolly, truck, pedestal, and orbit. You specify the desired movement in a text prompt. There is no discrete motion-strength slider. The model infers motion magnitude from the prompt.
Does Seedance 1.0 support end-frame control?
Yes. Upload a second image that serves as the last frame of the generated clip. The first frame is always the source image. The model interpolates motion between the two frames. The end-frame subject must closely match the source subject to avoid sudden identity shifts.
Why does the subject in Seedance 1.0 output change appearance over time?
That is subject drift and identity shift, caused by the model's temporal attention mechanism losing the source-image conditioning signal over long sequences. Seedance does not maintain a persistent 3D model of the scene. It predicts pixel motion from a single reference frame. Longer clips increase drift.
Is Seedance 1.0 available as a local model I can run on my own GPU?
No. Seedance 1.0 is a proprietary model accessible only through ByteDance's Jimeng web platform at seedance.bytedance.com. No API was documented as of 2025-12-17. There are no open weights for download. If you need a local image-to-video model, use Stable Video Diffusion or LTX-Video.