ByteDance
Seedance 2.0 is the upstream video model whose clean 1080p output makes frame extraction productive — its frames are well-composed, stable, and high-detail enough to serve as image-edit inputs or static deliverables. The pipeline is: generate or take an approved Seedance clip → route through the frame extraction tool node → pick the strongest frame(s) for the next step. Common downstream uses: feed the extracted frame as a reference image to a Nano Banana 2 or Flux Kontext image edit, use it as the starting frame for a different camera move on Sora 2 or Kling 3.0, or export as a static hero image. The companion `tools/frame-extraction` page covers tool routing and parameters; this how-to focuses on the Seedance-paired pipeline.
A 5-second Seedance clip at 24fps contains 120 frames; you don't need them all. Decide upfront what the extracted frames will be used for, and the right extraction strategy follows. (1) Single hero frame for image-edit chain: extract the strongest single frame, usually around the 60-80% mark of the clip when motion has settled and composition is final. (2) Starting frames for next-shot generation: extract the last frame for natural continuation, or the first frame for a camera-move variation. (3) Static deliverables: extract 2-3 candidate frames (early, middle, late) and pick the best on the canvas. (4) Storyboard panels: extract every 0.5-1 seconds to convert the clip back into a sequence of stills.
Add a Tool node and select the workspace's frame extraction route. Connect the Seedance Video node's output to the Tool input. Configure the extraction mode: "every frame" (full sequence, useful for storyboard conversion or animation reference), "every Nth frame" (1 frame per second is the common choice), "specific timestamp" (e.g., 2.4s, 3.1s, 4.0s), or "first/last frame only" (the lightest mode). Output is PNG by default for transparent compositing, or JPG for smaller file size. Each extracted frame becomes an Image node on the canvas, ready for the next step.
Lay all extracted frames out side-by-side on the canvas. Quick triage: is the subject sharply rendered? Is composition balanced? Is lighting clean? AI-generated motion sometimes produces 1-2 frames per clip with subtle artifacts (smeared limbs, drifting eyes, soft edges) that the eye catches at the still resolution but missed at video playback speed. Pick the frame where motion has settled and composition reads cleanly — typically not the very first or very last frame, but somewhere in the middle 60-80% of the clip. For a hero image deliverable, this is also where you'd optionally route through the image upscaler tool node (see upscale-images-to-4k how-to) to push the frame from 1080p source to 4K still.
The extracted frame becomes a regular Image node, so the canvas treats it like any other image. Common next steps: (1) Image edit chain — feed the frame into Nano Banana 2 or Flux Kontext as a reference for outfit swaps, scene changes, or aesthetic restyles. (2) Next-shot generation — feed the frame as the starting frame to a different video model (Sora 2 for narrative continuation, Kling 3.0 for human-motion shots) for a multi-shot sequence with consistent world. (3) Storyboard chain — extract a sequence of frames (every 0.5s) and feed each through Nano Banana 2 with a "match aesthetic" prompt to produce a clean storyboard from the AI clip. The Martini canvas template pattern lets you save this arrangement and re-run for the next clip.
Decide the use case before extracting. Single hero frame, starting frame for next shot, static deliverable, or storyboard sequence — each maps to a different extraction mode.
Common extraction modes: "every Nth frame" with N=24 (1 per second at 24fps) is the default; "specific timestamp" for hero shots; "first/last frame only" for image-to-video chaining.
Output PNG for transparent compositing or for the next step needing alpha; JPG for static deliverables and storyboard panels where file size matters.
Frames from the middle 60-80% of the clip usually have the best composition — motion has settled, framing is final. Avoid first/last frames unless specifically chaining for video continuation.
Companion tool page: `models/tools/frame-extraction` covers extraction modes, output formats, and routing in detail. This how-to is the Seedance-paired pipeline specifically.
Frame extraction from a Seedance 2.0 clip is the canvas's "video → image" bridge — clean 1080p Seedance frames serve well as image-edit inputs, next-shot starting frames, or static hero deliverables. Trade-off vs. Kling 3.0 paired: Seedance is the cleaner choice for product, environment, and modern lifestyle content where composition and stability matter; Kling 3.0 is the right pick when extracting frames featuring human performers because Kling's motion engine produces cleaner per-frame body/face detail. Both extract through the same tool node — the difference is which video source feeds the pipeline. The Martini canvas keeps the source clip, extracted frames, and downstream image/video nodes adjacent, so frame harvesting becomes a fluid step in a multi-shot or image-chain workflow rather than a context switch to an external tool.
Connect Seedance 2.0 with other AI models on Martini's infinite canvas. No GPU required — start free.
Get Started Free