Skip to content

first-last-frame-video.md

Latest commit

 

History

History
72 lines (49 loc) · 3.72 KB

first-last-frame-video.md

File metadata and controls

72 lines (49 loc) · 3.72 KB
Raw

Complex case: First–last frame video (flf2v)

Generate a short video (~5 s) that transitions from a first frame to a last frame. The model interpolates between the two frames. Use when the user asks for “首尾帧视频”, “first-last frame video”, or a short transition between two keyframes.

When to use

  • User wants a video that goes from one keyframe to another (e.g. camera push, time lapse, slight motion).
  • The transition should be short (~5 s); keep the change between first and last frame moderate (same scene/character, avoid full scene or character change).

Flow overview

Step Task Purpose
1 Get first frame t2i (if no image) or use given image as first frame
2 Get last frame i2i on first frame with a “last frame” edit prompt (small change: camera move, light change, small motion)
3 Generate video flf2v with first-frame and last-frame images → output video URL

Step-by-step

1. First frame

  • If user provides an image: Use that image as the first frame (local path or URL). You will pass it to i2i and then to flf2v.
  • If no image: Call t2i with a scene/character prompt and desired aspect_ratio (e.g. 9:16). Poll until done; get output_image URL → this is the first frame URL.

Example (script, no image):

# From skill dir; token/URL auto-loaded from openclaw.json if not set
./scripts/lightx2v_submit_and_poll.sh t2i "Qwen-Image-2512" "A person standing in a room, soft window light" --aspect-ratio 9:16
# → Save the printed URL as FIRST_FRAME_URL

2. Last frame (i2i on first frame)

  • Call i2i with:
    • input_image: first frame (URL from step 1 or base64 from user image)
    • prompt: edit that describes a small change (e.g. “Same scene, same character. Slight camera push-in; light shifts slightly as if 5 seconds passed.”). Avoid changing scene or character.
  • Poll until done; get output_image URL → this is the last frame URL.

Example (script, first frame as URL):

./scripts/lightx2v_submit_and_poll.sh i2i "Qwen-Image-Edit-2511" "Same scene and character. Slight camera push-in, soft light change, 5 second transition." --input-image "$FIRST_FRAME_URL" --aspect-ratio 9:16
# → Save the printed URL as LAST_FRAME_URL

3. flf2v (first–last frame to video)

  • Call flf2v with first-frame image (input_image) and last-frame image (input_last_frame). Use the same submit script with --input-image <first_frame_url> and --input-last-frame <last_frame_url>.
  • Poll until done; the script prints the output_video URL.

Example (after you have FIRST_FRAME_URL and LAST_FRAME_URL from steps 1 and 2):

./scripts/lightx2v_submit_and_poll.sh flf2v "Wan2.2_I2V_A14B_distilled" " " \
  --input-image "$FIRST_FRAME_URL" --input-last-frame "$LAST_FRAME_URL" --aspect-ratio 9:16
# → Prints the final video URL

So the full flow is: t2i (or user image) → i2i → flf2v, all with the same script; no extra curl or custom script needed.

Tips

  • Keep the change small: Same location and character; only camera, lighting, or small motion change so the ~5 s interpolation looks natural.
  • Prompt for i2i: Include “保持人物一致性” / “same character, same scene” when there is a character; describe only the small difference for the last frame.
  • Aspect ratio: Use the same ratio for t2i, i2i, and flf2v (e.g. 9:16 for portrait).
  • Deliver only the final video to the user; do not send intermediate first/last frame images unless asked.

See also

  • storyboard-video skill: flf2v flow with S3 and generate_flf2v_shot.sh.
  • LightX2V SKILL.md: task types, model list, and submit/poll workflow.