Skip to content

First Shot: Completing the Loop with Image-to-Video ​

When to Use This ​

With the storyboard done and scene reference images in place, the next step is turning the first shot into video. This page walks you through the Wan2.2 workflow in ComfyUI to run image-to-video: feed in one static reference image and get out a dynamic clip a few seconds long. The goal of this step is to complete the loop and get your first piece of footage—don't aim for perfection on the first try.

Haven't done shot planning yet? Start with the shot generation guide; missing reference images? See scene reference images.

Preparation ​

  • Environment: ComfyUI installed and the Wan2.2 models downloaded (via both HuggingFace and ModelScope; ModelScope works in China). If you haven't set up the environment, see ComfyUI Installation first.
  • Model: For the first shot, 5B (TI2V-5B) is recommended—under ComfyUI's native offloading it runs well with just 8GB of VRAM. The 14B I2V has a much higher VRAM threshold (the official repo scripts note at least 80GB on a single card), so skip it for now.
  • Cloud alternative: Comfy Cloud (cloud.comfy.org) offers templates such as video_wan2_2_5B_ti2v, no local deployment needed.@@I18N_KEEP_000018@@
  • Pick a shot: Choose a frame with a clear start and end state (as the course design suggests), such as "the character pushes open the door and enters the room"—it validates image-to-video motion far better than "the character stands still".

Follow Along: Three Steps to Get It Running ​

Step 1: Open the Image-to-Video Template ​

In the ComfyUI template library, search for "Wan2.2 5B" and load the TI2V workflow. It is configured for pure text-to-video by default, with the Load image node disabled—select it and press Ctrl+B to enable it. For 14B, search for "Wan2.2 14B I2V" instead.

Step 2: Upload the First Frame, Set Dimensions and Frame Count ​

Upload shot one's reference image to Load image, then set the width, height, and total frame count length in Wan22ImageToVideoLatent. The 5B 720P spec is 1280×704 (landscape) or 704×1280 (portrait), at 24 FPS; length divided by 24 gives the duration in seconds—for your first attempt, aim for around 5 seconds (about 120 frames).

In the 14B I2V workflow, the corresponding nodes are Load Image (for the first frame) and EmptyHunyuanLatentVideo (for adjusting length).

Step 3: Write the Prompt, Generate, Review ​

Write the prompt describing this shot's action and atmosphere; you can follow the official examples for structure. The official examples use shot type (close-up) + background elements + atmosphere, such as "a white cat with a surfboard on a summer beach." Once written, add it to the queue and generate: on a single consumer-grade GPU without special optimization, TI2V-5B can produce a 5-second 720P video within 9 minutes. After it finishes, review the result first and confirm it's fine before queuing the next one.

Checkpoint ​

  • Duration: Does frame count ÷ 24 equal the expected number of seconds?
  • Consistency: Are the first frame's composition and the character's appearance preserved?
  • Motion: Did the image move? Is the action natural, with no obvious glitches?
  • Out of VRAM: Confirm you're using the 5B workflow; with even lower VRAM, try the GGUF-quantized community versions (bullerwins / QuantStack repos + City96/ComfyUI-GGUF, community resources).
  • If the same shot fails repeatedly, first go to shot failure diagnosis to pinpoint the issue, then consider switching models.

Boundaries and Next Steps ​

  • Each generation produces a short clip of a few seconds; a complete short drama relies on editing multiple clips together—Runway's official long-video pipeline likewise works by "generating multiple short clips and then editing."
  • This course does not promise one-shot success or absolute character consistency; descriptions like Wan2.2's "cinematic-grade aesthetic control" come from official marketing—judge by your own testing.

Next, move on to action and camera-movement prompts to write image-to-video prompts more precisely and make shots move the way you intend.

gaodong.aiAboutPrivacyContact

Independent tutorials, not affiliated with any product mentioned. Product names and trademarks belong to their owners. Check official pages for prices and policies.