Skip to content

Shot Generation: From Character Reference Images to Usable Clips ​

💡 Goal: After finishing this station's roughly 35-minute course (rendering wait time not included), you will turn each shot on your storyboard into a usable clip—your first shot will run on the spot, and you will have everything at once: prompt writing for AI shot generation, transition control, failure diagnosis, and retry discipline.

🎯 Stage Goals ​

  • Run your first shot: use a scene reference image as the first frame to run image-to-video and produce your first usable clip of 3–5 seconds (3–5 seconds is the shot length design value for this course)—see 4.1 The First Shot
  • Write prompts that convey "motion" and keep transitions controllable: describe action, camera movement, and atmosphere separately, with the writing style broken down sentence by sentence from official example prompts; use the first-and-last-frame workflow for transitions and displacement shots, letting the shot move from frame A to frame B—see 4.2 Action and Camera Movement Prompts, 4.3 First and Last Frame Control
  • Diagnose failures and cap your retries: locate the problem within 60 seconds; at most 3 retries per shot—beyond that, replace it or let it go—see 4.4 Failure Diagnosis, 4.5 Retry and Replacement

⏳ Estimated Time ​

  • The five lessons total about 35 minutes (8 + 7 + 7 + 7 + 6); with a ±30% buffer reserved, about 25–46 minutes
  • Rendering wait time not included: official speed statement—without dedicated optimization, TI2V-5B can generate a 5-second 720P video on a single consumer-grade GPU within 9 minutes. Factoring in retries, budgeting a wait on the order of half an hour per shot is safer (this site's estimation recommendation)
  • Short on time route: 4.1 The First Shot is a must; read 4.2 when the first result disappoints; read 4.3 when you hit transition shots; come back to 4.4 and 4.5 when problems arise

🚀 Before You Start ​

  • Completed Character and Scene Assets: you have character reference images and scene reference images in hand—image-to-video will use them as first frames
  • Set your production route: 1.3 Cloud Tools or Local Deployment
  • Local route: ComfyUI updated to the latest version (required by the official tutorial; if you can't find templates in the template library, it's most likely an outdated version), and the Wan2.2 model files are in place—enter via ComfyUI Quick Start; for where to put models, see Models Folder Guide
  • Cloud route: your Comfy Cloud account is set up (cloud templates such as video_wan2_2_5B_ti2v can be opened directly), or the cloud platform account chosen on the 1.3 Routes page is ready
  • Storyboard in hand: the 2.4 Storyboard is finalized—this station is about turning it into video line by line
  • Assets archived per the 3.5 Naming Conventions: no matter how many retries, versions won't overwrite each other
  • A cup of water ☕—every shot involves waiting for rendering, a perfect time to drink water

📚 Lessons in This Stage ​

The five lessons answer five questions in order: How do you run the first shot? How do you write prompts that convey "motion"? How do you control transitions? How do you investigate failures? How many retries are enough?

LessonContentDuration
4.1 The First ShotRun your first shot with image-to-video: a scene reference image as the first frame8 minutes
4.2 Action and Camera Movement PromptsOfficial examples broken down sentence by sentence + camera movement vocabulary template7 minutes
4.3 Controlling Shot Direction with First and Last FramesFirst frame + last frame lock in the direction; use it for transitions and displacement shots7 minutes
4.4 Shot Failure DiagnosisLocate the problem within 60 seconds: triage environment issues vs. content issues7 minutes
4.5 Shot Retry and Replacement StrategyUp to 3 attempts per shot: retry, downgrade, or let it go6 minutes

Four Official Workflows: How to Choose for AI Shot Generation ​

Wan2.2's ComfyUI official tutorial provides four workflows, all of which can be found in the template library. This section only covers "how to use them for short-drama shots"; for node-level operations, see the dedicated page in the comfyui section:

Official WorkflowTemplate Library Search NameWhen to Use It in Short DramasNode-Level Tutorial
Wan2.2 5B (text-to-video + image-to-video in one)Wan2.2 5BMain path for practice films (recommended in this course): one model, two uses, runs on 8GB of VRAMWan2.2 5B workflow
Wan2.2 14B T2V (text-to-video)Wan2.2 14B T2VNo suitable reference image; generate visuals directly from text14B T2V workflow
Wan2.2 14B I2V (image-to-video)Wan2.2 14B I2VUse the reference image as the first frame; prioritize image quality14B I2V workflow
Wan2.2 14B FLF2V (first-last frame)Wan2.2 14B FLF2VTransitions and displacement shots: the first and last frames set the direction14B FLF2V workflow

In the 14B workflow you will see two diffusion model files (high-noise + low-noise) — this is Wan2.2's MoE dual-expert design: the high-noise expert handles overall layout, while the low-noise expert handles details (official README: 27B total parameters, 14B activated per step). The 5B workflow has only one diffusion model file, which is friendlier for practice films.

💡 Why is 5B the first choice for practice films?

  • Quoted from the official tutorial: ComfyUI under native offloading, the 5B "should fit well on 8GB vram" (it can run on 8GB of VRAM) — entry-level GPUs can handle it
  • One model supports both text-to-video and image-to-video, covering both needs of a practice film
  • The official positioning of the 5B is "rapid prototyping and creative exploration" — exactly what a practice film needs: fast trial and error, not one-shot perfection
  • 14B is officially positioned more toward image quality (workflow description: high-quality, cinematic-grade aesthetic control), but it is more demanding on hardware: the official repository's command-line scripts specify at least 80GB of VRAM on a single card for the 14B (ComfyUI repackaged version does not specify a number). When VRAM is tight, 4.4 Failure Diagnosis has a VRAM-reduction path via community GGUF quantized versions

Let's clarify the boundary up front (this is also the boundary for the whole course): AI shot generation means generating each shot separately, drawing multiple takes, and finally editing them together, not generating the entire film at once; reference images can improve character consistency but cannot guarantee "the face never changes across shots" — residual drift in face shape and clothing must be caught by 2.5 Shot Continuity Check and 4.3 First-Last Frame as fallbacks.

📋 Official Resources Quick Reference

🎒 What You Will Gain ​

  • A batch of usable 3–5-second clips archived by shot number from the storyboard, ready to go into the final cut for editing
  • A shot prompt-writing method of "action + camera movement + atmosphere" — derived from official examples, no mysticism involved
  • Practical use of the first-last frame workflow: transition shots go from frame A to frame B, reducing the parts that rely purely on luck
  • A retry discipline of "60-second diagnosis + at most 3 attempts per shot": spend the budget where it matters, and log costs back into 1.4 the Cost Log

Ready? Start Lesson One: Your First Shot →

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.