Building a Video Production Workflow

Bringing It All Together

Each lesson in this track covered one technique in isolation. A production video workflow combines them into a single pipeline that takes a text prompt and a starting image and outputs a polished, upscaled MP4 without manual intervention between stages. This lesson builds that pipeline using Wan 2.2 as the core video model, with a face detail pass, 4x upscaling, and proper MP4 export.

The workflow is organized into labeled groups so it is readable and maintainable. Once built, you reuse it by swapping prompts and images — not by rebuilding from scratch.

Planning the Pipeline

The production video workflow has seven stages:

1. Load: video model, VAE, CLIP, reference image (optional)
2. Condition: text prompts (motion + scene), optional I2V image conditioning
3. Generate: WanVideo Sampler at base resolution (480p or 720p)
4. Decode: WanVideo Decode → raw IMAGE batch
5. Restore faces (optional): FaceDetailer if the clip contains faces
6. Upscale: 4x-UltraSharp frame upscale
7. Export: RIFE interpolation (optional) → Video Combine → MP4

Assembling the Node Graph

Stage 1 — Loaders (purple group):

• WanVideo Loader: select your Wan 2.2 14B T2V model; VAE linked automatically
• Load CLIP (Wan): select the Wan text encoder
• Load Upscale Model: 4x-UltraSharp
• If using I2V: Load Image for the starting frame

Stage 2 — Conditioning (orange group):

• CLIP Text Encode (Wan): positive prompt — full scene and motion description
• CLIP Text Encode (Wan): negative prompt — optional, short (blurry, low quality)
• If using I2V: CLIPVisionEncode connecting Load Image output

Stage 3 — Base video generation (teal group):

• EmptySD3LatentImage or WanVideoEmptyLatent: set width, height, num_frames (81 for ~5 seconds at 16 FPS)
• WanVideo Sampler: connect model, text conditioning, latent; steps 30, cfg 6.0, shift 5.0

Stage 4 — Decode and preview (blue group):

• WanVideo Decode: connect sampler output and VAE
• Preview Image or Preview Animated: shows the raw generation before post-processing; lets you check quality before committing to the slow upscale pass

Stage 5 — Face restoration (red group, optional):

• FaceDetailer (Impact Pack): if the video contains faces, add FaceDetailer connected to the IMAGE batch. Use the same model/clip/vae connections as stage 3. Set guide_size 512, denoise 0.4.
• Note: FaceDetailer processes the full IMAGE batch frame by frame, so expect this to take 30–60 seconds for a 5-second clip

Stage 6 — Upscale (green group):

• Upscale Image With Model: connects IMAGE batch from decode (or FaceDetailer) and the 4x-UltraSharp model

Stage 7 — Export (gray group):

• Optional RIFE interpolation: RIFE node between upscale and Video Combine; multiplier 2 to double FPS
• Video Combine: frame_rate 24 (or 48 if RIFE was used), format video/h264-mp4, crf 18–23

Organizing and Saving

Select all nodes in each stage and create a group (Ctrl+G or right-click → Group Nodes). Name each group with its stage purpose. Color them consistently: purple for loaders, orange for conditioning, teal for sampling, blue for decoding, red for restoration, green for upscaling, gray for export.

Add Note nodes near any node with non-obvious settings — document which model file is loaded, what the motion prompt is targeting, and what RIFE multiplier you used. These notes make the workflow interpretable months later.

Export as JSON: Ctrl+S → save as "wan22-production-i2v.json." Keep a folder of named production templates. A well-organized library of templates (t2v-basic, i2v-portrait, v2v-style-transfer, long-form-stitched) lets you start any video project from a working baseline rather than from scratch.

Testing the Pipeline

Before running the full pipeline on a real project, run a short test:

• Set num_frames to 25 (about 1.5 seconds at 16 FPS)
• Run through all stages to confirm each node passes its output correctly
• Verify the MP4 saves to the output/ folder and plays correctly
• Check that FaceDetailer (if included) completes without error on the test frames

Once the test passes cleanly, increase num_frames to your target clip length and run the full generation. The per-frame upscale and face restoration stages are the slow parts — plan for 10–30 minutes for a 5-second fully post-processed clip depending on your GPU.