Wan2.2 I2V V2V Video Extend Quant 14B

Cascade Video Workflow — User Guide

This workflow is designed to generate long videos in chained segments using a Cascade → Cascade approach.

Each segment continues seamlessly from the last frame of the previous one.

You never regenerate from scratch — you extend the motion.

1. Core Concept (Read Once)

The workflow generates video in chunks.

Each chunk:

Starts from an image.

Produces a short video segment.

The last frame of one segment becomes the start image of the next segment.

This process can be repeated indefinitely.

This is called Cascade.

2. Required Models (Load Once)

Before running anything, make sure these are loaded:

High UNet model (structure / motion)

Low UNet model (detail / stability)

CLIP model compatible with WAN

VAE model compatible with WAN

Optional:

LoRA models (LightX, AccVid, or user-chosen LoRAs)

These models remain the same across all cascades unless you intentionally change them.

3. Step A — Initial Image-to-Video (First Segment)

Purpose

Create the first video segment from a starting image.

Steps

Load a starting image (this defines character, framing, style).

Enter Positive Prompt (Start).

Enter Negative Prompt (Start).

WanImageToVideo generates:

latent video seed

conditioning

Run High KSampler (motion / structure).

(Optional) Run RAM / VRAM Cleanup.

Run Low KSampler (detail refinement).

Decode frames using VAEDecodeTiled.

Save the video segment.

Output

A short video segment.

A batch of frames.

4. Extract the Last Frame (Critical Step)

Purpose

Create the entry point for the next Cascade.

Steps

Connect the decoded frame batch to ImageFromBatch.

Set:

batch_index = 999

length = 1

Output is the last frame of the video.

This image becomes the start image for the next cascade.

5. Step B — Cascade Segment (Extend the Video)

Purpose

Extend motion naturally from the previous segment.

Steps

Connect the last frame into a new WanImageToVideo node as start_image.

Enter Positive Prompt (Extend)

(describe continuation, not a new scene).

Enter Negative Prompt (Extend).

Run High KSampler.

(Optional) Run RAM / VRAM Cleanup.

Run Low KSampler.

Decode frames with VAEDecodeTiled.

Output

A second video segment that continues motion seamlessly.

6. Cascade → Cascade (Repeat for Long Videos)

To create longer videos:

Take the last frame of the current cascade.

Feed it into another WanImageToVideo.

Repeat the Cascade steps.

There is no hard limit to how many cascades you can chain.

7. Combining Video Segments

To produce a final continuous video:

Collect frame batches from:

Initial segment

Each Cascade segment

Merge them using ImageBatch.

Export using VHS_VideoCombine.

Result:

One continuous video file.

8. Prompt Writing Rules (Very Important)

Do

Describe continuation, not reset.

Keep:

same character

same camera

same environment

Use gradual changes.

Do NOT

Radically change scene, camera, or identity.

Increase denoise aggressively.

Change resolution or aspect ratio mid-cascade.

9. Common Problems & Fixes

Motion resets

Denoise too high

Prompt too different from previous

Identity drift

CFG too high

Low sampler too strong

Missing cleanup between High and Low

Visible seam between segments

Last frame not correctly extracted

Different frame count or FPS between segments

10. Best Practice Summary

One story, many cascades

Last frame always feeds next start

High = motion, Low = refinement

Prompts evolve slowly

Cleanup when extending long chains

Final Note

This workflow is built for:

long-form motion

stable identity

memory-efficient generation

modular extension