1. Workflow Overview

This workflow leverages WAN2.1 video generation model for timelapse/aerial nightscape scenarios, featuring:

• Converts static images to dynamic videos (e.g., city light transitions).

• 4x upscaling + RIFE interpolation (32FPS output).

• Custom LoRA (Nightscape Timelapse LoRA) for enhanced lighting.

• VRAM-optimized pipeline (runs on mid-tier GPUs).

2. Core Models

• WAN2.1 Base (Wan2_1-I2V-14B-480P_fp8): 480P image-to-video with FP8 precision.

• Text Encoders:

  • umt5-xxl-enc-bf16: Multilingual text understanding.

  • open-clip-xlm-roberta-large-vit-huge-14: Image semantics.

• Upscaler (4xRealWebPhoto_v4.pth): 4x resolution enhancement.

• Frame Interpolation (rife47.pth): Smart 32FPS conversion.

3. Node Details

Key Nodes

• WanVideoSampler: Core video generation (30-step DPM++ with SLG optical flow).

• CR Upscale Image: 4x super-resolution (lanczos anti-aliasing).

• RIFE VFI: Frame interpolation (strength=10).

Installation

• WAN2.1 Plugin: Manual install from GitHub.

• Video Suite: Install VideoHelperSuite via ComfyUI Manager.

Dependencies

• Model Files: Download from Aliyun OSS.

• CUDA 11.7+: Required for FP8 acceleration.

4. Workflow Structure

1. Input Group: LoadImage, WanVideoTextEncode.

2. Model Loading: WanVideoModelLoader, WanVideoVAELoader.

3. Generation Core: WanVideoSampler, WanVideoSLG (optical flow).

4. Post-Processing: Upscale + Interpolation + Video merge.

5. Inputs & Outputs

• Inputs: 480x640 image, prompts (e.g., "smooth camera zoom").

• Output: MP4 video (H.264, 32FPS) as Wan*.mp4.

6. Notes

1. Errors:

   • Missing FP8 support breaks model loading.

2. Optimization:

   • Enable offload_device for lower VRAM usage.

3. Compatibility: Requires Linux + ComfyUI ≥0.3.15.