1. Workflow Overview

A low-VRAM animation generator for architectural scenes featuring:

• Long Sequences: 60s generation with 6GB VRAM via FramePack

• Memory Optimization: Tiled decoding + temporal slicing

• Multimodal Control: CLIP vision + text prompts for motion

• Prompt Assistant: Built-in template in Note node

2. Core Models

3. Key Nodes

4. Pipeline Stages

Stage 1: Input Processing

• Image Input: Load via LoadImage (e.g., work-04.jpg)

• Resolution Matching: Auto-optimize with FramePackFindNearestBucket

• Feature Extraction: CLIPVisionEncode encodes visual cues

Stage 2: Animation Core

• Sampling:

  • 30 steps, CFG=10, UniPC-BH1 sampler

  • VRAM optimizations: teacache (0.15) + temporal_size=64

• Motion Control: Text prompts (e.g., "slow zoom-in")

Stage 3: Output

• Tiled Decoding: 128x128 blocks via VAEDecodeTiled

• Video Export: MP4 output (30FPS, H.264)

5. Inputs & Outputs

• Required Inputs:

  • Static architecture image

  • Motion prompts (Chinese preferred)

• Outputs:

  • MP4 video (FramePack_00001.mp4)

  • Resolution: 512x512 to 1024x1024

6. Critical Notes

1. VRAM Management:

   • Reduce total_second_length (1GB/10s)

   • Enable gpu_memory_preservation (default=6)

2. Dependencies:

   • Download FramePackI2V_HY and hunyuan_video_vae_bf16

   • CLIP models auto-download

3. Troubleshooting:

   • CUDA OOM → Lower latent_window_size (default=9)

   • Choppy video → Ensure temporal_overlap ≥8