Framepack with Video Input (Video Extension) + End Frame

Input Video

End Frame (Optional) - Reduce context frames if very different from input video or if it is jumpcutting/slowing to still image.

Prompt

Faster speed, but often makes hands and fingers slightly worse.

Use TeaCache

Might run out of VRAM (720p requires > 24GB VRAM).

Force Original Video Resolution (No Resizing)

Seed

Batch Size (Number of Videos)

Generate multiple videos each with a different seed.

1 1000

Additional Video Length to Generate (Seconds)

1 120

Distilled CFG Scale

Prompt adherence at the cost of less details from the input video, but to a lesser extent than Context Frames.

1 32

CFG Scale

Use instead of Distilled for more detail/control + Negative Prompt (make sure Distilled=1). Doubles render time.

1 32

Negative Prompt

Requires using normal CFG (undistilled) instead of Distilled (set Distilled=1 and CFG > 1).

Steps

Expensive. Increase for more quality, especially if using high non-distilled CFG.

1 100

Number of Context Frames (Adherence to Video)

Expensive. Retain more video details. Reduce if memory issues or motion too restricted (jumpcut, ignoring prompt, still).

2 10

VAE Batch Size for Input Video

Expensive. Increase for better quality frames during fast motion. Reduce if running out of memory

4 256

Latent Window Size

Expensive. Generate more frames at a time (larger chunks). Less degradation but higher VRAM cost.

9 49

GPU Inference Preserved Memory (GB) (larger means slower)

Set this number to a larger value if you encounter OOM. Larger value causes slower speed.

6 128

MP4 Compression

Lower means better quality. 0 is uncompressed. Change to 16 if you get black outputs.

0 100

Finished Frames

Share your results and find ideas at the FramePack Twitter (X) thread