First–last frame video

Official Wan2.1 FLF2V

Wan2.1 added FLF2V on April 17, 2025: inference code plus Wan2.1-FLF2V-14B-720P weights. The generator class WanFLF2V takes first_frame, last_frame, a prompt, and defaults such as frame_num = 81, max_area = 720×1280, UniPC sampler, 50 steps, guidance 5.5. If the two images mismatch, the last frame is cropped and resized to the first.

The README notes the model was trained primarily on Chinese text–video pairs and recommends Chinese prompts for best results. ComfyUI documents a native Wan FLF2V example (WanFirstLastFrameToVideo): encode both stills, build a latent canvas, mask the booked frames, denoise the middle.

What Wan2.2 does not add

• No Wan2.2-FLF2V row in the 2.2 model table.
• I2V-A14B / TI2V-5B --image condition on one still.
• Later closed Wan APIs (2.5+) reintroduce first/last-frame as a product control — that is not the 2.2 open ckpt.

Overlap with VACE

VACE can also interpolate or expand using frame references inside a Video Condition Unit. That is a different checkpoint (Wan2.1-VACE-1.3B / 14B) with extra mask/reference plumbing. If you only need two keyframes and a prompt, FLF2V is the narrower tool.

FAQ

Can I Fine-tune 2.2 T2V to FLF2V?

Community workflows exist, but they are not an official 2.2 release. Start from Wan2.1-FLF2V weights if you need the documented behavior.

Why 81 frames?

Wan latents use a temporal compression of 4. Frame counts of the form 4n+1 line up with the VAE grid (1 + 4×20 = 81 is the common 720P default).