Phr00t

/

WAN2.2-14B-Rapid-AllInOne

like 563

It is an interesting "problem" I've experimented with. Basically, v10 is less tied to the original image in I2V, which gives you more creative flexibility in how the video progresses. It also means the "default" behavior will be to emphasize the main thing in your prompt, which is typically a zoom. v9 and before fed a bit too much of the original image into the denoising process (something I think Lightx2v I2V does), which made results more static (and avoided zooms among other things).

With all that said, try adding stuff like "wide shot" or "smartphone camera recording from [location where you want the camera to stay]", which should discourage zooming.

you can try this:

"Fixed cam capturing a..."

I use this in V9 and it helps avoid zooming.

thx guys, i just tried. it's a bit better, 1 out of 6 will not zoom

Hello and thanks again for your excellent work!
Just a thing, when I use t2v example workflow I got this at the first generation:

"model_type FLOW
unet missing: ['text_embedding.0.scale_input', 'text_embedding.2.scale_input', 'time_embedding.0.scale_input', 'time_embedding.2.scale_input', 'time_projection.1.scale_input', 'blocks.0.self_attn.q.scale_input', 'blocks.0.self_attn.k.scale_input', 'blocks.0.self_attn.v.scale_input', 'blocks.0.self_attn.o.scale_input', 'blocks.0.cross_attn.q.scale_input', 'blocks.0.cross_attn.k.scale_input', 'blocks.0.cross_attn.v.scale_input', 'blocks.0.cross_attn.o.scale_input', 'blocks.0.ffn.0.scale_input', 'blocks.0.ffn.2.scale_input', 'blocks.1.self_attn.q.scale_input', 'blocks.1.self_attn.k.scale_input', 'blocks.1.self_attn.v.scale_input', 'blocks.1.self_attn.o.scale_input', 'blocks.1.cross_attn.q.scale_input', 'blocks.1.cross_attn.k.scale_input', 'blocks.1.cross_attn.v.scale_input', 'blocks.1.cross_attn.o.scale_input', 'blocks.1.ffn.0.scale_input', 'blocks.1.ffn.2.scale_input', 'blocks.2.self_attn.q.scale_input', ... etc"

The genration works anyway, just I wish to understand this, thank you!

sometimes i just type in breast or hips on my prompt will greatly help to the zoom-in face issue.