v1.1 much improved vs 0.2 on the use case I was playing with

#18
by tidalvibes - opened

v0.2 felt very heavy handed with the 'sharpening' on a video I was testing with that was quite dark overall - v.1.1 produced much more usable results.
My vid source is very dark, so it might be the case where this lora works well where there's already ton's of light, so I might be abusing it.
I came across this looking for a way to uncrap [wan 2.1 1.3b] renderings where it looks like a gorillion spheres/circles make up the objects due to camera motion.

v0.2 I could get rid of them using this one, but at the expense of making something completely new
v1.1

Vid:

Dark hallway lined with mirrors (appeared to have wet rough stone frames), a smooth rectangular pool on the floor lined with smooth wet stones, a purple-ringed opening at the end of the hall with a purple moon seen outdoors, and a green moon-ish orb above that - all under a arched ceiling with ambient green haze.

v0.2

the lora would go the the extremes on the edges, being mostly white, wet reflections on the stones also leaned toward white - and the still pool developed waxy highlights.

v1.1

Seems to be much more subtle, adding more overall ambient light but the 'sharpening' much more tame & allowing more color to come through.

Blur

  • things sucked major until I figured out I should unmute the blur in the workflow :) With that enabled, things got progressively better for me.
  • Radius 4/sigma 1.0 : toyed around with radius and the lora strength looking for a happy balance

Lora Strength

  • Added auto-increment to the lora strength to test
  • 1 created some really cool effects that were completely non-representative of the original source vid. Acts almost like a denoiser in this respect.
  • 0.5 ~ 0.8 seemed to be a happy place for the two similarly-styled vids I was working with.

Other

  • got some very cool 'retro movie' effects in some cases - transforming life-like video into a surreal old-school film or dreamy disassociated feel.
  • still mucking with things to meet my original need, but will be revisiting this lora for more improper use/abuse simply for the interesting effects it makes.

Thanks for sharing!

v0.2

lora strength 1

wan-v2v-2025-0812-232030-compare_00001.png

others

wan-v2v-2025-0812-222333-compare_00001.png
wan-v2v-2025-0812-220831-compare_00001.png
wan-v2v-2025-0812-231603-compare_00001.png
wan-v2v-2025-0812-232542-compare_00001.png
wan-v2v-2025-0812-233039-compare_00001.png

v1.1

wan-v2v-2025-0812-234854-compare_00001.png

wan-v2v-2025-0812-235250-compare_00001.png

wan-v2v-2025-0812-234410-compare_00001.png

Yes the blur is not optional, it's required. The preprocess for all tile control models is technically downscale->upscale->blur, although you can skip the down/up. It's the same with the older sdxl and sd1.5 tile controlnets.

The "correct" amount of blur to match my training settings is radius=12 or 13, sigma=0.3, but you can use different values to affect sharpness/local contrast.

Sign up or log in to comment