Keep the experimental look while a video model supplies the motion.
Video models such as MiniMax H3 render fluid, coherent motion - and smooth
away the boiling, flickering texture that makes Deforum, AnimateDiff and
Disco Diffusion feel alive. Feed the video-model render and the Feedback
Sampler look pass (same length or not; it is resampled in time):
* detail transfer - H3's motion, the look pass's high-frequency texture
* colour + detail - also pulls each frame's palette to the look pass
* flicker cuts - every Nth frame (seeded jitter) swaps to the look pass
* crossfade - plain mix
`step_fps` then holds frames to 8-12 fps for a stop-motion / AnimateDiff
stutter, independently of the blend (works without a look pass too).
By chillithebillis·Created 3 months ago·Updated about 7 hours ago· 8
Difforum · Look Mix
video
look_pass
frames
◄blenddetail transfer►
◄amount0.60►
◄detail_radius3►
◄flicker_every3►
◄step_fps0►
◄fps24.00►
◄seed0►
CategoryDifforum/7 · Post
Inputs (9)
Name
Type
Default
Description
video
IMAGE
The video-model render (H3, LTX...). Its motion and light are kept.
blend
COMBO
detail transfer
detail transfer: the look pass's fine texture on top. colour + detail: also its palette. flicker cuts: cut to the look pass every few frames. crossfade: plain mix.
amount
FLOAT
0.600–2
How much of the look pass comes through.
detail_radius
INT
31–16
Texture scale taken from the look pass.
flicker_every
INT
31–48
flicker cuts: roughly one cut every N frames.
step_fps
FLOAT
00–60
0 = off. 8-12 gives a hand-made stutter.
fps
FLOAT
24.001–120
Clip fps, for step_fps.
seed
INT
00–4294967295
flicker cuts: which frames flicker.
look_passopt
IMAGE
Frames in the look (Feedback Sampler or Restyle). Resampled to the video's length.