rompecabezas — aspero

the study is the micro-edit. the rest are variations, found by playing with the words.

N takes of one shot. same anchors, same prompt, different seed.
same length, aligned frame for frame. one is the base.

at chosen positions, the base frame is replaced by the SAME frame
index from another take. nothing stretched or blended. 34 for 34.

  pattern   decides WHERE          run mode decides HOW LONG

  fib              .xxx.x..x....x.......x............x......
  phase:3:8:0.12   x.................x.x...............x...
  euclid:3:8       x.................x.................x...

  single   1.................2.3...............4...1......
  grow     1.................22333.............444411111..

fib = an arc. euclid, phase = a pulse.
divergence = how much the takes disagree. drives all of it.

aspero    cut between the takes
velvet    average them
quantum   let them decide when to collapse
melt      re-diffuse until nothing resets

phase:3:8:0.12   two tresillos drifting. bjorklund. clave.
divergence       how much the takes disagree. drives everything.

GPT writes brief + frames  ->  pipeline.py  ->  takes  ->  micro-edit
D:\Dropbox\paul\LTX 2.5 FLF Prompts\
D:\Dropbox\paul\LTX Visual Development\   (highest -vN wins)

  python pipeline.py --shots S02D --n 8

velvet 5

average the takes. soft where they disagree

velvet.py --mode mean|median|ramp|breathe

mean · median · ramp · breathe = collapse on a clave clock

velvet_S02C_breathe_euclid

velvet_S02B_breathe_phase

velvet_S02D_median_8takes

velvet_S02B_ramp_13takes

velvet_S02B_mean_13takes

melt 9

re-diffuse each frame from the last. the image never resets

melt.py --strength 0.20 --feedback 0.58 --detail 0.85

peak/med 1.2 · spikes 0.0 · his 1.2-2.2 / 0-2.5 · sharp 46 -> 240

MELT_quantum_inverted

SHARP_detail85

SHARP_1280_detail55

MELT_promptsched

MELT_light_s0.25

MELT_light_s0.18

MELT_light_s0.12

MELT_sdxlturbo_lcmlora

S02B-base05_grow_s1_melt_s0.30_fb0.70

quantum 7

the takes time their own collapse

velvet.py --mode quantum [--invert]

S02B 69/145 · S02A 91/145 · inverted 1.4% spikes · nothing set by hand

QINV_S02E1

QINV_S02D

QINV_S02C

QINV_S02A

QUANTUM_S02A_boardwalk

QUANTUM_S02B_inverted

QUANTUM_S02B

agua 16

same swell, smooth and jagged. the two operations on one subject

velvet.py --pool S07E*_0000*.mp4 --mode M · fib_cut.py --run grow

3 families × 8 takes · mean median breathe quantum trail + aspero · two takes of one wave correlate at 0.07

water_S07E2_trail

water_S07E2_mean

water_S07E3_aspero

water_S07E3_quantum

water_S07E3_breathe

water_S07E3_median

water_S07E3_mean

water_S07E2_aspero

water_S07E2_quantum

water_S07E2_breathe

10 of 16 · water

chunkforum 12

the cut hides the zoom reset. travel from material that never travelled

chunkforum.py --rotate 1.8 --sat 3.6 --reset hold

curl 0.115 = his 0.115 · sat 143 = his 143 · no GPU

VX_slowspin

VX_vortex_velvet

VX_rot_sat_hi

CF_miranda_VELVET

CF_boardwalk_VELVET

CF_boardwalk_runaway

CF_faces_fast

CF_miranda_portrait

CF_boardwalk_travel

chunkforum_RUNAWAY_hold_s4

10 of 12 · chunkforum

deforum 1

FIRST=frame LAST=warp(frame). the model re-renders every link

deforum.py --links 10 --zoom 1.10

invents structures · 977 -> 36 sharpness by link 10

boardwalk_10links_z1.10

lora 8

2.3 loras load on 2.5. they act at generation, before any edit

lora_test.py --lora X.safetensors --trigger zhuanchang

transition = cut + wipe, 15% spikes · restyle = sat x2, 0 spikes

tr_neutral_ltx2.3-transition_s1.00

tr_sand_ltx2.3-transition_s1.00

tr_shatter_ltx2.3-transition_s1.00

tr_smoke_ltx2.3-transition_s1.00

tr_liquid_ltx2.3-transition_s1.00

restyle_852654_LTX2.3-22B_ReStyle__s1.00

crossview_LTX2.3-22B_IC-LoRA-CrossVi_s1.00

morphA_ltx2.3-transition_s1.00

women 25

3 live + 25 cropped. hard cuts only. order reshuffles per seed

build_women.py --variant crescendo|live --ramp 14:1

28 women · 7-frame inserts · no pattern, no hierarchy

S02E_women_live_s5_bed

S02E_women_live_s21_blk

S02E_women_live_s8_blk

S02E_women_live_s3_blk

S02E_women_crescendo_s21

S02E_women_crescendo_s5_bed

S02E_women_crescendo_s8_blk

S02E_women_crescendo_s3_blk

S02E_women_scattered_s34_bed

S02E_women_ordered_s21_bed

10 of 25 · outS02/sequences

S02 9

A boardwalk · B camera · C keep warm · D scar · E rupture

build_s02.py --women live --prefer grow

26-29s

S02_full_hard_live_s15

S02_full_hard_live_s12

S02_full_hard_live_s11

S02_full_hard_ordered_s14

S02_full_hard_crescendo_s13

S02_full_hard_scattered_s9

S02_full_hard_crescendo_s7

S02_full_hard_ordered_s4

S02_full_hard_crescendo_s1

aspero 128

cut hard between takes of one shot

pipeline.py --shots S02D --n 8

SHOT-baseNN_mode_seed · single|grow · phase:3:8:0.12

S02E8-base00010_grow_s5

S02E8-base00010_single_s5

S02E8-base00010_grow_s3

S02E8-base00010_single_s3

S02E8-base00010_grow_s2

S02E8-base00010_single_s2

S02E8-base00010_grow_s1

S02E8-base00010_single_s1

S02D-base00002_grow_s5_v2

S02D-base00002_single_s5_v2

10 of 128 · outS02

V3Y seq 22

the frame-swap as the cut. freeze, black, silence

seq_glitchcut.py --pattern phase:3:8:0.12 --trans 48

V3Y_repeat_build_silent

V3Y_repeat_early_vo

V3Y_repeat_build_vo

V3Y_repeat_weave_vo

V3Y_glitchcut_phase_silent

V3Y_glitchcut_phase_single_vo

V3Y_glitchcut_phase_vo

V3Y_glitchcut_grow_vo

V3Y_glitchcut_single

V3Y_glitchcut_grow

10 of 22 · outV3Y/sequences

V3Y 50

55 takes, five shots, bases by hand

pipeline.py

Y_S01E-base07_grow_s1_srcALL

Y_S01E-base07_single_s1_srcALL

Y_S01D-base08_grow_s1_srcALL

Y_S01D-base08_single_s1_srcALL

Y_S01C-base10_grow_s1_srcALL

Y_S01C-base10_single_s1_srcALL

Y_S01B-base10_grow_s1_srcALL

Y_S01B-base10_single_s1_srcALL

Y_S01A-base11_grow_s1_srcALL

Y_S01A-base11_single_s1_srcALL

10 of 50 · outV3Y

S01 51

where it started. fibonacci across takes

fib_cut.py

S01E_fib-grow_pick-random_s5_srcfib_bl2

S01E_fib-single_pick-random_s5_srcfib_bl2

S01E_fib-grow_pick-random_s3_srcfib_bl2

S01E_fib-single_pick-random_s3_srcfib_bl2

S01E_fib-grow_pick-random_s2_srcfib_bl2

S01E_fib-single_pick-random_s2_srcfib_bl2

S01E_fib-grow_pick-random_s1_srcfib_bl2

S01E_fib-single_pick-random_s1_srcfib_bl2

S01D-08fade_grow

S01D-08fade_single

10 of 51 · .

archaeological software excavation

raw auto1111 + deforum + parseq. the 2023-24 stack that made the old films.

rtx 5080. torch 2.1.2 has no sm_120. the old install cannot run.

GPT     engine copied to pixelsoup\deforum-app, torch 2.7.1+cu128.
        api up. first job dies on an empty controlnet namespace.
        writes HANDOFF_PIXELSOUP.md. applies the one-line fix 90 s later.
        the retry renders. the handoff still says FAILED.
Claude  reads the logs, not the note. finds the render. looks at it.
        env frozen. start/stop. comfy bridge. 80 parseq scores read.
        controlnet restored. 13 checkpoints. one checkpoint per process.

original install on D: never modified. everything new beside it.

broke on the way   inherited pipe hang . cp1252 vs a check mark .
                   model half on cpu . two loads racing . mediapipe 1.0 .
                   hashed names . two wrong guesses: motion, steadiness

excavation 3

raw auto1111 + deforum on a 5080. dead engine, new torch

pixelsoup\deforum-app\start.ps1 -Ckpt X · Deforum Render (PixelSoup)

torch 2.7.1+cu128 · 1 patched line · 48f in 20s · 15.34 / 1.31 / 0%

01_first_render_rv_2d

02_comfy_bridge_rv

03_parseq_leonsol22_rv_3d

jam 5

his parseq scores, decoded. prompts crossfade on AND weights

Parseq Score (PixelSoup) · parseq_tools.py · prompt_weights.py

80 scores · seed drift 0.0089/f · hard 14.9/1.36 · linear 13.2/1.43 · his 7.1/1.93 · pulse 7.1/1.84

A_hard

B_linear

C_his

D_pulse

compare_2x2

hyper 5

juggernaut x hyper alone in the loop becomes texture

checkpoint Juggernaut X Hyper · DPM++ SDE 6 · CFG 1.7 · leonsol22 score

spots from frame 1 · 2 steps / 6 steps / still camera / CFG pulse / noise 0.15 - none recover

01_leonsol_2steps

02_leonsol_6steps

03_still_camera

04_cfg_pulse

05_noise_015_fixed_seed

feed 15

feed it a video. the take becomes the structure

LTX take -> ControlNet Unit (depth / canny) -> Deforum Render

RV depth 22.8/1.20/0% · Jugg depth 13.4/1.10/0% · depth+canny 7.4/1.20 · looser 0.6+0.3 9.2/1.08 · release@60% 7.7/1.14 · breathing grip 9.6/1.09 · noise pulse 13.5/1.16 · trap: fixed seed 2.9 = frozen · song bass -> strength 12.7/1.64, change follows bass r=0.84 · bass -> strength + grip 12.3/1.73 r=0.86

03_S03A_jugg_depth_canny

04_S04B_jugg_depth

08_S03A_jugg_depth_noise_pulse

09_S03A_jugg_depth06_canny03

10_S03A_jugg_release_at_60pct

11_S03A_jugg_breathing_grip

12_S03A_drawn_strength_zoom_curves

13_S03A_subwayseq_bass_strength

14_S03A_subwayseq_bass_grip

15_S03A_strength_vs_grip_compare

10 of 15 · excavation/feed

mondongo 11

everything in one pot. the take as bones, motion and skin. mueca

Load Video -> ControlNet Unit x2 + Hybrid Video -> Deforum Render

comp 0.3 + RAFT flow · S03A 5.3/1.23 the source wins · S02DM scar 6.2/1.66 kintsugi, then stone · breathing blend 13.0/1.50 another woman · women cuts 11.6% spikes -> 0% (8.3/1.56)