topabaem commited on
Commit
09ebdf8
·
verified ·
1 Parent(s): f4fcb0f

Upload samples/README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. samples/README.md +37 -0
samples/README.md ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Samples
2
+
3
+ Five clips generated with `A3.packed.safetensors` driving
4
+ `LTX25-distilled-DiT-Q3_K_M.gguf` — 768×512, 24 fps, video and audio muxed,
5
+ 2.0–4.0 s each. Encoder loaded through the resident path (7.75 GiB), eight
6
+ distilled steps, CFG 1.0, no LoRA, no latent upscale. Seeds 20260813–20260817,
7
+ one seed per prompt.
8
+
9
+ The prompts are deliberately spread across the axes a text encoder has to keep
10
+ apart — camera instruction, shot scale, subject count, lighting, and a distinct
11
+ audio scene each — so that compression blurring any of them shows on screen.
12
+
13
+ | file | prompt |
14
+ |---|---|
15
+ | `sample-00.mp4` | A humanoid robot steps out of a freight elevator into a rain-slick loading bay, sodium lights flaring off its wet white shell. Locked-off low camera as the robot turns toward the lens. Audio: heavy rain on metal, one servo whine, distant thunder. |
16
+ | `sample-01.mp4` | Extreme close-up of a robotic hand closing around a cracked ceramic cup, actuator tendons flexing under translucent silicone skin. Warm workshop light, macro lens. Audio: fine servo clicks, ceramic scraping steel. |
17
+ | `sample-02.mp4` | A four-legged inspection robot trots along a narrow steel catwalk inside a power plant, camera tracking alongside, steam venting behind. Audio: metallic footfalls, hissing steam, turbine hum. |
18
+ | `sample-03.mp4` | Overhead wide shot of a warehouse floor where dozens of small wheeled robots weave between one another carrying yellow crates. Slow push-in, cold fluorescent light. Audio: rolling wheels, warning beeps. |
19
+ | `sample-04.mp4` | A battered rescue robot pushes through thick smoke in a collapsed corridor, headlamp cutting through dust. Handheld camera, orange emergency light. Audio: crackling fire, muffled alarm, grinding treads. |
20
+
21
+ ## What these do and do not show
22
+
23
+ `sample-00` is the one to watch: it renders an actual robot — dome head, two
24
+ amber eyes, segmented neck — and the BF16 encoder renders the same thing on the
25
+ same card and seed. The pre-2026-08-17 encoder produced a human-faced android
26
+ for this prompt, so that improvement belongs to Lightricks' checkpoint
27
+ alignment, not to this compression.
28
+
29
+ `sample-01`, `sample-02` and `sample-03` do **not** follow their prompts. There
30
+ is no robotic hand, no four-legged robot, no swarm of wheeled robots — the model
31
+ returns people instead. **The BF16 encoder fails these prompts identically**,
32
+ across seeds, phrasings and CFG up to 6.0. It is a limit of LTX-2.5 distilled,
33
+ not compression damage, and it is left in rather than cherry-picked away.
34
+
35
+ Frame-by-frame comparisons against BF16 and against the legacy-guard build, all
36
+ from the same card and seeds, are in `evidence/frames/cmp-*.png`, with the full
37
+ clip sets under `evidence/clips-{BF16,A0,A3}/`.