Replace the samples with a BF16 comparison set

.gitattributes

@@ -72,3 +72,33 @@ samples/sample-02.mp4 filter=lfs diff=lfs merge=lfs -text
 samples/sample-03.mp4 filter=lfs diff=lfs merge=lfs -text
 samples/sample-04.mp4 filter=lfs diff=lfs merge=lfs -text
 encoder-hf/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+samples/4bit-00.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/4bit-01.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/4bit-02.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/4bit-03.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/4bit-04.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/4bit-det-00.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/4bit-det-01.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/4bit-det-02.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/4bit-det-03.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/4bit-det-04.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/bf16-00.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/bf16-01.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/bf16-02.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/bf16-03.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/bf16-04.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/bf16-det-00.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/bf16-det-01.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/bf16-det-02.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/bf16-det-03.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/bf16-det-04.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/compare-00.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/compare-01.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/compare-02.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/compare-03.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/compare-04.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/compare-det-00.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/compare-det-01.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/compare-det-02.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/compare-det-03.mp4 filter=lfs diff=lfs merge=lfs -text
+samples/compare-det-04.mp4 filter=lfs diff=lfs merge=lfs -text

samples/{sample-00.mp4 → 4bit-00.mp4}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0c9523b8260a92eeb6ddd325fdd054ce24f89a0f653341642270927011629e9e
-size 741031
+oid sha256:246d0402e1731a748afdb3101efe52d048b8e8dd65f6d4cb8f65606fbfb15531
+size 658765

samples/{sample-01.mp4 → 4bit-01.mp4}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9cd49e47396ca2a8f20560a4fa9b84798887f8446775d314025a61f9a3e62788
-size 417342
+oid sha256:e120009a490789533b43864c667f2216564af5844c60034b600a021d2b5c876b
+size 736641

samples/{sample-02.mp4 → 4bit-02.mp4}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7830ebe37279864abb0e64e5bd033236a3ee83626b492042373169b4c0d1999f
-size 814823
+oid sha256:95f5f59400a4fec548d0c10b436d656ff7eab16ae260ef149df10e23b4b15b63
+size 462652

samples/{sample-03.mp4 → 4bit-03.mp4}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fe19cbe497839d43af7b04ce55f37377b079ba11372632a2337869079f78032e
-size 794800
+oid sha256:d7c89db141a9102f1d43c494a5e4f1bfc4ae6498792452b728ca5c43fb92eb96
+size 453177

samples/4bit-04.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74a7cd6a40ee72b5fee66cd986ec707591bde7c8546911cae406cae3fb316f09
+size 359739

samples/4bit-det-00.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a453393cb07c4fd098846c0bbfff893125713bdc1a56526f35887aff25e367e
+size 614764

samples/4bit-det-01.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a83ac652e05430e6a3ccdc934e62af306876256eb1a50f23e1f5260c1bfcfae0
+size 652452

samples/4bit-det-02.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e6c63077ee2a351e822e229b7d926f3451a3df0cf36e58ed65a4237d298889a
+size 595254

samples/4bit-det-03.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80c0b28292aba9d1dc9b2a7005fbd0a3af157c4a3a24b477984a39339a6af032
+size 587334

samples/4bit-det-04.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6064c5474a0c9a822370088d884f34cef8955c0bff948933ec58698811e3a142
+size 346015

samples/README.md

@@ -1,37 +1,131 @@
-# Samples
-
-Five clips generated with `A3.packed.safetensors` driving
-`LTX25-distilled-DiT-Q3_K_M.gguf` — 768×512, 24 fps, video and audio muxed,
-2.0–4.0 s each. Encoder loaded through the resident path (7.75 GiB), eight
-distilled steps, CFG 1.0, no LoRA, no latent upscale. Seeds 20260813–20260817,
-one seed per prompt.
-
-The prompts are deliberately spread across the axes a text encoder has to keep
-apart — camera instruction, shot scale, subject count, lighting, and a distinct
-audio scene each — so that compression blurring any of them shows on screen.
-
-| file | prompt |
-|---|---|
-| `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. |
-| `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. |
-| `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. |
-| `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. |
-| `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. |
-
-## What these do and do not show
-
-`sample-00` is the one to watch: it renders an actual robot — dome head, two
-amber eyes, segmented neck — and the BF16 encoder renders the same thing on the
-same card and seed. The pre-2026-08-17 encoder produced a human-faced android
-for this prompt, so that improvement belongs to Lightricks' checkpoint
-alignment, not to this compression.
-
-`sample-01`, `sample-02` and `sample-03` do **not** follow their prompts. There
-is no robotic hand, no four-legged robot, no swarm of wheeled robots — the model
-returns people instead. **The BF16 encoder fails these prompts identically**,
-across seeds, phrasings and CFG up to 6.0. It is a limit of LTX-2.5 distilled,
-not compression damage, and it is left in rather than cherry-picked away.
-
-Frame-by-frame comparisons against BF16 and against the legacy-guard build, all
-from the same card and seeds, are in `evidence/frames/cmp-*.png`, with the full
-clip sets under `evidence/clips-{BF16,A0,A3}/`.
+# Samples — 4-bit against the BF16 original
+
+Every clip here exists twice: once from the **BF16 original (26.264 GB)** and
+once from this **4-bit build (8.46 GB)**. `compare-NN.mp4` puts the pair side by
+side, BF16 on the left.
+
+Everything downstream of the text encoder is held identical — same DiT
+(`LTX25-distilled-DiT-Q3_K_M`), same seed, same 8-step distilled schedule, same
+VAE tiling, one process, one GPU. The encoder is the only thing that differs.
+
+768x512, 24 fps, 2.0 s, video and audio generated together, rendered on a
+**Tesla V100-SXM2-16GB** — the card this compression exists to fit.
+
+## Two sets, because they answer different questions
+
+`compare-NN.mp4` uses **`euler_ancestral`**, the vendor default and what you
+actually get. It injects fresh noise at every step, so a conditioning difference
+of cosine 0.998 does not stay small — it steers the trajectory elsewhere and the
+two builds return different **takes of the same scene**. Look at these to see
+what the model produces. They cannot tell you what the compression cost, and
+reading them that way will mislead you: in `compare-01` the BF16 side has a
+second figure on the dune that the 4-bit side does not, and that is the sampler
+re-rolling, not the encoder forgetting.
+
+`compare-det-NN.mp4` uses **`euler`**, which injects nothing. Same seed means
+the same starting noise, so the only thing left that can move a pixel is the
+conditioning. **This is the set that attributes a difference to the encoder.**
+
+It is worth saying what this did *not* do. Removing the ancestral noise did not
+reliably bring the pairs closer — measured below, it helped enormously on one
+clip and made three marginally worse. Fixing the starting noise does not fix
+the path: eight distilled steps through a 22B transformer are nonlinear enough
+that a cosine-0.998 conditioning difference can still land somewhere else. So
+the deterministic set is the honest attribution, not a flattering one.
+
+## Prompts
+
+Written to pull on what a text encoder has to keep apart: camera instruction,
+shot scale, lighting direction, surface material, and a distinct audio scene.
+
+| # | seed | prompt |
+|---|---:|---|
+| 00 | 20260813 | 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. |
+| 01 | 20260814 | A lone astronaut in a scuffed white suit walks across a red dune at sunset, a long shadow stretching behind, dust curling off each bootfall. Low tracking camera moving with the walk. Audio: wind over sand, suit fan hum, muffled breathing. |
+| 02 | 20260815 | A blacksmith hammers glowing orange steel on an anvil inside a dark forge, sparks bursting outward with every strike, warm rim light on soot-streaked arms. Medium shot, camera static. Audio: ringing hammer strikes, roaring bellows, a hiss as the blade is quenched. |
+| 03 | 20260816 | A vintage sports car slides through a wet mountain hairpin at night, headlights sweeping across the guardrail, spray lit red by the tail lights. Camera pans to follow the car through the turn. Audio: engine roar, tyre squeal, rain on asphalt. |
+| 04 | 20260817 | A battered rescue robot pushes through thick smoke in a collapsed corridor, headlamp cutting a cone through the dust, orange emergency light pulsing on the walls. Handheld camera following from behind. Audio: crackling fire, a muffled alarm, grinding treads. |
+
+## Drift, measured before any of it reaches the DiT
+
+4-bit conditioning against BF16 conditioning, same prompts, same tokenizer path:
+
+| # | video relL2 | audio relL2 | video cosine | audio cosine |
+|---|---:|---:|---:|---:|
+| 00 | 0.05820 | 0.05010 | 0.99831 | 0.99875 |
+| 01 | 0.05614 | 0.04925 | 0.99843 | 0.99879 |
+| 02 | 0.07002 | 0.06274 | 0.99755 | 0.99803 |
+| 03 | 0.05959 | 0.05222 | 0.99823 | 0.99864 |
+| 04 | 0.05203 | 0.04624 | 0.99865 | 0.99893 |
+| **mean** | **0.05919** | **0.05211** | | |
+
+Prompt 02, the forge, is hardest on both branches. It also has the most
+simultaneous physical events — sparks, glow, bellows, quench — which is where a
+text encoder has the most to lose.
+
+## Difference on screen
+
+Mean absolute error per channel over every decoded frame, 4-bit against BF16.
+Not relative L2: these scenes run from a night forge to a sunlit dune, and
+relative error divides by the frame's own norm, so a dark clip would score worse
+than a bright one for the same visible difference. MAE has its own bias — it
+flatters dark frames — so **only the ratio compares across rows**, because
+within one scene the brightness cancels.
+
+| # | scene | `euler_ancestral` | `euler` | ratio |
+|---|---|---:|---:|---:|
+| 00 | robot, rain | 0.0172 | 0.0182 | 0.94x |
+| 01 | dune | 0.0596 | 0.0580 | 1.03x |
+| 02 | forge | 0.0098 | 0.0129 | 0.76x |
+| 03 | night road | 0.0702 | 0.0252 | **2.79x** |
+| 04 | smoke | 0.0128 | 0.0165 | 0.77x |
+| | **mean** | **0.0339** | **0.0261** | 1.26x |
+
+## The one that actually diverged
+
+Under deterministic sampling, four of the five pairs hold together: same
+composition, same lighting, same timing, differing in surface detail. Clip 00
+gives the 4-bit side slightly more panel detail on the hull; clips 02, 03 and 04
+are close enough that finding the difference takes looking.
+
+**Clip 01 is not.** Same terrain, same sun, same shadow, same walk — and BF16
+renders a soldier in helmet and fatigues while the 4-bit renders a man in a
+business suit. The sampler is deterministic and the seed is identical, so that
+belongs to the encoder.
+
+It deserves one qualification rather than being either buried or oversold.
+Neither build followed the prompt: it asked for an astronaut in a scuffed white
+suit and got neither. The model has no confident answer here, so it lands on an
+arbitrary human, and in that regime a small conditioning change flips which one.
+The difference is real and attributable; its size reflects a prompt that was
+already unstable on both sides.
+
+## What these prove, and what they do not
+
+They show that at 8.46 GB the encoder drives the same DiT to a comparable result
+on five prompts, at one resolution, one length and one seed each.
+
+They are **not** a claim that the 4-bit encoder is indistinguishable from BF16.
+No blind study, no listening test, no sweep over seeds or resolutions.
+
+**Prompt adherence is partial, and mostly on both builds equally.** Clip 00 has
+no freight elevator and frames far tighter than "locked-off low camera" implies;
+clip 04 reads closer to a firefighter than a rescue robot; clip 01 produces no
+astronaut on either side. The BF16 original does the same thing on the same
+prompts, so that is a limit of LTX-2.5 distilled rather than compression damage.
+
+The clips are published unedited — including clip 01, where the compression
+does visibly cost something — rather than replaced with prompts that happen to
+flatter the result.
+
+## Files
+
+* `compare-NN.mp4` — BF16 left, 4-bit right, `euler_ancestral`.
+* `compare-det-NN.mp4` — the same pairing under deterministic `euler`.
+* `bf16-NN.mp4`, `4bit-NN.mp4` — the individual `euler_ancestral` renders.
+* `bf16-det-NN.mp4`, `4bit-det-NN.mp4` — the individual `euler` renders, if you
+  want to step through them frame by frame rather than trust the stack.
+
+Audio in the paired clips is taken from the 4-bit side rather than mixed: the
+two tracks are different renders of the same prompt, and mixing them would
+produce a phasing artefact belonging to neither build.

samples/bf16-00.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:098ec89ce20cc538d38bccc38a76410190c8ccc83ac38b243b961dc1d7113466
+size 655599

samples/bf16-01.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87f6df79837c1dbb1f5eca8771a5d543c54091e7f6c6cb22c7c54d4faf7bbc1a
+size 722695

samples/bf16-02.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13ae68bf658b71eaf1fd099ecddc7e9d237dd3e07a4ba191a275509057ac6dc5
+size 439369

samples/bf16-03.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a41fedd7e9caa88785c078e9ca86b01c7ca12d9714f92150e101c8bd0c06db8c
+size 451430

samples/bf16-04.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aaa5f3b2cc5adb6a9ae1ca6d49de5dfb04e32bc0ab9b5df19bd17f6f58c15000
+size 362430

samples/bf16-det-00.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27beaf579b4842bed8e2a1be5872b66730f34ad811d1d2ecd57caa453b177683
+size 626961

samples/bf16-det-01.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7ca148ea743c51ef3cc433334cf37b24c6c076d66d4c004bcdbc53e09012e0c
+size 804162

samples/bf16-det-02.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a74323540aea1f1cc019892c91a18cd33f06cf50918c5c63f3d3df3bcbe2e58e
+size 599848

samples/bf16-det-03.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:806a95ec65fe69b60414914a0fcdf42277f868fe4f34a66a02f27bbf279b64a1
+size 564347

samples/bf16-det-04.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:710042e4dfc30d27afe204233c3c63778d7c6cd4167824a2039f496498520106
+size 342896

samples/compare-00.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e19beebce63e7a6ad77ec6e617bb0015389e8e41e24d7891f2845b73e7a33407
+size 1078844

samples/compare-01.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:819a4602a19fc13eaca625edec94a5f5df4af77dbf770f7ed062b148d355e38f
+size 1172206

samples/compare-02.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:180d71a6148d23f55aad7560d734768199eea660a084bacfb48d69546e20a1cf
+size 757909

samples/compare-03.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc76a01fdd5a788763f83845beb66c28527e9ac2178b7b33da5c410663cead7f
+size 757951

samples/compare-04.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19cde35ee8069d079e116ad3a8136da9370424ad29b786519e47f3ab15999cda
+size 634342

samples/compare-det-00.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4c8922ea2d7338badcc20a679ef6b8f478b06ad3245c468360ab996cdbce7e0
+size 1032779

samples/compare-det-01.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1759fc635fba7ca08d01962e8a2fd3fcc0be6e78c4c9a7b53c7a62497d2cc14
+size 1188623

samples/compare-det-02.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60a5d715cc5e5eb97de70d925367a1ff160964411100538d97da7492c62c4c7d
+size 1042953

samples/compare-det-03.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68656a2a168a3882a6fd3a7b20f30ebfbe6e810fa24123fb206f0fab991c45d6
+size 974405

samples/compare-det-04.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2ffc7182d833234da7a17b84421e337c56faa74254c3e6a813f7ac85b029c06
+size 594604

samples/sample-04.mp4

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:cf535d3764e53267f99f8aeb984b1fcc9c8fb41d79776a5ccb7e3e6520131f36
-size 730651