A100 / current-upstream ladder (2026-08-20)
Rebuilt against the encoder Lightricks published on 2026-08-17 (commit
1b92891c, "Aligns the published encoders with the LTX-2.5 model
checkpoints"). Everything measured before that date used the superseded,
misaligned encoder and is not comparable to what is here.
Runtime: Colab A100-SXM4-40GB (cc 8.0), torch 2.11.0+cu128, transformers 5.14.1
pinned, plan r45c, calibration calib-large.txt, reference regenerated on this
card from the current source.
Gate — A3 wins on every axis
A0 legacy |
A3 bypass + group bounded 12:1 |
|
|---|---|---|
| video mean relL2 | 0.06061 | 0.05204 (−14%) |
| audio mean relL2 | 0.06095 | 0.04827 (−21%) |
| video worst | 0.14733 | 0.09187 (−38%) |
| audio worst | 0.14440 | 0.10573 (−27%) |
| video cosine | 0.99763 | 0.99848 |
||Q||/||W|| |
302.654 | 4.216 (72x) |
| columns zeroed | 5 | 0 |
| max restore gain | 469.56 | 19.00 |
| damping escalation | 1x | 10x |
The trade this project spent two sessions quantifying is gone. On the superseded source and a V100, the structural fix cost about 6% of conditioning drift; here it costs nothing and improves the worst case most.
Both figures reproduced to five decimals across two independent A100 runtimes, so builds are deterministic within an architecture even though they are not across architectures.
Render prompts, against a BF16 conditioning from the same card
| # | A0 | A3 | prompt |
|---|---|---|---|
| 0 | 0.05870 | 0.05073 | humanoid robot, freight elevator |
| 1 | 0.07487 | 0.06596 | robotic hand, cracked ceramic cup |
| 2 | 0.05551 | 0.08213 | four-legged inspection robot |
| 3 | 0.05061 | 0.04489 | dozens of small wheeled robots |
| 4 | 0.08887 | 0.05286 | rescue robot in smoke |
| mean | 0.06571 | 0.05931 | |
| worst | 0.08887 | 0.08213 |
A3 is better on four of five. Prompt 2 is the exception, and it was also the exception on the superseded source with a V100 - the same prompt moving the same way after both the weights and the architecture changed is a signal, not noise, though five prompts cannot say what it means.
Visual
frames/cmp-*.png puts BF16, A0 and A3 side by side - same card, same DiT, same
seeds, same prompts.
- Prompt 0 now renders an actual robot. Dome head, two amber eyes, segmented neck, panelled torso, in all three. The superseded encoder produced a human-faced android for the same prompt. That improvement comes from the source alignment, not from anything this project did.
- BF16 and A3 track each other closely - same badge colour on prompt 0, glasses present on prompt 1 - while A0 differs on both.
- Prompt 1 still fails to a human face in all three including BF16. Model limit, not compression damage.
Artifacts
On the Hub at topabaem/LTX-2.5-TextEncoder-nvfp4-portable (private), pushed by
the build worker as each stage finished rather than downloaded afterwards:
A0.packed.safetensors, A3.packed.safetensors, and evidence/ carrying the
gate JSON, build logs with per-layer drift and ||Q||/||W||, the A100 BF16
reference, all three conditionings and all fifteen clips.
That push-as-you-go arrangement exists because three Colab runtimes died
mid-build, twice destroying finished artifacts. The render stage then failed
three times on a missing gguf package - and cost only the render, because the
builds were already safe on the Hub.