LTX 2.3 video benchmarks

Test hardware and platform limits.

Leaf control: Windows and native Linux use Diffusers leaf offloading with use_stream=True. WSL uses use_stream=False because streamed leaf offloading was incompatible with the tested WSL environment; this is not a limitation established for every WSL setup. This setting concerns weight offloading, not video VAE decode streaming.

Distilled BF16, 1280×704, 5 seconds at 24 fps (121 frames), 8 steps, seed 43; real guitar image/prompt, no input audio, generated stereo audio. Shared official leaf encoder/auxiliaries and resident temporal GPU VAE decode.

Presets resolve to the executed transformer runtimes; the VAE route is independent. Scenario names link directly to measurement details. Times are seconds; GPU is sampled absolute driver usage, RAM is process RSS, and memory units are GiB. Average is denoise wall divided by steps, including warmup, not completed warm GPU timing. Tracked total excludes command startup and downloads. Before reading the tables, see the labels and memory glossary.

In comparable overview tables, official Diffusers leaf is the first row and the reference for that table: 🟢 ↓ is lower, 🔴 ↑ is higher. Equal displayed values have no arrow. Arrows describe measurements, not an overall preset ranking; higher memory can accompany faster execution. Phase tables and separate workloads without a leaf control have no arrows. Rows with different driver/cache sessions or a reused non-side-by-side observation also have no arrows.

Windows — BF16

The primary comparison uses the default PyTorch SDPA path for video self-attention. The external FlashAttention 2 and Sol-Attn options were not selected. The specific kernel chosen internally by SDPA was not recorded.

Phase-table rows identify the complete pipeline by its transformer route. Encoder, conditioning and VAE strategies are shared as stated above; a row labeled vmm_fast does not imply native VMM execution in every phase.

A separate two-step residual CPU source ownership validation recovered checkpoint-related commit headroom with exact output equivalence. Its larger conditioned workload and experimental backing policy are not rows in the matched performance table below.

Phase measurements

These are selected separately monitored phases, not an additive decomposition of the tracked total. Cleanup, saving and other unlisted setup remain in the total.

Text encoder

Connectors

Image conditioning encoder

Audio conditioning preparation

Transformer load/prepare

Transformer attachment

Denoise

Video VAE decode

Audio VAE decode

Vocoder

Pipeline total

Measured tracked totals and global peaks for the same complete runs above. These are not sums of the phase tables. The encoder and VAE routes remain as stated for this workload; the runtime column identifies the transformer.

Linux — BF16

The NVMe observation uses a different driver/cache session from the route trio; it is not a controlled allocator-only A/B.

Phase measurements

These are selected separately monitored phases, not an additive decomposition of the tracked total. Cleanup, saving and other unlisted setup remain in the total.

Text encoder

Connectors

Image conditioning encoder

Audio conditioning preparation

Transformer load/prepare

Transformer attachment

Denoise

Video VAE decode

Audio VAE decode

Vocoder

Pipeline total

Measured tracked totals and global peaks for the same complete runs above. These are not sums of the phase tables. The encoder and VAE routes remain as stated for this workload; the runtime column identifies the transformer.

WSL — BF16

Leaf controls are non-streamed; linear uses an explicit 3 GiB override and native uses bounded registration. These are not Windows/Linux parity claims.

Phase measurements

These are selected separately monitored phases, not an additive decomposition of the tracked total. Cleanup, saving and other unlisted setup remain in the total.

Text encoder

Connectors

Image conditioning encoder

Audio conditioning preparation

Transformer load/prepare

Transformer attachment

Denoise

Video VAE decode

Audio VAE decode

Vocoder

Pipeline total

Measured tracked totals and global peaks for the same complete runs above. These are not sums of the phase tables. The encoder and VAE routes remain as stated for this workload; the runtime column identifies the transformer.

VAE strategies — separate native-encoder workload

Five seconds, 1280×704, 24 fps, 8 steps, seed 43. Native text encoder; temporal GPU streaming in both decoder routes.

Phase measurements

Saved FP8 weight-only — separate precision pairs

Five-second video; auxiliary components remain unchanged within each platform pair. Lower transformer storage does not imply faster denoising or lower peak GPU usage.

Phase measurements

WSL video with conditioning audio

Separate five-second workload with guitar image and input audio; not matched to the no-input-audio WSL comparison above.

Phase measurements

Recommendations

low_ram_safe was not the lowest-memory Windows route. These five-second tables do not establish capacity or performance for longer videos.

Single observations, not statistical or arbitrary-model guarantees. Missing measurements are unavailable, not zero; phase peaks must not be summed. See presets, reproduction and measurement details before comparing different recipes.