LTX 2.3 image — measurement details

Back to comparison tables. Scenario links in those tables open the sections below; no local report download is required to read the results.

Labels and units: read the labels and memory glossary before comparing runtime names, phase timings, tracked totals or memory peaks.

Workload and measurement scope

Distilled BF16, 1280×704, 8 steps, seed 43, real prompt, CFG 1/rescale 0.7, no PAG. Shared official leaf text encoder and resident untiled VAE in the primary comparison.

Checkpoint: elismasilva/ltx2.3-image-distilled-1.1, snapshot 5cd2f057b18d3964f7dc1ce9ab061e9dd22754b9. Official Diffusers 0.41.0 in the primary platform comparisons. Encoder setup differs in the separately labeled native-encoder VAE pairs. No LoRAs, compute compilation, PAG or tiling are added to the primary cases.

The text-encoder row is the monitored Pass 0: Encode prompts: loading, tokenizer/pipeline setup, encoding and cleanup, plus any purge inside that phase. It is not the pure encode_prompt_call duration. Connectors run in the Pass 1: Generate aggregate together with transformer loading, attachment, denoise and cleanup. Its memory peak cannot be relabeled as a denoise-only peak.

The overview's denoise wall is denoise_modular_pipe_call. VAE rows use separate load/decode monitors where recorded. INT4 reports instead monitor Pass 2: Decode VAE as a load/decode/cleanup aggregate: it must not be compared to a pure decode time as if their scopes were identical.

All displayed phase rows are existing observations, not new runs or inferred phase estimates. Memory comes from each recorded phase's peak_vram_absolute_gb and peak_ram_gb with ram_metric=rss. Times use elapsed_sec; global peaks and tracked totals come from the same report as that row. Values are rounded to two decimals for display only. Missing fields remain unavailable. Phase tables show selected boundaries, not an additive decomposition; do not add aggregates to their child events or sum memory peaks. Host callbacks do not establish completed warm GPU timing. See metric definitions.

Platform conditions

Windows comparisons recorded successful standby purge before the transformer; phase-local purge can also be included in tracked totals. Linux and WSL do not execute the Windows purge. WSL auxiliary leaf hooks are non-streamed and native registration is bounded; those routes are not equivalent to Windows/Linux streamed controls. Do not treat different platform observations as causal A/Bs.

Linux protected NVMe observations dated 2026-10-13 use driver 595.99.02 and a fresh post-reboot/cache session, unlike the earlier route trios. This prevents attributing their timing difference solely to the allocator. Primary cases are single observations, not repeated statistical estimates or low-RAM certification. The compared completed outputs matched within the recorded route sets; this is not a universal quality or cross-platform byte-identity claim.

BF16 scenarios

Windows / Official leaf

The diffusers_leaf_offload_compat control executes official Diffusers leaf hooks, not DDO dynamic denoising.

Windows / one_shot_fast

--ddo-backend linear --ddo-preset one_shot_fast executes linear_runtime.

Windows / low_ram_safe

--ddo-backend linear --ddo-preset low_ram_safe executes linear_runtime.

Windows / vmm_fast

--ddo-backend vmm --ddo-preset auto resolves to vmm_fast, executing vmm_fast_runtime.

Fixed transformer capacity: 4 GiB active weights; 1536 MiB total staging. Retained source / registration ceilings: 24.625 / 24.625 GiB. Native allocation is process-wide: encoder/VAE driver baselines can differ even when their component offload policies are unchanged.

Windows / nvme_baseline

--ddo-backend vmm --ddo-preset nvme_baseline executes protected vmm_runtime.

Fixed transformer capacity: 4 GiB active weights; two 640 MiB staging slots. Retained source / registration ceilings: 0 / 0 GiB. Normal PyTorch allocation plus native weight arena, not the process-wide native Torch allocator. Zero persistent source cache is not zero OS file-cache usage.

Linux / Official leaf

The diffusers_leaf_offload_compat control executes official Diffusers leaf hooks, not DDO dynamic denoising.

Linux / one_shot_fast

--ddo-backend linear --ddo-preset one_shot_fast executes linear_runtime.

Linux / vmm_fast

--ddo-backend vmm --ddo-preset auto resolves to vmm_fast, executing vmm_fast_runtime.

Fixed transformer capacity: 4 GiB active weights; 1536 MiB total staging. Retained source / registration ceilings: 24.625 / 24.625 GiB. Native allocation is process-wide: encoder/VAE driver baselines can differ even when their component offload policies are unchanged.

Linux / nvme_baseline

--ddo-backend vmm --ddo-preset nvme_baseline executes protected vmm_runtime.

Fixed transformer capacity: 4 GiB active weights; two 576 MiB staging slots. Retained source / registration ceilings: 0 / 0 GiB. Normal PyTorch allocation plus native weight arena, not the process-wide native Torch allocator. Zero persistent source cache is not zero OS file-cache usage.

WSL / Official leaf (no stream)

The diffusers_leaf_offload_compat control executes official Diffusers leaf hooks, not DDO dynamic denoising.

WSL / wsl_compat (3 GiB override)

--ddo-backend linear --ddo-preset wsl_compat executes linear_runtime.

The resident-module budget is explicitly 3 GiB, CPU pinning is disabled and automatic budgeting is off. This validates the override, not preset-default capacity; a configured ceiling is not measured residency.

WSL / vmm_fast (bounded)

--ddo-backend vmm --ddo-preset auto resolves to vmm_fast, executing vmm_fast_runtime.

Fixed transformer capacity: 4 GiB active weights; 384 MiB total staging. Retained source / registration ceilings: 24.625 / 0.5 GiB. Native allocation is process-wide: encoder/VAE driver baselines can differ even when their component offload policies are unchanged.

VAE strategies — separate native-encoder workload

1280×704, 8 steps, seed 43. These pairs use a native text encoder, unlike the shared-leaf-encoder comparisons above.

Windows / vmm_fast / resident VAE

Both encoder and transformer use native VMM. Only the independently selected VAE route changes within each platform pair; outputs match within each pair. These runs predate the primary shared-leaf-encoder cases and are a separate VAE workload, not replacement transformer speed measurements.

Windows / vmm_fast / leaf VAE

Both encoder and transformer use native VMM. Only the independently selected VAE route changes within each platform pair; outputs match within each pair. These runs predate the primary shared-leaf-encoder cases and are a separate VAE workload, not replacement transformer speed measurements.

Linux / vmm_fast / resident VAE

Both encoder and transformer use native VMM. Only the independently selected VAE route changes within each platform pair; outputs match within each pair. These runs predate the primary shared-leaf-encoder cases and are a separate VAE workload, not replacement transformer speed measurements.

Linux / vmm_fast / leaf VAE

Both encoder and transformer use native VMM. Only the independently selected VAE route changes within each platform pair; outputs match within each pair. These runs predate the primary shared-leaf-encoder cases and are a separate VAE workload, not replacement transformer speed measurements.

Windows SDNQ INT4

Saved INT4 transformer, 1280×704, 8 steps, seed 43; resident VAE. Compare within this pair, not as an isolated BF16-versus-INT4 speedup.

Windows / SDNQ / one_shot_fast

Saved LTX image INT4 transformer; BF16 auxiliaries, stock resident VAE and official Diffusers 0.41.0. Encoder/setup choices differ from the BF16 tables. one_shot_fast executes packed sdnq_runtime, preserving SDNQ math. Normal PyTorch allocation, not native VMM. Outputs match within the INT4 pair; BF16-versus-INT4 quality equivalence was not established.

Windows / SDNQ / official leaf

Saved LTX image INT4 transformer; BF16 auxiliaries, stock resident VAE and official Diffusers 0.41.0. Encoder/setup choices differ from the BF16 tables. diffusers_leaf_offload_compat executes official streamed leaf with SDNQ math. Normal PyTorch allocation, not native VMM. Outputs match within the INT4 pair; BF16-versus-INT4 quality equivalence was not established.

Windows / warm_process repeated workload

1280×704 BF16, real prompt, eight steps per generation, two transformer preparations and two generations in one process. Shared streamed leaf encoder and resident untiled VAE. Initial seed 43; the shared generator advances between generations, so the final image need not equal a single-generation seed-43 image. The plan and pinned-weight cache are reused. The total covers both generations and final decode, not one image. No comparative repeated-preset ranking follows.

Installed rc3 functional validation

The same local DDO/allocator rc3 Linux wheels completed separate native-Linux and WSL image checks without rebuilding. Both used installed packages, Python 3.12.12, Torch 2.12.1+cu132 and Diffusers 0.41.0. The workload was the BF16 checkpoint/revision above, 1280 x 704, eight steps, seed 43, the runner's real cat prompt, CFG 1/rescale 0.7 and no PAG. Each run included one generation, one transformer preparation and resident VAE decode. Transformer runtime was vmm_fast_runtime, requested through vmm/auto with base one_shot_fast; the text encoder used official group offload and VAE route was none.

The table rounds for display. Exact reproduction values are weights 4, source cache 24.625 and pages 32 MiB on both platforms; --ddo-vmm-total-staging-gb 1.5 --ddo-vmm-source-registered-gb 24.625 on native Linux, and --ddo-vmm-total-staging-gb 0.375 --ddo-vmm-source-registered-gb 0.5 on WSL. Use scripts/ltx2_image/run_dynamic_modular_vmm_distilled.py, local model files, --ddo-backend vmm --ddo-preset auto, --ddo-vmm-resident-weight-budget-gb 4, --ddo-vmm-source-cache-gb 24.625, dimensions/steps/seed above, one pre-VAE cleanup and level-2 metrics. No compression, activation eviction, ETW, GPU timing or automatic execution-backing reclaim was enabled. Windows standby purge is a no-op on these targets. The controller allowed 1,800 s with 10 s progress and 16 MiB per log stream, without retry or automatic termination. The runner guide explains total staging. The current spelling above selects the same public runtime; API selection, the one_shot_fast base settings and 32 MiB pages are implicit. Original recorded CLI spellings remain available through the separate diagnostic entry; the table reports the original runs, not new validation of the CLI revision.

Each rc3 run exited 0 after eight steps and decode. Its PNG and decoded RGB matched the retained platform reference exactly. Available pre-close native failure and quarantine counters were zero; decode and cleanup returned. Final native counters, latent finiteness and final driver-memory release were not established by the unchanged runner. Native Linux emitted a preserved group-offload lazy-prefetch warning for unused vision-tower layers; this text-prompt check does not validate image-conditioned vision-tower execution.

These single historical observations close selected functional checks, not statistical performance acceptance. Platform capacities, host/cache conditions and registration differ; do not use the table as a Linux-versus-WSL ranking or a causal speed comparison. Neither full video nor arbitrary models are covered. Rejected preparations and staging-mismatch captures were retained separately and are excluded from these accepted observations. Package availability and artifact delivery remain separate release gates.

Reproduce this workload

Use the runner guide and comparison controls, retaining the model, schedule, inputs, component routes and the scenario-specific capacities above. Runner defaults or a different custom-block revision do not guarantee equivalent conditions. Review trusted custom-block code and keep platform environments separate. Package/build validation and raw engineering investigations are developer concerns, not additional timing rows or prerequisites for interpreting these tables.