LTX 2.3 video

Model and requirements

The runners load LTX video custom blocks. Complete installation, including FFmpeg configuration on Windows when loading audio.

Available runners

Inspect an entry without generating video:

python -B scripts/ltx2_video/run_dynamic_modular_vmm_distilled.py --help

Run the model

Choose the platform-specific native recipe in video benchmarks. The comparison workload is 1280×704, 5 seconds at 24 fps, 8 steps and seed 43, with guitar-image conditioning. Input-audio conditioning is a separate workload.

For that Windows native workload, the public command uses the documented capacities and computation options below. Run in an elevated terminal for the explicit Windows purge phases. This command generates video; it is not a dry run.

python -B scripts/ltx2_video/run_dynamic_modular_vmm_distilled.py `
  --ddo-backend vmm --ddo-preset auto `
  --width 1280 --height 704 --seconds 5 --frame-rate 24 --steps 8 --seed 43 `
  --ddo-vmm-resident-weight-budget-gb 1.5 `
  --ddo-vmm-total-staging-gb 1.5 `
  --ddo-vmm-source-cache-gb 24.625 --ddo-vmm-source-registered-gb 24.625 `
  --ddo-vmm-execution-backing-reclaim `
  --feed-forward-chunk-size 1024 --normalization-chunk-size 1024 `
  --phase-modulation --compact-timesteps --inplace-residuals `
  --decode-device cuda:0 --stream-video-decode `
  --purge-before-run --purge-after-text-encoder --purge-before-transformer --no-purge-after-run `
  --memory-monitor rss --metrics-level 2 --show-profile --show-metrics --save-metrics

This usage recipe explicitly enables execution backing reclaim. That opt-in does not change the recorded conditions or results of historical benchmark rows.

To condition on input audio, add --use-audio-condition --audio C:/media/song.mp3 with your actual file. Keep the same image and audio sources when comparing durations; --seconds 10 changes the duration without replacing either source. Conditioned results are separate from the image-only benchmark above.

For GPU temporal decoding, include these options in the selected recipe:

--decode-device cuda:0 --stream-video-decode

They are explicit because the current parser still defaults to CPU decode without streaming. Temporal streaming reduces the VAE's working set; it does not change the transformer route. Audio VAE and vocoder execution are staged separately.

Runtime and preset selection

Linear recipe

run_dynamic_modular_distilled.py defaults to the recorded linear computation settings: feed-forward and normalization chunks of 1024, phase modulation, compact timesteps, inplace residuals and RoPE reuse. These are model computation options, separate from the DDO memory plan. Minimal and SDNQ entries do not inherit these defaults.

python -B scripts/ltx2_video/run_dynamic_modular_distilled.py `
  --ddo-backend linear --ddo-preset one_shot_fast `
  --width 1280 --height 704 --seconds 5 --frame-rate 24 --steps 8 --seed 43 `
  --decode-device cuda:0 --stream-video-decode `
  --purge-before-run --purge-after-text-encoder --purge-before-transformer --no-purge-after-run `
  --memory-monitor rss --show-profile --show-metrics --save-metrics

Input audio remains optional: add --audio C:/media/song.mp3 --use-audio-condition to preserve your conditioning source. That changes the workload relative to the historical image-conditioned benchmark without input audio.

The memory plan remains automatic. The recorded Windows linear run selected full CPU weight pinning and a zero automatic resident-module budget. This is not a fixed 6 GiB GPU recipe: available RAM determines whether the same plan is admissible today. pin=skip_insufficient_ram indicates a different plan and can affect latency. Computation defaults alone do not guarantee the historical time. CPU decode remains the parser default; the recipe explicitly selects temporal GPU decoding. Run elevated on Windows to preserve the requested purge phases.

linear/auto selects one_shot_fast; reducing a resident ceiling does not select low_ram_safe. In particular, DDO_MAX_RESIDENT_MODULE_BUDGET_GB=3 only caps the automatic resident-module budget. Full CPU pinning can still select a zero budget. To request the conservative linear policy, explicitly use --ddo-preset low_ram_safe: its defaults disable automatic budgeting and CPU pinning and set a 3 GiB resident-module budget. Neither value limits total GPU memory. Explicit environment settings can override preset defaults; remove a temporary override after its test. The startup log distinguishes the selected preset, automatic ceilings and fixed budgets.

Use presets for backend and memory tradeoffs. --distilled-model selects the BF16 pipeline/auxiliary source; --model-path overrides only the transformer. Native encoder offload is a separate option with its own capacity. Windows standby purge uses the shared controls.

Optional LoRA

Add these arguments to a BF16 recipe to load Soft Enhance at strength 0.8:

--lora vrgamedevgirl84/LTX_2.3_Soft_Enhance_Style_LoRa --lora-weight-name LTX2.3_Soft_Enhance.safetensors --lora-scale 0.8

--lora also accepts a local safetensors file or directory. Use --lora-revision to pin a Hub revision. No adapter is loaded by default. For multiple adapters, repeat --lora and --lora-scale in the same order. When specifying filenames or revisions, provide one per adapter; use auto for an automatically selected filename or default revision. For example:

--lora vrgamedevgirl84/LTX_2.3_Soft_Enhance_Style_LoRa --lora-weight-name LTX2.3_Soft_Enhance.safetensors --lora-scale 0.8
--lora Lightricks/LTX-2-19b-LoRA-Camera-Control-Static --lora-weight-name ltx-2-19b-lora-camera-control-static.safetensors --lora-scale 0.5

Both adapters stay active. Additional adapters can increase memory and execution cost; loading compatibility does not establish visual quality or CUDA acceptance. The adapter stays unfused and is loaded before offload; the base checkpoint is not rewritten. Only transformer adapters are accepted. Saved FP8 plus LoRA is not enabled by this recipe.

The console confirms the active adapter and strength; saved metrics include lora_adapters. Loading checks do not establish native VMM execution support: that combination still needs a CUDA comparison. Visual changes alone are not a reliable activation check.

Optional weight formats

SDNQ

Install only the SDNQ optional backend; FP8 is not a prerequisite. Use --bits 4 or --bits 8 in the SDNQ runner to choose the corresponding saved OzzyGT transformer. Auxiliary components remain BF16. --ddo-preset diffusers_leaf_offload_compat selects its Diffusers leaf control. Image measurements do not establish video performance.

FP8 weight-only

Install the optional fp8 dependencies and supply a previously saved supported TorchAO FP8 weight-only transformer:

--model-path F:/models/your-converted-model --transformer-fp8-weight-only

Add these to the VMM recipe, keeping its auxiliary BF16 pipeline source. The flag does not convert BF16 weights or accept arbitrary FP8 checkpoints. It reduces weight storage; it does not promise FP8 compute or faster generation. Compilation is not required. See DDO's FP8 guide.

Metrics and profile

Add --show-profile --show-metrics --save-metrics. --output-dir selects media and JSON output; the runner prints saved paths. Use --memory-monitor rss for comparisons with the RSS benchmark tables.

Runner-specific parameters

Linear capacity profile

Calibrate the conditioned five-second guitar workload below. Use your actual image/audio paths and an elevated terminal on Windows for purge. This runs the eight denoise steps with calibration enabled and stops before decode/export.

python -B scripts/ltx2_video/run_dynamic_modular_distilled.py `
  --ddo-backend linear --ddo-preset one_shot_fast `
  --width 1280 --height 704 --seconds 5 --frame-rate 24 --steps 8 --seed 43 `
  --image C:/media/man_guitar.png --audio C:/media/song.mp3 --use-audio-condition `
  --ddo-profile ltx-video-bf16-1280x704-5s-conditioned --build-ddo-profile `
  --decode-device cuda:0 --stream-video-decode `
  --purge-before-transformer --memory-monitor rss --save-metrics `
  --output-dir outputs/ltx_video_linear_profile

The decode options are included for subsequent generation; calibration stops before decode and does not measure decoder time or memory. To generate using the profile, remove --build-ddo-profile and choose a new output directory. Keep --decode-device cuda:0 --stream-video-decode for the GPU streaming decoder. The linear computation defaults remain chunks of 1024, phase modulation, compact timesteps, inplace residuals and RoPE reuse. Recalibrate under a different name when changing duration, resolution, conditioning or math options. See profile storage and scope.

See video parameters and shared controls. Compute experiments are optional, not requirements for DDO offloading.