LTX video parameters

The four public entries use canonical names only. Use the selected runner's --help for its supported subset and shared controls for lifecycle and reporting semantics.

Interface scope by runner

Controls Minimal Linear Native VMM SDNQ
Generation, conditioning, custom blocks, LoRA, purge and reports Yes Yes Yes Yes
--ddo-backend vmm linear vmm linear
--ddo-preset auto Linear/compatibility presets auto or nvme_baseline Linear or leaf compatibility presets
Separate transformer --model-path No Yes Yes Yes
--max-sequence-length, --reuse-rotary-embeddings No Yes Yes Yes
Native capacities and execution backing reclaim Yes No Yes No
Chunking, phase modulation, compact timesteps and inplace residuals No Yes Yes No
Native encoder/VAE controls No No Yes No
--bits No No No Yes
--metrics-level No No Yes No
--ddo-profile NAME, --build-ddo-profile No Yes No No

Unsupported options are rejected, including abbreviated names. Backend and preset selection do not change the input audio or decode device automatically.

Sources and generation

Control Purpose
--distilled-model Defaults to OzzyGT/LTX-2.3-Distilled-1.1; BF16 pipeline/auxiliary source. Also supplies the BF16 transformer unless overridden; SDNQ selects its quantized transformer separately.
--model-path Optional transformer-only override; auxiliary components still use --distilled-model.
--custom-blocks Defaults to elismasilva/ltx2.3_video_custom_blocks; trusted model code, or a local development directory.
--custom-blocks-revision Optional Hub branch, tag or commit; omission uses main. Does not switch a local clone. With --local-files-only, only the cached snapshot is available.
--local-files-only False; require cached model/code files when enabled. Input-media URLs are separate.
--prompt Defaults to the man plays the guitar and sings on an old western town.
--image Defaults to the guitar image URL from the OzzyGT Diffusers examples dataset; accepts a path or URL.
--audio, --use-audio-condition No default audio; conditioning is off. Enabling requires an explicit audio file. Accepts --no-use-audio-condition.
--width, --height 1280 × 704; positive multiples of 32.
--seconds, --frame-rate 5 seconds, 24 fps; finite positive values.
--steps, --seed 8 steps, seed 43; distilled entries accept 1–8 steps.
--max-sequence-length 256; positive text token limit in linear, VMM and SDNQ.
--bits 8 in SDNQ; choose 4 or 8 for the corresponding saved transformer. Auxiliary components remain BF16.
--lora, --lora-weight-name, --lora-revision, --lora-scale Optional transformer adapters; see LoRA usage for repeated adapters and defaults.

Input-audio conditioning and generated output audio are different. The default image does not supply an audio file. Custom blocks execute trusted Python code.

Native capacity and components

The VMM entry exposes the same DDO route/capacity and reclaim controls as the image entries. Capacities use GiB; they are independent limits, not total GPU usage.

Control Default and meaning
--ddo-backend, --ddo-preset vmm, auto; experimental native fast. nvme_baseline selects the protected NVMe route.
--ddo-vmm-resident-weight-budget-gb 4 GiB; active transformer weight capacity.
--ddo-vmm-total-staging-gb Omitted: native fast uses 0.625 GiB total host staging; NVMe uses two 0.625 GiB slots. Explicit values specify total staging for either route.
--ddo-vmm-source-cache-gb 0; native retained-source cache disabled.
--ddo-vmm-source-registered-gb 0; native registered-source ceiling inherits the source-storage ceiling.
--ddo-vmm-execution-backing-reclaim False; opt into reclaim of eligible idle execution backing under pressure. Requires vmm/auto; the environment variable does not enable it in public VMM entries. See behavior and tradeoffs.
--text-encoder-native-vmm Off; opt into native text-encoder ownership on vmm/auto.
--text-encoder-vmm-resident-weight-budget-gb Required positive finite encoder capacity with the preceding switch; separate from the transformer ceiling.
--transformer-fp8-weight-only Off; load supported saved TorchAO FP8 weights. Requires native fast, a local --model-path and --local-files-only; does not convert BF16 weights.
--release-pytorch-host-cache-before-transformer Off; request PyTorch host-cache release at the phase boundary. Separate from Windows standby purge and native source-cache capacity.

Native API and transport selection are automatic. Source-cache settings apply to native fast, not the protected NVMe runtime. Benchmark capacities are explicit overrides rather than these defaults.

Model computation

Control Default and meaning
--reuse-rotary-embeddings True in linear, VMM and SDNQ; prepare generation-local RoPE once. Accepts --no-reuse-rotary-embeddings.
--feed-forward-chunk-size 1024 in linear, 0 in VMM. Nonnegative token chunk size; 0 disables chunking.
--normalization-chunk-size 1024 in linear, 0 in VMM. Nonnegative token chunk size; 0 disables chunking.
--phase-modulation On in linear, off in VMM; use phase-local modulation buffers. Linear accepts --no-phase-modulation.
--compact-timesteps On in linear, off in VMM; project spatially uniform video timesteps without the dense repeated representation. Unsupported layouts remain dense. Linear accepts --no-compact-timesteps.
--inplace-residuals On in linear, off in VMM; reuse model-owned residual states in inference mode. Linear accepts --no-inplace-residuals.

These are custom-block computation options, not generic DDO presets. Preserve the exact choices when comparing measurements. Compilation and experimental fused RMS/rotary kernels belong in the diagnostics entry.

Decode and output

Control Default and meaning
--decode-device cpu; choose cuda:0 explicitly for GPU decoding.
--stream-video-decode Off; temporal window decoding requires --decode-device cuda:0.
--vae-leaf-offload Off, VMM only; requires GPU streaming. Audio VAE and vocoder routes stay separate.
--output-type video, VMM only; latent skips output decoding and cannot be combined with streaming.
--output-dir outputs/ddo_modular_video, or outputs/ddo_modular_sdnq_video for SDNQ.

Lifecycle and reports

All four entries expose --purge-before-run, --purge-after-text-encoder, --purge-before-transformer and --purge-after-run, with negative forms. Omission inherits DDO environment/preset policy; it does not force purge off. See Windows standby purge.

Control Default and meaning
--memory-monitor rss; full additionally collects USS and can add overhead.
--metrics-level 2 in VMM; accepts 0/1/2. Level 2 records public DDO component snapshots.
--show-profile Off; display the component profile.
--show-metrics, --save-metrics Both on; display observations and save JSON.
--dry-run Off; print a phase plan without weights, network or CUDA.
--help List the selected entry's arguments without generating video.

Display/save switches accept their --no-... forms. Phase observations do not prove GPU completion or local VRAM placement. CPU parser checks do not establish GPU or performance acceptance.

Video self-attention backend

All LTX video runners accept --video-self-attention with default, flash-attn2, or sol-attn. The default preserves the model's configured attention. flash-attn2 uses the optional external FlashAttention 2 package. sol-attn uses Triton Sol-Attn (diag, tau=1.0) and uses FlashAttention 2 for the first transformer block. These choices apply only to eligible video self-attention. Audio, cross-attention, and masked or perturbed calls retain their original routes. Saved metrics include per-layer dispatch counts.

Install the optional FlashAttention 2 wheel as described in Setup. For Sol-Attn, make the Sana techniques/sparse_backends directory importable in the runner environment. The tested Sana revision is 781e834. The selected GPU validation used the VMM Distilled runner on an RTX 3060 Ti. Sol-Attn accepts dynamic token counts and requires CUDA BF16 video attention with head dimension 128. The validated recipe used 14,080 video tokens (1280 × 704, 5 seconds at 24 fps). A 10-second run encountered an illegal CUDA memory access during denoising. Its cause remains under investigation.