LTX image parameters

This page lists the public LTX image controls, with native VMM defaults in the detailed tables and differences for the other entries below. See shared controls for the DDO route/capacity and reclaim options, reporting defaults and Windows purge switches. --help / -h prints the installed checkout's interface without generating an image. Each entry exposes only implemented controls; shared controls keep one spelling.

Interface scope by runner

Entry Code source selection CLI scope
run_dynamic_modular_vmm_distilled.py --custom-blocks / --custom-blocks-revision Native route selection, capacities, component loading and reporting.
run_dynamic_minimal.py --custom-blocks / --custom-blocks-revision Minimal native VMM transformer integration; canonical names, basic sampling and reporting.
run_dynamic_modular_distilled.py --custom-blocks / --custom-blocks-revision Linear/compatibility entry; canonical names, component loading and PAG.
run_dynamic_modular_sdnq_distilled.py --custom-blocks / --custom-blocks-revision Packed SDNQ/leaf entry; canonical names and quantization-specific options.
run_dynamic_modular_vmm_diagnostics_distilled.py Existing fixed local clone Separate full native experiment interface, intentionally broader than the public entry.

All four public entries use the shared parameter vocabulary. Select the runtime with --ddo-backend and --ddo-preset; the three non-VMM entries support backend linear only and default to preset auto. SDNQ preserves its packed forward and supports auto, one_shot_fast, low_ram_safe, wsl_compat, warm_process and diffusers_leaf_offload_compat. The latter selects the official leaf control. This naming cleanup does not replace quantized math with a dense linear forward.

Controls Minimal Linear SDNQ VMM
Checkpoint, custom blocks, device, prompt, dimensions, steps, seed, guidance, output Yes Yes Yes Yes
Unified --lora options Yes Yes Yes Yes
Purge phases, memory monitor, metrics display/save, denoise progress Yes Yes Yes Yes
--metrics-level, decode timestep/noise, PAG, loading controls No Yes Yes Yes
--text-encoder-dynamic-offload No Yes No Yes
VMM capacities and execution backing reclaim Yes No No Yes
--local-files-only, --show-profile, native encoder/VAE overrides No No No Yes
--bits, --text-encoder-bits No No Yes No
--ddo-profile NAME, --build-ddo-profile No Yes No No

The minimal entry uses its existing component defaults. Linear and SDNQ retain their preset-driven profile display; they do not expose a separate profile display switch. The BF16 modular linear entry separately exposes persistent capacity calibration as shown below. Their output directories default to outputs/ltx_image_modular and outputs/ltx_image_modular_sdnq_distilled; minimal also uses outputs/ltx_image_modular. SDNQ accepts --bits 4 (default) or 8 for transformer checkpoint selection; --text-encoder-bits is unset by default (original encoder) and accepts 8 only. In SDNQ, --model-path selects the auxiliary pipeline weights; the packed transformer and text encoder sources remain selected by their bit-depth settings. The default text encoder group offload is enabled. No unimplemented dynamic encoder option is accepted.

Execution

Control Purpose and default
--model-path PATH_OR_ID Pipeline checkpoint. Default elismasilva/ltx2.3-image-distilled-1.1, overridable by the MODEL_PATH environment variable.
--custom-blocks SOURCE Trusted executable LTX image code: a local directory containing modular_config.json or a Hub repository. Defaults to the existing modular_diffusers/custom_blocks/ltx2_image clone under the examples root.
--custom-blocks-revision REVISION Optional Hub code revision; prefer a commit SHA. Rejected for local directories; does not switch a local Git branch.
--local-files-only VMM only, default off; require cached weights/code rather than downloading.
--device Execution device; default cuda:0. Does not establish support for other devices.
--prompt TEXT Text conditioning; default is the calico-cat/fisheye/flower-crown prompt in the runner.
--negative-prompt TEXT Negative conditioning; default empty.
--width, --height Dimensions; defaults 1280 and 704.
--steps, --seed Defaults 8 and 43.
--guidance-scale Guidance strength; default 1.0.
--guidance-rescale Guidance rescaling; default 0.7.
--decode-timestep VAE decode timestep; default 0.0.
--decode-noise-scale Optional decode noise scale; default unset (None).
--pag Enable perturbed-attention guidance; disabled by default.
--pag-scale PAG strength when enabled; default 0.2.
--pag-layers Comma-separated transformer layer indices for PAG; default 28.
--output-dir DIRECTORY Default outputs/ltx_image_modular_vmm, relative to the working directory; output images and enabled reports.

The dense entries use BF16; SDNQ keeps its packed backend. None exposes Qwen's --variant selector. Custom blocks are Python code, not weights: loading them executes trusted code. All LTX image phases, transformer and connectors come from the same resolved code namespace. Missing required exports fail instead of falling back to a different checkout. Reports record the requested source/revision and resolved location/snapshot. In VMM, --local-files-only also applies to Hub code resolution. Other entries retain their existing download behavior. Source selection has CPU validation; GPU equivalence for a different revision still requires separate validation.

Components and loading

Control Purpose and default
--low-cpu-mem-usage Default on; use the lower-CPU-memory transformer loading path.
--text-encoder-low-cpu-mem-usage Default on; lower-CPU-memory text encoder loading.
--text-encoder-group-offload Explicit encoder group/leaf compatibility override. Omitted inherits environment/preset resolution; ordinary default enabled.
--text-encoder-dynamic-offload Explicit encoder DDO override. Omitted inherits environment/preset resolution; ordinary default disabled.
--text-encoder-native-vmm Default off. Select native encoder ownership independently of the transformer; requires the native vmm/auto route and cannot be combined with explicit group-on or dynamic-off compatibility overrides.
--vae-leaf-offload Default off; explicitly select no-stream leaf VAE offload instead of the resident decoder.

Boolean loading/group/dynamic controls also accept --no-.... The VAE leaf override accepts --no-vae-leaf-offload. This entry does not expose native idle cache trim, cuBLAS clearing, VAE tiling or the Qwen single-frame VAE controls.

LoRA

No adapter is loaded by default. Adapters are loaded on the CPU transformer before offload hooks, without fusing or replacing the backend forward.

Control Purpose
--lora SOURCE Transformer-only LoRA file, directory or Hub source; repeat for multiple adapters.
--lora-weight-name NAME Filename per source; auto selects automatically. Not accepted with a direct local file.
--lora-revision REVISION Hub revision per source; auto uses the default revision. Not accepted with a direct local file.
--lora-scale VALUE Finite strength per adapter; default 1.0 for each.

These four LoRA controls are shared by all public LTX image entries; SDNQ does not expose separate named soft/crisp adapter switches. When providing filenames, revisions or scales, supply each option once per --lora, in the same order. These options require --lora. Native VMM with LoRA still requires separate CUDA validation; see the model guide.

Keep encoder, VAE and transformer choices fixed when reproducing a benchmark.

Reporting and diagnostic entries

The public entry accepts metrics levels 0, 1 and 2. Level 2 includes public DDO component reports; lower levels omit those detailed snapshots. Display/save switches are independent; level 0 does not mean all monitoring is disabled.

Among the LTX image entries, --fake-prompt is available only in run_dynamic_modular_vmm_diagnostics_distilled.py; it skips real prompt encoding and is not a full-pipeline benchmark. Native adapter and workload calibration for native VMM also belongs to that diagnostic entry. The BF16 modular linear runner separately supports --ddo-profile NAME and --build-ddo-profile; see the linear calibration recipe. Profiles and metrics use the shared switches. See developer diagnostics before interpreting timings.