LTX-2.5 pipelines: Distilled, DFR or two-stage, which to run

For a first LTX-2.5 run, use DistilledPipeline: the repository describes it as the fastest text/image-to-video option, started with python -m ltx_pipelines.distilled. Choose DFRPipeline for production-quality generation with spatial detailing, and TI2VidTwoStagesPipeline when you want guided generation with CFG and STG.

What pipelines does the README list?

What else does the README set?

The README lists an optional duration head, ltx-2.5-duration-head-bf16.safetensors, so you can omit --num-frames and have length predicted from the prompt. Memory flags are --quantization fp8-cast and --offload to cpu or disk. It says gradient estimation reduces steps from 40 to 20 to 30. A prompt enhancer is on by an enhance_prompt parameter; the README gives no further detail on it, so test with it on and off.

What are the nearest hosted fields?

Sume does not list LTX (catalog code, read 2026-10-09), so these are the nearest request shapes on listed models, not equivalents.

• Keyframes: frame_images with first_frame and last_frame, on models whose supported_frame_images include both.
• Audio-driven: input_references with an audio_url, on models that list audio references such as minimax-h3 and wan-3.0.
• Edit an existing clip: the Video Router video_url edit on gemini-omni-flash-1.1 (whole-clip edit, not a region retake).
• Guided steps, CFG or STG: not exposed.

What is a sensible order to try them?

Run DistilledPipeline first to prove the install and get a baseline time. Move to DFRPipeline or TI2VidTwoStagesPipeline for the clips that matter, and compare against the distilled output on the same prompt. Use the specialised pipelines (keyframes, audio-to-video, dubbing, retake) only when the task calls for them; each one has its own inputs. For example, A2VidPipelineTwoStage needs an audio file, and KeyframeInterpolationPipeline needs the keyframes you want to hit, so gather those before you start.

Write the pipeline name, the checkpoint and the flags beside every output. With eleven pipelines and several checkpoints, that note is what keeps your tests comparable.

Check each id's fields at GET /v1/videos/models before you build; the video docs list them.