Preserve the upstream artifacts under sources/upstream/isi-dev/ as read-only references.
Prefer the scripted bootstrap path first.
- local setup launcher: scripts/run-remote-gpu-setup.ps1
- local start launcher: scripts/run-remote-gpu-start.ps1
- remote installer: scripts/setup-remote-ltx23-comfyui.sh
- remote starter: scripts/start-remote-comfyui.sh
- detailed usage: references/scripted-setup.md
Read references/source-materials.md if provenance or upstream comparison matters.
Read references/setup-and-models.md before environment or model work.
Read references/experiment-tracking.md when running real sweeps or comparing model variants on the remote GPU machine.
Read references/operational-lessons.md before repeating a setup or generation workflow that previously failed.
Read references/usage-and-parameters.md before changing prompts or generation settings.
Use sources/api/ltx23-ti2v-audio-api-prompt.json as the committed API-format base prompt when queueing runs directly through ComfyUI /prompt.
If you add Python helpers in this repo, run them with uv run ....

Choose the Source of Truth

Use the App JSON when the task is "import this into ComfyUI" or "explain which controls the App exposes".
Use the notebook when the task is "bootstrap the remote GPU machine / ComfyUI stack" or "recover the dependency list".
Use the committed API prompt when the task is "run this workflow through ComfyUI API without repeating the UI export step".
When the task is actual environment setup on a remote Linux GPU host, prefer running the scripted bootstrap that was derived from the notebook instead of replaying notebook shell lines manually.
When notebook notes and graph internals disagree, prefer the actual JSON graph for exposed controls and use the notes as soft operating guidance. Call out mismatches instead of flattening them.

Scripted Setup Rules

Use scripts/run-remote-gpu-setup.ps1 from Windows when the operator has an SSH path to the remote GPU machine.
The setup launcher streams and runs scripts/setup-remote-ltx23-comfyui.sh on the remote side.
Start the service with scripts/run-remote-gpu-start.ps1, which streams scripts/start-remote-comfyui.sh.
Keep the notebook as the source of truth for dependency and model selection, but centralize operational fixes in the script.
Preserve the script guards for:
- LD_LIBRARY_PATH=/usr/lib64-nvidia
- PIP_CONFIG_FILE=/dev/null
- writable install-directory resolution for /content/ComfyUI versus ~/ComfyUI
If setup fails, inspect which step broke and patch the script rather than falling back to ad hoc command replay.

Experiment Tracking Rules

When you run a real generation batch, save a committed manifest under experiments/<date>-<slug>/.
If prompt, image, audio, or duration are fixed, record them once as top-level fixed inputs.
For each run, record:
- model file
- width and height
- seed
- key switches such as USE ONLY VOCALS
- prompt_id
- output filename and remote path
- media duration and size
- GPU peak memory, total memory, headroom, and peak utilization
Prefer comparing one or two variables at a time so the manifest stays readable.
If the user says prompt or image should stay fixed, do not change them across the batch.

Recurrence Guards

If sources/api/ltx23-ti2v-audio-api-prompt.json exists and the upstream workflow has not changed, use it instead of repeating UI-to-API conversion.
If the operator is on Windows, prefer the launcher scripts over hand-running remote bash. They already normalize line endings and protect against stale install directories.
Re-check CUDA visibility before long runs, not only after setup.
Treat stale /content/ComfyUI directories and broken host pip config as known failure modes.
When a comparison batch is about models, keep prompt, image, audio, and duration fixed unless the user explicitly asks to vary them.
Record generation time and GPU headroom together. Peak VRAM without timing context is not enough for capacity planning.
If output files are copied back locally, verify local file sizes against the remote originals before trusting the comparison set.

Default Operating Model

Run the scripted setup path unless the user explicitly wants a manual notebook walkthrough.
Start ComfyUI after setup instead of assuming the installer leaves a service running.
Verify ComfyUI plus the required custom nodes are installed.
Verify the expected model files exist before blaming the graph.
Import sources/upstream/isi-dev/LTX_2.3_Image_or_Text_&_Audio_2_Video_App_V3.json.
Prepare prompt text, reference audio, and optionally a reference image.
Treat the packaged App V3 as audio-driven by default. Uploaded audio is on the active path, and USE ONLY VOCALS is App-exposed.
For text-to-video, flip Text To Video (no image ref) to true, but still keep an image available because the notebook explicitly warns that loading an image helps avoid errors.
Keep first smoke runs short, with moderate size and a fixed seed.
Validate the produced MP4, audio, lip-sync, and actual output size or duration before claiming the workflow is working.

Prompt Rules

Describe motion and sound chronologically.
Do not restate details already fixed by the reference image unless the scene is intentionally changing.
Do not invent camera motion unless the user asks for it.
Do not invent dialogue unless the user asks for speaking, singing, or conversation.
Keep the prompt as one concise English paragraph if you are targeting the built-in prompt-enhancer path.

Parameter Rules

Distinguish extra.linearData App inputs from graph-only defaults.
The packaged App V3 exposes prompt, image, audio, audio start trim, width, height, length, text-to-video mode, USE ONLY VOCALS, seed, and an Enhanced Prompt surface. It does not expose App-side FPS control or the internal custom-audio versus LTX-audio latent switch.
length_seconds is converted into frame count with 1 + 8 * round(seconds * fps / 8). Expect internal rounding.
The custom-audio trim duration is derived from frames / fps, so audio segment length follows the generated clip length.
Upstream notes warn that invalid size or frame settings can be silently nudged to nearby valid values. In practice, the current defaults and resizers point toward clean multiples of 32 for width and height. Recommend clean multiples of 32 and verify the actual output size after generation.
fps exists in the graph with default 24, but it is not one of the App inputs exposed by this JSON. Do not promise an App-side FPS control unless you are editing the graph.
USE ONLY VOCALS is useful when the reference audio has busy accompaniment and lip-sync quality matters more than preserving the full mix.

Audio Path Rules

The graph loads audio once, trims a segment downstream, optionally separates vocals with MelBandRoFormer, then encodes audio latent.
Preserve that pattern if you adapt the workflow. Do not duplicate audio loads per segment.
The graph contains both custom-audio and LTX-audio latent paths, but the packaged App V3 does not expose the custom-audio switch. If a user asks for generated LTX audio instead of uploaded audio, say that workflow editing is required.
If a user-made LoRA introduces noisy audio, the upstream note recommends the KJNodes advanced LTX-2 LoRA loader with non-video strength set to zero.

Validation

Setup validation: confirm the setup script completed and torch.cuda.is_available() is true on the remote host.
Service validation: confirm ComfyUI started and the configured port is open.
Static validation: confirm the JSON imports and all custom nodes resolve.
Asset validation: confirm image, audio, and model files exist in the expected folders.
Smoke run: 3-5 seconds, modest resolution, fixed seed.
Output validation: inspect video, audio, lip-sync, and any silent parameter rounding.
Only call it "working" after at least one real generation run or after clearly labeling the result as unverified.

Repository References

Upstream workflow: sources/upstream/isi-dev/LTX_2.3_Image_or_Text_&_Audio_2_Video_App_V3.json
Upstream notebook: sources/upstream/isi-dev/ComfyUI_LTX2_3_TI2V.ipynb

ltx23-comfyui-ti2v-audio

LTX2.3 ComfyUI TI2V + Audio

Start Here