LTX-2.3-Foley-LoRA — LTX Video¶

Development line: project:ltx-foley-lora · thread ltx-video
Last event: 2026-07-01 · 1 dated since 2026-07-01 · Researched: 2026-09-05 · confidence: medium

What it is¶

LTX-2.3-Foley-LoRA is a 400-step video-to-audio LoRA for LTX-2.3 base or distilled models. It takes a silent video and an action prompt to generate synchronised, non-speech Foley. It is not a general music or dialogue model. Use a LoRA multiplier of 1–3. It fixes LTX-2.3 generations that add background score, but speech behaviour remains insufficiently documented.

Development line¶

• 2026-07-01 — LTX-2.3-Foley-LoRA model page was recorded. On 2026-07-01, FuzzPuppy published the Hugging Face model page for LTX-2.3-Foley-LoRA. The page provides the Foley LoRA artifact, but does not detail its capabilities, provenance, compatibility, or release details.

What changed¶

• 2026-07-01 — LTX-2.3-Foley-LoRA was released as a 400-step adapter to suppress music overlays and generate visually matched sound effects. Its author trained it without speech examples and did not test speech.
• 2026-07-04 — The companion ComfyUI repository added a long-video sliding-window workflow with overlap and audio stitching, extending the original short-clip workflow.

How to use this¶

From 2026-07-01, evaluate and verify LTX-2.3-Foley-LoRA on Hugging Face before relying on it in production; this line alone does not support operational claims.

1. Download ltx-2.3-foley-400-steps.safetensors, load it with an LTX-2.3 base or distilled video-to-audio model, and start at LoRA strength 1.0. — https://huggingface.co/FuzzPuppy/LTX-2.3-Foley-LoRA
2. For ComfyUI, install the official LTXVideo nodes and the workflow repository's ltx_foley_v2a helper nodes; put the LoRA in ComfyUI/models/loras. — https://huggingface.co/FuzzPuppy/LTX-2.3-Foley-Workflow
3. Use the short workflow for a quick clip; use foley-sliding-window.json for longer video so windows are overlapped and stitched. — https://huggingface.co/FuzzPuppy/LTX-2.3-Foley-Workflow

Best practices¶

• Describe only the visible action, then add “No speech is present. No music is present”; use an anti-music negative prompt. — https://huggingface.co/FuzzPuppy/LTX-2.3-Foley-LoRA
• Increase LoRA strength from 1 toward 2 or 3 only if score remains in the output. — https://huggingface.co/FuzzPuppy/LTX-2.3-Foley-LoRA
• Start ComfyUI with --cache-classic for the sliding-window workflow to avoid large LTX node outputs being evicted and recomputed between windows. — https://huggingface.co/FuzzPuppy/LTX-2.3-Foley-Workflow

Superseded by this¶

• 2026-07-04 — Treating the original short-clip workflow as the only supported route is obsolete for long videos; use the sliding-window workflow instead.

Still unknown¶

• The model card does not document training data, evaluation metrics, exact VRAM requirements, or reliable dialogue preservation.
• The 2026-07-04 workflow update appears in repository history and long-video documentation, but lacks an immutable release note.

Sources¶

Agent brief¶

• Subject: project:ltx-foley-lora, thread ltx-video, 1 dated events 2026-07-01 → 2026-07-01.
• Practical note: From 2026-07-01, practitioners should treat LTX-2.3-Foley-LoRA as a separately linked Hugging Face Foley-LoRA artifact to evaluate and verify before relying on it; this line alone does not support operational claims.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.