LTX2.3 Audio Reactive LoRA — LTX Video Models¶

Development line: project:ltx2-3-audio-reactive-lora · thread ltx-video-models
Last event: 2026-06-13 · 1 dated since 2026-06-13 · Researched: 2026-09-05 · confidence: medium

What it is¶

LTX2.3 Audio Reactive LoRA is a LoRA adapter in the fal repository for Lightricks/LTX-2.3, not a standalone base model, built for music visualizer creators.

• Sound response: amplifies sound-driven pulse, particles, light, deformations, and camera movement.
• Pipeline modes: operates in audio-to-video and image+audio-to-video with an audio track, a prompt, and an optional initial frame.

The model card recommends scale 1.0–1.5, 24 fps, and clips of 5–15 s; exact beat detection is not guaranteed. The adapter fits short visualizers with clear geometry, but it does not replace frame-accurate synchronization.

Development line¶

• 2026-06-13 — LTX2.3 Audio Reactive LoRA appeared on Hugging Face under fal. On 2026-06-13, the LTX2.3 Audio Reactive LoRA line was cataloged with a direct Hugging Face link under fal. This provides a public reference point for the line, but available evidence does not establish model capabilities, version details, training data, or usage terms.

What changed¶

2026-06-12 — card metadata changed private: true to private: false; this marks card visibility rather than an independent runtime check. 2026-06-13 — primary history shows no separate dated revision for V1, V2, or the API on this date; the link led to the already visible adapter. 2026-06-15 — a separate V2 file was added; the author's card claimed stronger audio reactivity without providing an independent benchmark.

How to use this¶

From 2026-06-13, treat this as a distinct fal-linked LTX video-model line and inspect its Hugging Face model card before adoption; public sources provide no validated configuration or performance guidance.

1. Use LTX-2.3 instead of replacing the base with LTX-2.5: official documentation states that files are incompatible and the LoRA works only with the model on which it was trained. — https://github.com/Lightricks/LTX-2
2. Send prompt, mandatory audio_url, and loras in the current fal API; image_url is optional, and the endpoint allows up to three LoRAs simultaneously. — https://fal.ai/models/fal-ai/ltx-2.3-quality/audio-to-video/lora/api
3. Add V1 or V2 safetensors from the model card to loras, set transformer: "both", and start within the claimed scale range of 1.0–1.5. — https://huggingface.co/fal/ltx2.3-audio-reactive-lora
4. Pass image_url as the initial frame for a more directed clip; keep match_audio_length enabled if video duration must follow audio length. — https://fal.ai/models/fal-ai/ltx-2.3-quality/audio-to-video/lora/api
5. Review the result for text legibility, logo fidelity, and synchronization before publishing. — https://huggingface.co/fal/ltx2.3-audio-reactive-lora

Best practices¶

• Do not mix the adapter with another LTX release line: official documentation states that the LoRA works only with the model on which it was trained. — https://github.com/Lightricks/LTX-2
• Give the initial frame structures that can move: geometry, layers, particles, light seams, or visualizer shapes. — https://huggingface.co/fal/ltx2.3-audio-reactive-lora
• Treat clean, controlled visual direction as a practical defense against chaotic artifacts: this reflects a single user run, not a general benchmark. — https://www.reddit.com/r/StableDiffusion/comments/1un6vh3/followup_followup_more_experimentation_with_the/
• Generate several variations for a complex track and pick the best one; one user test reported better results on sparser arrangements when choosing among roughly three renders. — https://www.reddit.com/r/StableDiffusion/comments/1ulq1jk/followup_i_take_it_back_the_ltx_23_audioreactive/

Superseded by this¶

• 2026-06-12 — The early card state private: true is obsolete: a commit changed the visibility flag to private: false; this supersedes only the card state.
• 2026-06-15 — The note that only one adapter file exists is obsolete: V2 was added as a separate file; replacing V1 with V2 by default remains unproven.

Still unknown¶

• No external primary source dated 2026-06-13 adds facts to the step itself; event_findings remains empty.
• Training data, rank/alpha, target modules, and a reproducible training recipe for this specific adapter are not published.
• No independent benchmark compares V1 against V2; the claim of stronger audio reactivity comes from the author's card.
• Local execution, weight downloads, and fal API calls were not tested in this pass; current documentation does not serve as runtime verification.
• The exact applicability of the LTX-2 Community License to specific use cases has not been analyzed.
• Chinese-language searches yielded no independent primary or practical sources for this exact adapter.

Sources¶

Agent brief¶

• Subject: project:ltx2-3-audio-reactive-lora, thread ltx-video-models, 1 dated events 2026-06-13 → 2026-06-13.
• Practical note: From 2026-06-13, treat this as a distinct fal-linked LTX video-model line and inspect its Hugging Face model card before adoption; public sources provide no validated configuration or performance guidance.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.