LTX-2.5 — LTX Video¶

Development line: project:ltx-2-5 · thread ltx-video
Last event: 2026-08-11 · 1 dated since 2026-08-11 · Researched: 2026-09-07 · confidence: medium

What it is¶

LTX-2.5 is Lightricks' 22B audio-video model family for local ComfyUI and Python setups or hosted generation.

• Synchronized video and audio generation from text, image, video, and audio conditioning.
• Local tooling for video-to-video, LoRA/IC-LoRA, HDR, and fine-tuning paths.
• Hosted API access via ltx-2-5-fast and ltx-2-5-pro without retake, extend, or reframe.

The technical local guide requires an NVIDIA GPU with 32 GB+ VRAM, 32 GB RAM, 100 GB free storage, CUDA 12.7+, and Python 3.12+. The product page separately lists a 16 GB minimum. Use LTX-2.5 when synchronized sound, self-hosting, or local adaptation matter. Keep LTX-2.3 for a hosted workflow that needs retake, extend, or reframe.

Development line¶

• 2026-08-11 — LTX-2.5 official model and application links were shared. On 2026-08-11, a dated item in the LTX Video thread linked to the official LTX-2.5 model page and the LTX application. This marks a specific LTX-2.5 model and its official access path. The available evidence does not establish whether this was a launch, an update, or a change in availability.

What changed¶

• 2026-08-10 — ltx-2-5-fast gained automatic duration for text-to-video and image-to-video before the named release.
• 2026-08-11 — LTX introduced hosted Fast and Pro variants for text-, image-, and audio-to-video; LTX-2 v1.2.0 added the local 2.5 runtime and training support.
• 2026-08-12 — ComfyUI added day-one LTX-2.5 templates and both local-weight and hosted Partner Node routes.
• 2026-08-16 — The older ltx-2-fast and ltx-2-pro API IDs were removed, making migration to 2.3 or 2.5 mandatory.
• 2026-08-18 — Audio-to-video gained Fast 4K and Pro 1080p resolution tiers.
• 2026-08-19 — Audio-to-video gained frame-rate, last-frame, and camera-motion controls.
• 2026-08-25 — LTX-2 v1.3.0 changed local DFR with tiled 4K, multi-GPU execution, keyframe-aware decoding, and new DFR flags.
• 2026-09-02 — ltx-2-5-pro expanded to 1440p/4K and 48 fps across text-, image-, and audio-to-video.

How to use this¶

1. Choose the surface first. Use ltx-2-5-fast for speed and lower cost, or ltx-2-5-pro for higher fidelity in hosted text-to-video, image-to-video, or audio-to-video routes. Use a local path if you need self-hosting or adaptation. — https://docs.ltx.io/models/ltx-2-5
2. For a hosted image-to-video request, pass duration: null only when the end frame is not fixed. Automatic duration cannot be combined with last_frame_uri. — https://docs.ltx.io/models/ltx-2-5
3. For the visual local path, install ComfyUI-LTXVideo, select an LTX-2.5 template, download the listed matching files, add a prompt or references, and queue the job. — https://docs.ltx.io/open-source-model/getting-started/quick-start
4. For code or automation, clone LTX-2, run uv sync, obtain access to the model repository, download the split components, and provide each component path to ltx-pipelines. — https://docs.ltx.io/open-source-model/integration-tools/pytorch-api
5. For local generation, keep width and height divisible by 32. Use an 8n + 1 frame count, or omit the frame count to let the LTX-2.5 duration head choose it. — https://docs.ltx.io/open-source-model/integration-tools/pytorch-api

Best practices¶

• Keep the checkpoint components matched, especially the LTX-specific fine-tuned Gemma 4 text encoder. Do not substitute Google's vanilla Gemma 4. — https://github.com/Lightricks/LTX-2/blob/main/packages/ltx-trainer/docs/configuration-reference.md
• Start the ComfyUI text-to-video template at 768x512, 97 frames, and 24 fps. Increase length or resolution only after hardware handles the baseline, and fix the seed when you need reproduction. — https://docs.ltx.io/open-source-model/usage-guides/text-to-video
• Write prompts as a chronological scene. Name the shot, scene, action, character, camera movement, and audio. Rewrite prompts from other video models rather than pasting their tags or shot-list syntax unchanged. — https://docs.ltx.io/open-source-model/usage-guides/prompting-guide
• For native multi-shot work, use a chronological paragraph. Explicitly name every cut, restate recurring identifiers and audio continuity, and prefer two to four shots per generation. — https://docs.ltx.io/open-source-model/usage-guides/prompting-guide
• Use the native two-stage pipeline for local production. Use the distilled pipeline for speed or batch runs. — https://docs.ltx.io/open-source-model/integration-tools/pytorch-api
• Provision local work against the technical baseline: 32 GB+ NVIDIA VRAM, 32 GB RAM, 100 GB storage, CUDA 12.7+, and Python 3.12+. Validate the actual workflow on target hardware. — https://docs.ltx.io/open-source-model/getting-started/system-requirements

Superseded by this¶

• 2026-08-16 — The API model IDs ltx-2-fast and ltx-2-pro remain valid. Source: https://docs.ltx.io/api-changelog/2026/8/16
• 2026-08-18 — Audio-to-video on ltx-2-5-fast is limited to 1080p. Source: https://docs.ltx.io/api-changelog/2026/8/18
• 2026-08-19 — Audio-to-video cannot set frame rate, camera motion, or a last frame. Source: https://docs.ltx.io/api-changelog/2026/8/19
• 2026-09-02 — LTX-2.5 Pro is limited to 720p/1080p. Source: https://docs.ltx.io/api-changelog/2026/9/2
• 2026-08-25 — DFR uses --checkpoint-path, --distilled-lora, and --temporal-upsample-rounds. Source: https://github.com/Lightricks/LTX-2/blob/main/CHANGELOG.md

Still unknown¶

• A second unresolved LTX-2.5 entry dated 2026-08-12 uses provisional key ltx_2_5 and different Hugging Face URLs. It may cover a downstream artifact rather than this base release. Do not merge its history without a reviewed exact identity.
• Hosted API names fast/pro and local checkpoint names dev/distilled are separate documented surfaces. No first-party source found here declares a one-to-one mapping between them.
• First-party sources conflict on local VRAM: the product page says 16 GB minimum, while the technical system-requirements page says 32 GB+. No reconciliation was found; the latter is used for the practical baseline.
• The supplied hosted-app URL did not yield readable documentation in this research. Current UI controls, account flow, and pricing were not independently verified.
• The Simplified-Chinese search lane produced community pages but no verified, dated, reproducible LTX-2.5 practice source. No Chinese-language parameter claim is included.

Sources¶

Agent brief¶

• Subject: project:ltx-2-5, thread ltx-video, 1 dated events 2026-08-11 → 2026-08-11.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.