---
language:
- en
license: other
license_name: ltx-2-community-license
license_link: https://huggingface.co/AhmadRH/PACT/blob/main/ltx/LICENSE
base_model: Lightricks/LTX-2.3
base_model_relation: adapter
library_name: pytorch
pipeline_tag: text-to-video
tags: [pact, flow-matching, distillation, lora, image-to-video, audio-video]
---

# PACT for LTX-2.3 audiovisual generation

Inference adapter for **PACT: Preserving Accurate Continuous Trajectories via Flow-Map Distillation**, released by the PACT authors. This is the **raw step-30,000**, rank/alpha-384 joint audiovisual checkpoint, trained with **equal video and audio loss weights**. It is a modified derivative of Lightricks' LTX-2.3 22B dev model, not an official Lightricks release.

## Base and contents

Download separately:

- [`Lightricks/LTX-2.3`](https://huggingface.co/Lightricks/LTX-2.3), file `ltx-2.3-22b-dev.safetensors`, revision `3c6a4e66e5d0a684231950b9c74dd4ded7b6fadc`.
- [`google/gemma-3-12b-it`](https://huggingface.co/google/gemma-3-12b-it), verified release revision `96b6f1eccf38110c56df3a15bffe176da04bfd80`. Accept Google's access and license terms. The original historical Gemma download was unpinned; this pin identifies the tested release snapshot.

`lora_weights_step_30000.safetensors` contains AV/cross-modal LoRA plus **both** destination/guidance conditioners. `adapter.yaml` supplies construction metadata; `inference.json` records the exact sigma schedule, base/adapter hashes and sanitized-export receipt. The tensor payload is unchanged from the selected paper checkpoint. `SHA256SUMS` covers this complete folder. No optimizer or full training-resume state is included.

The paper setting uses **eight joint evaluations, learned video CFG 3, no external guidance, 768×512, 121 frames, 24 fps, and 48 kHz stereo audio**. The adapter supports text-to-video and first-frame image-to-video. Generic upstream LoRA inference omits PACT conditioning and is incompatible. The approximately 15 GB FP32 adapter requires substantial GPU memory in addition to the base model; release validation used GH200 GPUs.

## Usage

Install all three packages from the accompanying PACT `LTX-FlowBank-Distillation/` source tree following its README (source publication is pending). Use its CUDA environment and FFmpeg; keep it separate from the SD3.5 environment. From the LTX source directory, after downloading the base and Gemma:

```bash
export PACT_DATA=/absolute/path/to/pact-ltx-data
hf download AhmadRH/PACT --include 'ltx/*' --local-dir "$PACT_DATA/checkpoints"
python scripts/configure.py \
  --data-root "$PACT_DATA" \
  --base "$PACT_DATA/models/ltx-2.3/ltx-2.3-22b-dev.safetensors" \
  --gemma "$PACT_DATA/models/gemma-3-12b" \
  --adapter "$PACT_DATA/checkpoints/ltx/lora_weights_step_30000.safetensors" \
  --adapter-config "$PACT_DATA/checkpoints/ltx/adapter.yaml" \
  --output "$PACT_DATA/configs"
cat > "$PACT_DATA/prompts.jsonl" <<'JSONL'
{"prompt":"A red fox walks through a snowy pine forest. The camera tracks smoothly beside it as soft snow crunches under its paws and a quiet wind blows through the trees.","seed":42}
JSONL
python scripts/generate.py --config "$PACT_DATA/configs/vbench_student.yaml" \
  --prompts "$PACT_DATA/prompts.jsonl" --output "$PACT_DATA/samples" --cache-only
python scripts/generate.py --config "$PACT_DATA/configs/vbench_student.yaml" \
  --prompts "$PACT_DATA/prompts.jsonl" --output "$PACT_DATA/samples"
```

For I2V, add an `"image":"/absolute/path/reference.png"` field to a prompt record. Generation creates MP4s with audio and JSON receipts. The separate cache step releases text-encoder memory before loading the audiovisual transformer. Training is unnecessary.

## Training data and method

The paper uses 25,000 [VidProM](https://huggingface.co/datasets/WenhaoWang/VidProM) text prompts and 25,000 [TIP-I2V](https://huggingface.co/datasets/WenhaoWang/TIP-I2V) image/prompt cases, with 10% of each task held out. Preparation filters, embeds, clusters, selects and enhances conditions using Gemma. The teacher stores 30-step joint trajectories; PACT learns average velocity with separate masked video/audio losses of equal weight. Training data and Gemma/base weights are not redistributed here.

Both dataset cards declare **CC-BY-NC-4.0**. See [NOTICE](https://huggingface.co/AhmadRH/PACT/blob/main/ltx/NOTICE) for source attribution and preparation changes. This release is intended for noncommercial research and evaluation and does not assert clearance for commercial use of its training sources. The datasets' terms are distinct from the model agreement; training on them does not by itself establish that every output or adapter inherits a Creative Commons license.

## Evaluation and limitations

Release checks include paper-adapter T2V/I2V generation and visual review, audiovisual integrity checks, and bounded evaluation execution. A separate fresh 500-update smoke test reduced total loss from 0.548355 to 0.060541; tiny held-out errors increased, so this is not evidence of generalization. Visual VBench metrics do not measure audio fidelity or audiovisual synchronization; no subjective listening evaluation is claimed. Full paper benchmarks were not rerun during packaging.

Generated motion, fine details and composition may differ from the teacher, particularly for difficult image-conditioned dynamics. The adapter inherits its base model's limitations and biases. Preserve the complete AV conditioners, image masks, paired sigma schedule and learned guidance when using it.

## License

The adapter is distributed under the included **[LTX-2 Community License Agreement](https://huggingface.co/AhmadRH/PACT/blob/main/ltx/LICENSE), dated January 5, 2026**, exactly as present in the pinned base-model repository. Its full Attachment A, acceptable-use requirements, commercial restrictions and redistribution obligations apply. [NOTICE](https://huggingface.co/AhmadRH/PACT/blob/main/ltx/NOTICE) contains prominent modification and upstream attribution notices, plus the PACT authors' permission for their controlled adapter modifications subject to that agreement. The separate Gemma encoder is governed by [Gemma terms](https://ai.google.dev/gemma/terms), and dataset licenses remain applicable. This release is not a blanket commercial-use clearance.