PACT: Preserving Accurate Continuous Trajectories via Flow-Map Distillation

PACT trains adapters on an offline bank of teacher trajectories to predict interval-average velocity, enabling few-step image and audiovisual generation while preserving teacher correspondence. This repository contains the two selected paper inference adapters, including their learned destination-time and guidance conditioners. Base models are separate downloads.

Adapter	Base model	Paper setting	Download size	Details
sd35/	Stable Diffusion 3.5 Medium, 2.5B	4 steps, 1024×1024; raw step 190,000	471 MB	Model card and usage
ltx/	LTX-2.3 22B dev, joint audio/video	8 steps, 768×512, 121 frames at 24 fps; raw step 30,000	15.0 GB	Model card and usage

Both use rank/alpha 384. Use the PACT implementation and load the complete adapter folder. Generic Diffusers or upstream LTX LoRA loaders omit required PACT conditioners and do not implement its flow-map sampling path. These packages contain inference weights, not optimizer state for exact training resumption.

Download

With the Hugging Face CLI installed, choose one or both:

hf download AhmadRH/PACT --include 'sd35/*' --local-dir ./pact-checkpoints
hf download AhmadRH/PACT --include 'ltx/*' --local-dir ./pact-checkpoints

Read the corresponding model card for base-model pins and inference commands. Source code publication on GitHub is pending; inference requires the accompanying PACT release source tree (SD35-FlowBank-Distillation/ or LTX-FlowBank-Distillation/). The weights alone are not a standalone pipeline. Public paper and code links will be added when available.

Validation and limitations

The release adapters were tested with the PACT loaders and representative few-step generations. Small preparation, trajectory-bank, 500-update training, checkpoint-reload and evaluation checks passed for both implementations. These are functional release checks, not a rerun of the paper benchmarks or a guarantee of quality for every prompt. No subjective audio-quality claim follows from the audio integrity checks. See each card for data provenance and reproducibility limitations.

Licenses and attribution

The two adapters have different governing agreements: SD3.5 Stability AI Community License and LTX-2 Community License. Read the license scope, SD3.5 Notice and LTX NOTICE before use. These notices include the authors' permission for their controlled adapter modifications under the corresponding agreement. This repository does not grant rights to the separate base models, text encoders or datasets. The LTX training sources declare noncommercial terms; no blanket commercial-data clearance is asserted.

Powered by Stability AI. SD3.5 is based on Stability AI's Stable Diffusion 3.5 Medium. The audiovisual adapter is a modified derivative of Lightricks' LTX-2.3, using a separate Google Gemma text encoder. These adapters are provided by the PACT authors; upstream attribution does not imply endorsement.