language:
- en
license: other
license_name: stabilityai-community-license
license_link: https://huggingface.co/AhmadRH/PACT/blob/main/sd35/LICENSE
base_model: stabilityai/stable-diffusion-3.5-medium
base_model_relation: adapter
library_name: pytorch
pipeline_tag: text-to-image
tags:
- pact
- flow-matching
- distillation
- lora
PACT for Stable Diffusion 3.5 Medium
Powered by Stability AI. Inference adapter for PACT: Preserving Accurate Continuous Trajectories via Flow-Map Distillation, released by the PACT authors. This is the raw step-190,000 paper checkpoint, with rank/alpha 384 attention LoRA and the learned destination-time/guidance conditioner. It is a modified derivative of Stable Diffusion 3.5 Medium, not an official Stability AI release.
Base and contents
Download the base separately after accepting its access terms:
stabilityai/stable-diffusion-3.5-medium, revision b940f670f0eda2d07fbb75229e779da1ad11eb80.
flowmap_lora.safetensors and flowmap_conditioner.safetensors are both required. flowmap_manifest.json records architecture and tensor-file hashes; training_config.yaml preserves the portable final recipe; paper_schedules.json stores the exact schedules. All 390 trainable tensors passed bitwise zero/reload export verification, recorded in export_validation.json. SHA256SUMS covers the complete published folder.
The main setting is four student evaluations at 1024×1024, learned CFG 4.5, and no additional external CFG. Use the PACT loader; generic LoRA loading is insufficient.
Usage
Install the accompanying PACT SD35-FlowBank-Distillation/ source tree following its README (source publication is pending), and authenticate for the gated base with hf auth login. Training is unnecessary. From that source directory:
export SD35_DATA=/absolute/path/to/pact-sd35-data
hf download AhmadRH/PACT --include 'sd35/*' --local-dir "$SD35_DATA/pact-checkpoints"
export SD35_ADAPTER="$SD35_DATA/pact-checkpoints/sd35"
python scripts/infer.py \
--checkpoint "$SD35_ADAPTER" \
--prompt 'A wooden rowboat on a turquoise lake beneath snowy peaks.' \
--seed 42 --nfe 4 --output "$SD35_DATA/rowboat.png"
The source README documents CUDA/PyTorch and pinned Diffusers/Transformers dependencies. Use a separate environment for this image implementation and the LTX implementation. These adapters do not contain the base weights or optimizer/RNG/EMA state.
Training data and method
The final method uses raw prompts from DiffusionDB (CC0-1.0) and Pick-a-Pic v1 (no license field in the inspected card). Prompts are filtered, clustered and selected without enhancement. A frozen teacher constructs 28-step trajectories; PACT trains destination/guidance conditioning and rank-384 attention LoRA on interval-average velocity plus endpoint MSE. See Notice for pinned source attribution.
The exact earlier 50k prompt manifest was not preserved in the release. The code reconstructs the selection method without claiming byte-identical historical data. Training data are not redistributed here.
Evaluation and limitations
The raw paper adapter passed release inference and visual checks. A separate 500-update fresh-training smoke test reduced mean loss from 0.124115 to 0.013757. These small checks establish execution and loading; they do not reproduce the full paper's image-quality, diversity or teacher-similarity benchmarks. The code provides the paper evaluation entrypoints with their original metric backbones.
The adapter inherits base-model biases and limitations. Few-step outputs may differ from the teacher in fine detail, color and composition. Research and evaluation are the intended uses; users must comply with upstream restrictions and applicable rights in inputs/outputs.
License
The adapter is distributed under the included Stability AI Community License, dated July 5, 2024, with its incorporated Acceptable Use Policy. The license covers LoRA derivatives, requires its attribution and notices on redistribution, and imposes commercial registration/revenue conditions. Read the full terms; this release does not waive them or clear dataset rights. Notice includes the PACT authors' permission for the adapter modifications whose rights they control, subject to that agreement. The separate original PACT code license is not assigned by this model card.