LUVE — 2K and 4K inference checkpoints

Inference checkpoints for LUVE: Latent-Cascaded Ultra-High-Resolution Video Generation with Dual Frequency Experts (ICML 2026), for the LUVE inference 0.3.0 runtime.

Paper · Code and installation

LUVE generates a low-resolution video latent, upsamples it with the trained video latent upsampler (VLU), and refines it with high-resolution content refinement (HCR). Each resolution has one model.safetensors containing the HCR backbone, both frequency experts, and VLU parameters. HCR uses the hcr. tensor namespace and VLU uses vlu..

The native model applies FFT frequency filtering to activations. It selects the low-frequency attention expert when the original scheduler timestep is t >= 875 and the high-frequency FFN expert when t < 875, on a 1000-step training-time scale. Routing occurs before the timestep is converted to BF16. The experts are included in the checkpoint and require no separate LoRA files.

Checkpoint contents

weights/
  2k/
    model.safetensors
    config.json
    vlu_config.yaml
    model_index.json
    weights_manifest.json
  4k/
    model.safetensors
    config.json
    vlu_config.yaml
    model_index.json
    weights_manifest.json

Default	2K	4K
Output height × width	1472 × 2560	2176 × 3840
Low-resolution height × width	720 × 1280	720 × 1280
Frames / FPS	49 / 15	49 / 15
Low-resolution generation steps	50	50
HCR scheduler steps / skipped prefix	50 / 10	50 / 10
CFG / seed	5 / 0	5 / 0

The original-format Wan2.1-T2V-1.3B components are required separately: the low-resolution generator, T5 text encoder, tokenizer and VAE. They are not included in these LUVE checkpoints. Generation loads local files; it does not download components automatically.

Download and generate

Install the inference code following the GitHub instructions. The runtime targets Linux, Python 3.10 and a CUDA GPU supporting BF16; required GPU memory depends on the resolution, frame count and attention implementation.

Download one LUVE profile and the required Wan components:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="zhihefang/LUVE",
    local_dir="./models/LUVE",
    allow_patterns=["weights/2k/**"],
)
snapshot_download(
    repo_id="Wan-AI/Wan2.1-T2V-1.3B",
    local_dir="./models/Wan2.1-T2V-1.3B",
    allow_patterns=[
        "diffusion_pytorch_model.safetensors",
        "models_t5_umt5-xxl-enc-bf16.pth",
        "Wan2.1_VAE.pth",
        "google/umt5-xxl/**",
    ],
)

luve-generate \
  --model ./models/LUVE/weights/2k \
  --wan-model-dir ./models/Wan2.1-T2V-1.3B \
  --prompt "A cinematic landscape with mountains and a slow-moving river." \
  --output ./outputs/luve-2k.mp4 \
  --gpu 0

For 4K, download weights/4k/** and select ./models/LUVE/weights/4k with --model. The command saves one MP4. luve-generate --help lists generation controls. Add --check-only --verify-weights to validate local component paths and the LUVE checkpoint hashes without generating a video.

Validation scope

Release checks cover checkpoint loading and numerical consistency:

• A prior 2K, 49-frame comparison using the same cached low-resolution latent matched the separate-expert FFT implementation and the integrated native implementation bit for bit.
• The current release passed 27 CPU checks. Actual checkpoint predictions matched bit for bit at both the low-frequency and high-frequency stages for both profiles.
• A 4K check used the full 49-frame latent shape and verified finite output for one denoising step. A complete 4K video generation run has not been verified for this release.

These are implementation checks, not video-quality benchmarks. They do not establish equivalence to a non-FFT implementation or performance across prompts, seeds, hardware and attention kernels. This release does not provide a new training-data description or new quality-evaluation results; refer to the paper for the research method and its reported experiments.