Osc7/WaveForcing_Stage2_ODE_pair

WaveForcing Stage 2 — Wan2.1-T2V-14B ODE Endpoint Pairs

2,176 text-to-video endpoint pairs from Wan2.1-T2V-14B, the frozen-teacher target for the stage-2 distillation of WaveForcing. Each pair is one prompt, the initial Gaussian noise z_ref, and the latent y_ref the teacher produces from them in a 50-step UniPC run — the two ends of the sampling trajectory, without the 50 intermediate states.

Files

File	Content
pairs/000000.pt … pairs/002175.pt	one pair per file, pair_index in the filename
manifest_train.jsonl	train split, pair_index 0–2047 (2,048 pairs)
manifest_val.jsonl	held-out split, pair_index 2048–2175 (128 pairs)
integrity_check.json	verification record of all 2,176 files
SHA256SUMS	file-level hashes

Tensor format

z_ref and y_ref are both [21, 16, 60, 104], i.e. [T, C, H, W] without a batch dimension; z_ref is float32, y_ref is float16. Decoded with the matching Wan VAE they give 81 frames at 832×480, about 5.06 s at 16 FPS. Latents only — no pixels, no audio.

Load

from huggingface_hub import snapshot_download
import torch

snapshot_download(repo_id="Osc7/WaveForcing_Stage2_ODE_pair",
                  repo_type="dataset", local_dir="./wf_stage2_pairs")

item = torch.load("./wf_stage2_pairs/pairs/000000.pt",
                  map_location="cpu", weights_only=True)
z_ref, y_ref, prompt = item["z_ref"], item["y_ref"], item["prompt"]

Each .pt is a dict with version, pair_index, prompt_index, prompt, seed, z_ref, y_ref, z_ref_sha256, y_ref_sha256 (SHA256 of each tensor's raw bytes, not of the file), and the generation settings teacher_name, sampling_steps, guidance_scale, shift, solver, teacher_input_dtype.

Generation settings

Item	Setting
Teacher	Wan2.1-T2V-14B, bidirectional, bf16 input
Sampler	UniPC, 50 steps, CFG 3.0, timestep shift 5.0
Prompts	2,176 non-empty lines of the filtered and extended VidProM list, drawn with random.Random(1504386)
Noise	CPU torch.randn, float32, seed = 1504386 + pair_index

Checks

All 2,176 records were verified at generation time: metadata, tensor shapes, finiteness and tensor SHA256. Replaying the teacher on pair_index=2048 reproduced the stored tensors bit-exactly. On 2026-09-15 every file was re-verified on CPU — numbering, split sizes, manifest agreement, sampling settings, dtypes, finiteness, tensor hashes.

Scope

• One recipe: Wan2.1-T2V-14B at 832×480 and 81 frames with the settings above. Long video and other resolutions are outside the distribution.
• One noise sample per prompt: the pairs cover prompts, not each prompt's output distribution.
• Targets are the teacher's outputs as generated; pairs were not filtered by quality.

Links

• Code: SJTU-DENG-Lab/WaveForcing
• Weights: SJTU-DENG-Lab/WaveForcing-1.3B
• Project page: sjtu-deng-lab.github.io/WaveForcing
• Paper: 2611.00859