percool/causal-rcm-wan-eval

Causal-rCM Wan2.1 1.3B: generated videos and results report

This dataset holds every text-to-video sample from an independent re-evaluation of the 16 public 1.3B checkpoints in worstcoder/rcm-Wan. The checkpoints come from Causal-rCM (Zheng et al., arXiv:2606.25473, code NVlabs/rcm). It also includes the static web pages that show the samples:

• a video grid,
• a results report with VBench scores, bootstrap intervals and speed,
• a paper tutorial,
• a checkpoint map.

This evaluation is not affiliated with the paper's authors.

Viewing the pages

hf download percool/causal-rcm-wan-eval --repo-type dataset --local-dir causal-rcm-wan-eval   # about 45 GB
cd causal-rcm-wan-eval
python -m http.server 8100

Then open these pages:

• the video grid: http://localhost:8100/
• the results report: http://localhost:8100/mds/causal_rcm_report.html

The pages fetch data.json and load videos by relative URL, so serve them over HTTP. Opening the files directly (file://) does not work.

The report and the checkpoint map need only the pages (under 2 MB): --include "index.html" "data.json" "mds/*". The grid and the tutorial's side-by-side videos also need videos/. You can fetch a single configuration with, for example, --include "videos/sfdmd_dcm_c1_2step/*".

Layout

Path	Contents
index.html, data.json	Video grid: pick a prompt; rows are configurations, columns are seeds; per-video VBench scores
mds/causal_rcm_report.html, mds/causal_rcm_report_data.js	Results report: reproduction against the paper, leaderboard with 95% intervals, speed, all 16 VBench dimensions, pairwise tests, dynamics, open questions
mds/causal_rcm_tutorial.html	Tutorial on the paper, with interactive attention-mask and rollout explorers
mds/causal_rcm_pipeline.html	Map of the 16 checkpoints and the training stage each comes from
mds/causal_rcm_results.md	Short Markdown summary of the VBench results
videos/<configuration>/<prompt>-<seed>.mp4	70,048 generated videos

Videos

• Format: 832×480, 81 frames, 16 fps, H.264 (5 s).
• Prompts: the official sample_videos.py protocol in NVlabs/rcm, using VBench's augmented T2V prompts (944 prompts × seeds 0–4). Files are named <original VBench prompt>-<seed>.mp4, VBench's standard naming, so a folder can be scored by VBench directly.
• Teachers (50-step diffusion, about 100 network evaluations per chunk): seed 0 on every prompt, plus seeds 1–4 on the 12 prompts shown in the grid, which is 992 videos each.
• Batching: samples were generated five at a time on H200 GPUs, with each sample's noise seeded as in the official script. At batch size 1 the output is bit-identical to the official script. At batch size 5, GPU kernel selection changes the rounding, and the autoregressive rollout amplifies it: later frames differ from the official script about as much as a different seed would.

Folder	Configuration	Checkpoint (in worstcoder/rcm-Wan)	Sampling	NFE per later chunk	Videos	VBench Total (ours)	Paper
rcm_bidir_4step	rCM bidirectional	rCM_Wan2.1_T2V_1.3B_480p.pt	4 steps, whole clip	4	4,720	83.37	–
tf_diffusion_c1	TF diffusion teacher c1-1	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF_Diffusion_c1-1.pt	50-step Euler, CFG 3.0, shift 3.0	102	992	82.82 (seed 0)	–
df_diffusion_c1	DF diffusion teacher c1-1	Causal_rCM_Wan2.1_T2V_1.3B_480p_DF_Diffusion_c1-1.pt	50-step Euler, CFG 3.0, shift 3.0	102	992	82.72 (seed 0)	–
tf_diffusion_c3	TF diffusion teacher c3-3	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF_Diffusion_c3-3.pt	50-step Euler, CFG 3.0, shift 3.0	102	992	83.05 (seed 0)	–
df_diffusion_c3	DF diffusion teacher c3-3	Causal_rCM_Wan2.1_T2V_1.3B_480p_DF_Diffusion_c3-3.pt	50-step Euler, CFG 3.0, shift 3.0	102	992	83.36 (seed 0)	–
tf_dcm_c1	TF-dCM c1-1	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM_c1-1.pt	4 steps per chunk	5	4,720	80.96	–
tf_scm_c1	TF-sCM c1-1	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-sCM_c1-1.pt	4 steps per chunk	5	4,720	82.02	–
tf_dcm_c3	TF-dCM c3-3	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM_c3-3.pt	4 steps per chunk	5	4,720	81.99	–
tf_scm_c3	TF-sCM c3-3	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-sCM_c3-3.pt	4 steps per chunk	5	4,720	83.24	–
sfdmd_dcm_c1_4step	Causal-rCM c1-1 4-step (dCM init)	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c1-1_step4.pt	4 steps per chunk	5	4,720	84.23	84.29
sfdmd_dcm_c1_2step	Causal-rCM c1-1 2-step	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c1-1_step2.pt	chunk 0: 4 steps, later: 2 steps	3	4,720	84.65	84.63
sfdmd_dcm_c1_2step_noisyctx	Causal-rCM c1-1 2-step, noisy context	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c1-1_step2_noisy_ctx.pt	chunk 0: 4 steps, later: 2 steps, noisy context from chunk 1	2	4,720	83.12	83.11
sfdmd_dcm_c1_1step	Causal-rCM c1-1 1-step (step-2 checkpoint)	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c1-1_step2.pt	chunk 0: 4 steps, later: 1 step	2	4,720	84.77	84.63
sfdmd_scm_c3_4step	Causal-rCM c3-3 4-step (sCM init)	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-sCM-init_SF-DMD_c3-3_step4.pt	4 steps per chunk	5	4,720	84.53	84.37
sfdmd_dcm_c3_4step	Causal-rCM c3-3 4-step (dCM init)	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c3-3_step4.pt	4 steps per chunk	5	4,720	84.30	84.33
sfdmd_dcm_c3_2step	Causal-rCM c3-3 2-step	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c3-3_step2.pt	chunk 0: 4 steps, later: 2 steps	3	4,720	84.30	84.30
sfdmd_dcm_c3_2step_noisyctx	Causal-rCM c3-3 2-step, noisy context	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c3-3_step2_noisy_ctx.pt	chunk 0: 4 steps, later: 2 steps, noisy context from chunk 1	2	4,720	84.23	84.24
sfdmd_dcm_c3_1step	Causal-rCM c3-3 1-step (step-2 checkpoint)	Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c3-3_step2.pt	chunk 0: 4 steps, later: 1 step	2	4,720	84.10	84.01

Chunk patterns: c1-1 generates one latent frame per chunk (21 chunks); c3-3 generates three latent frames per chunk (7 chunks).

Scoring: VBench Total is VBench-T2V over all 16 dimensions, averaged over 5 seeds, except for the teachers (seed 0 only). The report gives per-dimension scores, intervals and pairwise tests.

Notes

• Licences: the videos were generated from Wan2.1-based checkpoints in worstcoder/rcm-Wan using VBench prompts. Follow the licences of those projects when you reuse them.
• Optical-flow study: report §9 reruns VBench's RAFT on every frame pair of the 72 dynamic-degree prompts. It covers motion over the clip, camera direction, chunk seams and appearance drift. The computing scripts are not part of this dataset.