The Open-Sora team (Peng, Zheng et al.), HPC-AI Tech (2025) — ARXIV-2503.09642
Key takeaways
-
The cost accounting is explicit, and narrow. Table 2 covers “a single full training run, assuming the rental price of H200 is $2 per GPU hour”:
- 256px text-to-video: 2240 GPU-days, $107.5K
- 256px text/image-to-video: 384 GPU-days, $18.4K
- 768px text/image-to-video: 1536 GPU-days, $73.7K
- total: 4160 GPU-days, $199.6K
This is the only report in the line that states its training cost at all.
-
It is initialized from a distilled image model. Training starts from FLUX. The authors “empirically find that, despite it being a distilled model, this initialization is effective” (§4.1.1). No comparison is shown.
-
The objective is SD3’s, stretched to video. §4.2.1 calls its flow matching “similar to that used in Stable Diffusion 3” (LIT-449): velocity regression along the straight noise-to-data path, with logit-normal timesteps, then a shift t ← αt/(1+(α−1)t) of the same form as SD3’s resolution shift. The change is that α is proportional to T×H×W, so the shift grows with duration as well as resolution. That extension is asserted, not tested. The MMDiT-style dual- and single-stream blocks come by way of FLUX.
-
Images are in every stage, as a buried detail. Appendix D gives image buckets at 256, 768 and 1024px in stages 1–2 and 768px images in stage 3 (Tables 4–5). In v3 the main text never says so. In v1 the same tables are main-text Tables 5–6 (§4.2).
-
High resolution is reached through image-to-video. “We find that adapting” from 256px to 768px “is significantly more efficient” as image-to-video than as text-to-video (§4.1.4). It is stated as a finding, with no numbers. The hypothesis is the explanation: that conditioning lets the model “focus more on motion generation”. The 768px stage is T/I2V with 12.5% image-condition dropout (§5.1). It is short in iterations (13K) but not in compute: 1,536 GPU-days, 37% of the total. At inference the pipeline is text to image (FLUX) to video.
-
Video DC-AE. A non-causal 4×32×32 autoencoder cuts tokens from 76K to 19K and raises training throughput 5.2× (Fig. 4). Its reconstruction matches HunyuanVideo’s VAE, with LPIPS 0.049–0.051 against 0.046 (Table 1). It is a further experiment (§4.3), not the $200k run, and its speedups and adaptation cost are outside Table 2.
Where the hedges are
Per DP-010:
-
“$200k” (title) leaves out several costs:
- HunyuanVideo’s VAE, which the run uses and did not train
- data filtering and captioning, partly with a proprietary model
- failed runs and hyperparameter search
- FLUX’s pretraining, which the initialization inherits
The comparison figure labels GPU-hours “in H100” for a run on H200, and the Movie Gen and Step-Video costs it compares against are estimates.
-
**The DC-AE speedups are not in the 200k run is the HunyuanVideo-VAE model (patch 2, CP = 4, 76K-token buckets; Tables 2, 3 and 5), so the figure does not rest on the DC-AE. The DC-AE model was “initially trained on HunyuanVideo’s VAE and later adapted to our Video DC-AE” (§3). Its 5.2× and >10× speedups and its adaptation cost fall outside Table 2. It “does not fully converge… underperforms the original” (§4.3.3), shown only qualitatively.
-
“Comparable to HunyuanVideo” (abstract) comes from a text-to-image-to- video pipeline, with FLUX supplying the first frame, evaluated against text-to-video baselines (Fig. 10 caption).
-
Asserted, not ablated: FLUX initialization, low-resolution motion learning, image-to-video-first adaptation, and “>99%” GPU utilization.
Standing in the anthology
The record keeps this report for its cost accounting, which no other video report gives. The honest reading of the figure is the price of the last run, not the price of the model. That distinction is worth writing down because “$200k” is the sentence that travels.
On SOTA-386 it is adoption in both halves: image initialization, and images in every stage. It is one more report that does the practice without testing it. The initialization from a distilled image model is new, and untested.
Its comparisons with HunyuanVideo (LIT-620) and Step-Video (LIT-624) run on three fronts. In Table 1, its own evaluation of the autoencoders, it puts HunyuanVideo’s VAE first on LPIPS (0.046) and Step-Video’s 8×16×16 VAE well behind (0.082, PSNR 28.7 against 30.2), contrary to Step-Video’s own Table 15. In the human evaluation (100 prompts, 10 evaluators, Fig. 1) it claims to win at least two of three aspects against each of six models, both of these among them, and its VBench score is above HunyuanVideo’s (Fig. 10); both use the FLUX-first text-to-image-to-video pipeline against text-to-video baselines. In the cost comparison (Fig. 6), Step-Video’s 500k GPU-hours against its 100k is an estimate from public information.
Read in full as NOTE-345.