Where open video models actually fall short
The gap between open weights and the closed frontier is mostly post-training and data curation, not architecture. Here is the evidence, and what it costs to close.
TL;DR — the quality gap between open-weight video models and the closed frontier is mostly a post-training and data-curation gap, not an architecture gap. The open literature already contains the missing pieces, and the measured gains from applying them are large. This note lays out the evidence and the parts that remain genuinely hard.
The claim, stated precisely
When people compare Wan 2.2 or LTX against Seedance or Veo and conclude that open models are “a generation behind”, the implied explanation is usually architectural — that the closed labs know some structural trick. The published record does not support that.
What the technical reports actually differ on is the stage after pre-training:
- Wan 2.1 and 2.2 describe their post-training as continued training on higher-quality data. Neither report discloses a reward model or any reinforcement learning from human feedback stage.
- Seedance 1.0 (arXiv:2506.09113) lists supervised fine-tuning plus multi-dimensional reward-model RLHF as a core pillar of its quality. It uses three separate reward models — a foundational one for prompt alignment and structural stability, a motion one that suppresses artefacts and rewards motion amplitude, and an aesthetic one operating on key frames. Preference labels are collected one dimension at a time, with the constraint that the chosen winner must not be worse than the loser on any other dimension.
That is a difference in method, not in model shape. And method transfers.
How much it is worth, measured
The strongest public number comes from DanceGRPO (arXiv:2505.07818), the first framework to make GRPO work across both diffusion and rectified-flow models. Applied to HunyuanVideo using the VideoAlign reward model:
| Dimension | Before | After | Change |
|---|---|---|---|
| Visual quality | 4.51 | 7.03 | +56% |
| Motion quality | 1.37 | 3.85 | +181% |
Motion is the axis where open models are most often described as “floaty” or “plasticky”, and it is the axis that moves most.
The second data point is a whole-recipe one. HunyuanVideo 1.5 — 8.3B parameters, with a complete post-training stack of SFT, DPO and MixGRPO — was evaluated against Seedance Pro in side-by-side human preference tests (GSB). At 720p text-to-video it came out +11.02% ahead. On image-to-video it was still 5.77% behind.
Read those two numbers together, because the pair is the actual finding: a serious post-training effort on an open base is enough to pass the previous generation of a closed flagship on text-to-video, and not yet enough on image-to-video.
What post-training does not fix
Two things stay hard, and it is worth being blunt about them.
Generational distance in pre-training. On blind human-preference Elo (Artificial Analysis, mid-2026 snapshot), Seedance 2.0 sat at 1222 for text-to-video while the strongest widely-available open weights of that moment, LTX-2.3, sat around 961–976 — roughly 250 Elo. No reward model closes 250 Elo. What post-training realistically buys is the distance to the previous flagship: Seedance 1.5 pro had by then fallen to 1000, and LTX-2.3 was within about 50 Elo of it on image-to-video.
Reward hacking. Optimising against a learned scorer teaches the model to please the scorer. DanceGRPO’s two stabilising tricks are worth copying directly: share the initial noise across samples of the same prompt, and train on a best-of-N selection — sample 16 trajectories, keep the top 8 and bottom 8.
The one that changed the picture
In August 2026 MiniMax H3 released open weights and posted a text-to-video Elo of 1240 — above Seedance 2.0 (1224), and within 4 Elo of the top of the entire board. The “open weights are 250 Elo behind” framing stopped being true at that moment.
It comes with real constraints: self-hosted deployment tops out at 768p, the training recipe is not disclosed, and the licence carries territorial exclusions and limits on how outputs may be used. We wrote those up separately in MiniMax H3 and the open-weight frontier.
Why we care
CFLabs serves open-weight video models commercially, through Phosor. Which base model we run, and whether it is worth post-training one ourselves rather than waiting for the next release, is a question we have to answer with numbers rather than vibes. Publishing the working is the cheapest way to find out we are wrong.
Sources
- Seedance 1.0 technical report — arXiv:2506.09113
- DanceGRPO — arXiv:2505.07818
- Flow-GRPO — arXiv:2505.05470
- VideoAlign / VideoReward — arXiv:2501.13918
- Working notes, source list and confidence tags: aoraki-labs-ai/Awesome-Video-AI-Research
Elo figures are a snapshot of a leaderboard that moves; treat them as of mid-2026 and re-check before relying on them.