This repo holds the 🌍 General Physics checkpoint of the Kandinsky WM 1.0 family — DiT soar, published here in two layouts: a diffusers pipeline (model_index.json, transformer/, vae/, text_encoder*/) and the GitHub-code DiT at model/soar.safetensors (same weights). The other two variants are in the Kandinsky WM 1.0 collection.
Kandinsky WM (World Model) 1.0 is a family of Image-to-Video models that adapt
Kandinsky 5.0 Video Lite — a 2B-parameter
latent video diffusion model (a Kandinsky5Transformer3DModel DiT paired with a HunyuanVideo VAE
and Qwen2.5-VL + CLIP text encoders, trained with flow matching) — to Physical AI: video
generation that is not only visually convincing but physically plausible — consistent scene
geometry, object dynamics, interactions and cause-and-effect — so the model can serve as a source
of synthetic training data and as a building block for world models and simulators.
The base model was domain-adapted on large corpora of 5-second scenes (first text-to-video, then a
mixed text/image-to-video regime that preserves first-frame continuation), spanning autonomous
driving, robotics, and a general domain (industrial processes, physical phenomena, and
human–object / human–human interaction). Each checkpoint is then reinforcement-learning post-trained with GRPO
against a reward model that scores the physical plausibility of the generated scene, steering the
generator toward more faithful geometry, object behaviour and interactions than domain fine-tuning
alone.
Each checkpoint generates 5-second, 121-frame clips at 768×512.
Model Zoo
Three domain checkpoints, same architecture — only the DiT weights differ (the VAE / text encoders are
byte-identical across all three). Each is published as a one-line diffusers pipeline; a
GitHub-code (DiT-only) copy lives in the same Hub repo used purely as file storage.
Four Kandinsky WM 1.0 image-to-video generations per domain (top-ranked by an internal visual
review). Each grid shows the first frame of a clip — click it to play the generated video
(hosted on the static_videos
dataset). The first frame and prompt for every clip also live in assets/: under
assets/<domain>/, frame_N.jpg and prompt_N.txt correspond to grid cell N (left→right, top→bottom).
Kandinsky WM 1.0 on three physical-AI video benchmarks (our row in bold). RBench and
Physics-IQ were run with Qwen3-VL prompt enhancers — noted above each table, with the scripts under
prompt_enhancers/. Sizes are total parameter
counts where publicly disclosed (MoE models note active params; proprietary/undisclosed left as —).
PAI-Bench-G (Physical AI Bench — Generation)
Leaderboard. Column
abbreviations follow the PAI-Bench-G dimensions — see the leaderboard for exact definitions.
Run without prompt enhancement.
Rank
Model
Size
Overall
Domain
Quality
SC
BC
MS
AQ
IQ
OC
IS
IB
CS
AV
RO
IN
PH
HU
1
Cosmos3-Super
64B
83.9
89.5
78.2
92.7
94.1
99.2
52.7
70.8
20.5
97.7
98.1
94.4
77.5
90.0
90.7
95.0
87.6
2
Cosmos3-Nano
16B
83.7
89.4
78.1
92.3
93.8
99.2
52.7
70.1
20.4
97.9
98.3
95.0
75.4
90.2
89.7
94.5
88.0
3
Veo-3
—
82.1
86.7
77.6
91.4
93.1
99.2
51.9
69.8
21.7
97.0
96.9
94.4
68.7
86.9
89.7
91.6
84.4
4
Kandinsky WM 1.0
2B
81.7
86.0
77.4
91.8
94.1
99.0
53.2
64.9
21.6
97.2
97.7
93.6
72.6
82.1
88.0
90.7
86.0
5
k5 Lite FT
2B
81.4
85.8
77.1
91.3
93.9
98.7
52.7
64.3
21.7
96.6
97.3
92.9
72.6
82.1
87.7
88.8
86.4
6
Cosmos-Predict2.5-14B
14B
81.0
83.8
78.1
93.4
94.8
99.1
52.5
70.0
20.1
97.2
97.9
94.2
67.8
79.9
87.7
93.5
80.0
7
Cosmos-Predict2.5-2B
2B
81.0
84.0
77.9
92.5
94.2
99.1
52.4
70.8
20.1
96.6
97.4
94.1
66.1
80.8
87.8
93.9
81.4
8
Wan2.2-I2V-A14B
27B (14B active)
80.6
84.1
77.2
91.6
93.7
98.3
51.2
69.6
20.4
96.0
96.6
93.2
66.3
81.7
89.2
91.8
82.1
9
K5 Lite
2B
80.5
83.0
77.9
91.7
94.4
99.3
54.1
65.6
21.7
98.1
98.6
89.2
66.3
77.3
86.3
87.2
84.6
10
Wan2.2-TI2V-5B
5B
80.4
83.4
77.4
91.8
93.7
98.8
51.9
69.9
20.3
95.9
96.7
93.1
65.2
79.3
88.4
91.5
83.0
Physical-AI post-training lifts the base K5 Lite by +1.2 Overall (80.5 → 81.7) and
+3.0 on the Domain axis (83.0 → 86.0) — a 2B model landing between Veo-3 and the
Cosmos-Predict2.5 family.
RBench
Leaderboard, Qwen evaluator tab.
RBench evaluates robot-oriented image-to-video generation across five task categories and four
robot embodiments.
Two ways to run — the checkpoints ship in both formats. The GitHub-code path is primary; the
diffusers path is a one-liner that already works off the Hub.
