Known operation. New example.
Main evaluation
The task type appears in project fine-tuning. Evaluation asks the model to perform that operation on separate samples.
101 / 170 tasks passedPublished Mini examples; these clips are separate from the final manuscript evaluation batch.
Read the paper’s selective-editing, circuit, and failure cases →
Select a model to explore its nine capability families.
November 1 manuscript · Table 1
Evaluation protocol →03 / TRANSFERSame checkpoint. No further fine-tuning.
A task type held out of fine-tuning.
Additional evaluation
The operation is excluded from project fine-tuning. The final model attempts it with unchanged weights and inference settings.
16 / 170 planned tasks passedAppearance, identity, and action.
Where transfer appears
Reported examples include reference-guided colorization, wildlife identity masks, and action-conditioned continuation.
See the transfer evidence →166 tasks / 827 samples have evaluable outputs; 22 planned samples are unsupported. Held out of project fine-tuning does not mean absent from base-model pretraining.
340tasks. Click a family to open its tiles. Click a tile for Input / Target.
Model
Omni-Reason-1
13.125B video diffusion Transformer
one checkpoint
Condition
TV2V
Input video
Instruction
Optional references
Embeddings
Role
Slot (Position)
Alignment
Modality
How the condition reaches the target, per interface
TV2V and RV2V are separate interfaces, not two labels for the same task. Choose an interface or family to load a matching dataset example. Directed attention lets targets read conditions. One loss: conditional flow matching on target tokens only.
Families
* Project lead.
2026 research manuscript · PDF forthcoming.