RoboRender teacher video model (LoRA)
The 50-step teacher used to render
Ravenh97/roborender_anchors.
| file | step-5294.safetensors |
| bytes | 699,995,040 |
| md5 | 20e940246337b3a73156d647cef1a82d |
| base model | PAI/Wan2.1-Fun-V1.1-1.3B-Control (pull separately) |
| sampling | 50 steps, CFG 5.0 |
| throughput | ~32 s per 81-frame 3-view clip on an RTX PRO 6000 |
This is a LoRA only — it needs the Wan foundation weights above. It is a 3-view dual-control (depth + robot mask) model with previous-frame conditioning, trained on mixed DROID/AgiBot data.
Loaded by the render server as:
run_raven_server.sh --host 127.0.0.1 --port <port> --teacher_lora step-5294.safetensors
TensorRT is disabled automatically for the teacher: the shipped engines are exported from the 5-step student's fused weights and are not valid for this adapter, so it runs the PyTorch DiT path.
If you only need anchors for FT-FIXED conditioning, use the anchor dataset instead — it is 2.1 GB against the ~20 GB of foundation weights this model requires.