| # Public inference contract for the generator-only release. | |
| # This file intentionally does not reference the private training manifest. | |
| algorithm: | |
| generator_is_causal: false | |
| model_kwargs: | |
| model_name: Wan2.2-TI2V-5B | |
| timestep_shift: 5.0 | |
| num_frame_per_block: 8 | |
| local_attn_size: -1 | |
| controlnet_model_path: TheDenk/wan2.2-ti2v-5b-controlnet-depth-v1 | |
| controlnet_weight: 0.8 | |
| controlnet_stride: 3 | |
| checkpoints: | |
| lora_ckpt: generator_lora.pt | |
| inference: | |
| sampling_steps: 4 | |
| guidance_scale: 1.0 | |
| width: 832 | |
| height: 480 | |
| raw_frames: 93 | |
| latent_frames: 24 | |
| fps: 24 | |
| control_guidance_start: 0.0 | |
| control_guidance_end: 0.8 | |
| sink_size: 0 | |
| multi_shot_rope_offset: 0 | |
| adapter: | |
| type: lora | |
| rank: 64 | |
| alpha: 64 | |
| dropout: 0.0 | |
| notes: | |
| - The LoRA adapts the Wan transformer; it is not a ControlNet checkpoint. | |
| - The TheDenk depth ControlNet is a separate frozen runtime dependency. | |
| - Use a genuinely dynamic depth video; do not repeat one still depth map. | |
| - Other frame counts are temporal extrapolation from 93 raw / 24 latent frames. | |