| # Human/VLM-evaluated step-80 generator: four trained stages + terminal t=0. | |
| # The auxiliary seed controls context-cache noise, independently of the video seed. | |
| attention_type: softmax | |
| causal: true | |
| denoising_step_list: | |
| - 1000 | |
| - 750 | |
| - 500 | |
| - 250 | |
| warp_denoising_step: true | |
| num_train_timestep: 1000 | |
| timestep_shift: 5.0 | |
| denoising_loss_type: flow | |
| mixed_precision: true | |
| independent_first_frame: false | |
| same_step_across_blocks: true | |
| context_noise: 0 | |
| num_frame_per_block: 3 | |
| latent_video_shape: | |
| - 1 | |
| - 21 | |
| - 16 | |
| - 60 | |
| - 104 | |
| guidance_scale: 1.0 | |
| negative_prompt: '' | |
| model_kwargs: | |
| model_name: Wan2.1-T2V-14B | |
| timestep_shift: 5.0 | |
| inference_terminal_zero_cleanup: true | |
| inference_train_matched_context_cache: true | |
| inference_aux_noise_seed: 12345 | |