Pipeline config of the Kandinsky6 video super-resolution (VSR) pipeline.
Video -> video with no text encoder (the SR DiT is text-free), so the text-encoder lists are empty. The bundle's scheduler component drives the denoising loop and the transformer config's sr_params hold the noise level and RoPE scale; per-request knobs live in Kandinsky6SROptions and travel through ForwardBatch.extra.
Classes¶
dataclass ¶
Kandinsky6SRPipelineConfig(model_path: str = '', pipeline_config_path: str | None = None, embedded_cfg_scale: float = 6.0, flow_shift: float | None = None, flow_shift_sr: float | None = None, disable_autocast: bool = False, scheduler_step_in_fp32: bool = False, is_causal: bool = False, dit_config: DiTConfig = Kandinsky6SRConfig(), dit_precision: str = 'bf16', upsampler_config: UpsamplerConfig = Kandinsky6SRLatentUpscalerConfig(), upsampler_precision: str = 'bf16', vae_config: VAEConfig = Kandinsky6SRVAEConfig(), vae_precision: str = 'bf16', vae_decode_precision: str | None = None, vae_tiling: bool = False, vae_sp: bool = False, image_encoder_config: EncoderConfig = EncoderConfig(), image_encoder_precision: str = 'fp32', image_encoder_configs: tuple[EncoderConfig, ...] | None = None, image_encoder_precisions: tuple[str, ...] | None = None, text_encoder_configs: tuple = tuple(), text_encoder_precisions: tuple = tuple(), preprocess_text_funcs: tuple = tuple(), postprocess_text_funcs: tuple = tuple(), dmd_denoising_steps: list[int] | None = None, ti2v_task: bool = False, lucy_edit_task: bool = False, boundary_ratio: float | None = None)
Bases: PipelineConfig
Kandinsky6 SR: source video -> KVAE encode -> latent upscaler -> tiled SR DiT -> KVAE decode -> stitch.