Diffusers documentation
KandinskyS2VTransformer3DModel
KandinskyS2VTransformer3DModel
The Diffusion Transformer (DiT) of Kandinsky S2V 1.0, conditioned on a reference frame and a speech track.
KandinskyS2VTransformer3DModel
class diffusers.KandinskyS2VTransformer3DModel
< source >( in_visual_dim: int = 16out_visual_dim: int = 16in_text_dim: int = 3584in_text_dim2: int = 768in_audio_dim: int = 3584time_dim: int = 512patch_size: tuple[int, int, int] = (1, 2, 2)frame_buckets: tuple[int, int, int] = (2, 4, 24)enable_framepack: bool = Trueindex_rope_ref_image: int = 40patch_proj1x_coef: tuple[int, int, int] = (1, 1, 1)patch_proj2x_coef: tuple[int, int, int] = (2, 2, 2)patch_proj4x_coef: tuple[int, int, int] = (4, 4, 4)model_dim: int = 1792ff_dim: int = 7168num_text_blocks: int = 2num_audio_blocks: int = 2num_visual_blocks: int = 32axes_dims: tuple[int, int, int] = (16, 24, 24)audio_inject_layers: tuple[int, ...] = ()text_rope_max_pos: int = 4096 )
Speech-to-video diffusion transformer.
Parameter names match the training checkpoints, so a converted .safetensors loads with no key remapping. Inference is batch size 1: the sequence is packed and the speech cross-attention is frame-aligned, so forward takes a leading batch axis of 1 and squeezes it.
The visual stream is one packed sequence [memory | ref | visual]:
memory— FramePack context from the previous chunk at 4x / 2x / 1x temporal downsampling, placed at negative temporal RoPE positions;ref— the reference frame, pinned to temporal RoPE indexindex_rope_ref_image;visual— the noisy latent being denoised.
memory and ref are clean, so they are modulated with t = 0 while visual uses the real timestep; the
modulation holds both rows and an index selects per token.
forward
< source >( hidden_states: Tensorencoder_hidden_states: Tensoraudio_hidden_states: Tensorreference_hidden_states: Tensormemory_hidden_states: tuple[Tensor, Tensor, Tensor]timestep: Tensorpooled_projections: Tensorvisual_rope_pos: list[Tensor]reference_rope_pos: list[Tensor]memory_rope_pos: tuple[list[Tensor], list[Tensor], list[Tensor]]text_rope_pos: Tensoraudio_rope_pos: Tensorscale_factor: tuple[float, float, float] = (1.0, 1.0, 1.0)return_dict: bool = True ) → [~models.modeling_outputs.Transformer2DModelOutput] or tuple
Parameters
- hidden_states — noisy latent,
(1, T, H, W, C). - encoder_hidden_states — text embeddings,
(1, S_text, in_text_dim). - audio_hidden_states — speech features,
(1, S_audio, in_audio_dim). - reference_hidden_states — reference frame latent,
(1, 1, H, W, C). - memory_hidden_states — FramePack memory at 4x / 2x / 1x, each
(1, T_i, H_i, W_i, C). - timestep — current timestep,
(1,). - pooled_projections — pooled CLIP text embedding,
(1, in_text_dim2). - visual_rope_pos — per-axis RoPE positions of the noisy latent.
- reference_rope_pos — per-axis RoPE positions of the reference frame.
- memory_rope_pos — per-axis RoPE positions of each FramePack memory scale.
- text_rope_pos — 1-D RoPE positions of the text tokens.
- audio_rope_pos — latent frame each speech token belongs to.
- scale_factor — RoPE scale per axis, set by the resolution bucket.
- return_dict — return a
Transformer2DModelOutputinstead of a tuple.
Returns
[~models.modeling_outputs.Transformer2DModelOutput] or tuple
the predicted velocity, shaped like
hidden_states.
Denoise one chunk.
Batch size must be 1: the sequence is packed and the speech cross-attention is frame-aligned.