minimax-h3 / diffusers /models /autoencoders /autoencoder_kl_minimax_h3.py
multimodalart's picture
multimodalart HF Staff
Sync the split MiniMax-H3 Spaces
186aa49 verified
Raw
History Blame Contribute Delete
41.4 kB
# Copyright 2026 The MiniMax and HuggingFace Teams. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from ...configuration_utils import ConfigMixin, register_to_config
from ...utils import logging
from ...utils.accelerate_utils import apply_forward_hook
from ..attention import AttentionMixin, AttentionModuleMixin, FeedForward
from ..attention_dispatch import dispatch_attention_fn
from ..modeling_outputs import AutoencoderKLOutput
from ..modeling_utils import ModelMixin
from .vae import AutoencoderMixin, DecoderOutput, DiagonalGaussianDistribution
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class MiniMaxH3VideoCausalConv3d(nn.Conv3d):
r"""
3D convolution used throughout the MiniMax-H3 video encoder.
Spatial padding is symmetric and uses `spatial_padding_mode` (`"reflect"` in the released checkpoint); temporal
padding is causal, i.e. `kernel_size_t - 1` zero frames are prepended and nothing is appended.
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int | tuple[int, int, int],
stride: int | tuple[int, int, int] = 1,
spatial_padding: int = 0,
temporal_padding: int = 0,
spatial_padding_mode: str = "reflect",
) -> None:
super().__init__(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=0)
self.spatial_padding = spatial_padding
self.temporal_padding = temporal_padding
self.spatial_padding_mode = spatial_padding_mode
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
if self.spatial_padding > 0:
padding = self.spatial_padding
hidden_states = F.pad(
hidden_states, (padding, padding, padding, padding, 0, 0), mode=self.spatial_padding_mode
)
if self.temporal_padding > 0:
hidden_states = F.pad(hidden_states, (0, 0, 0, 0, self.temporal_padding, 0), mode="constant")
return F.conv3d(hidden_states, self.weight, self.bias, stride=self.stride, padding=0, dilation=self.dilation)
class MiniMaxH3VideoGroupNorm(nn.GroupNorm):
r"""
Group normalization applied to each latent frame in isolation (`use_t_isolated_gn` in the original config): the
temporal axis is folded into the batch axis so statistics never mix across frames.
"""
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
batch_size, num_channels, num_frames, height, width = hidden_states.shape
hidden_states = hidden_states.permute(0, 2, 1, 3, 4).contiguous()
hidden_states = hidden_states.view(batch_size * num_frames, num_channels, 1, height, width)
hidden_states = super().forward(hidden_states)
hidden_states = hidden_states.view(batch_size, num_frames, num_channels, height, width)
return hidden_states.permute(0, 2, 1, 3, 4).contiguous()
class MiniMaxH3VideoResnetBlock3d(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
norm_num_groups: int = 32,
norm_eps: float = 1e-6,
spatial_padding_mode: str = "reflect",
) -> None:
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.norm1 = MiniMaxH3VideoGroupNorm(norm_num_groups, in_channels, eps=norm_eps, affine=True)
self.conv1 = MiniMaxH3VideoCausalConv3d(
in_channels,
out_channels,
kernel_size=3,
spatial_padding=1,
temporal_padding=2,
spatial_padding_mode=spatial_padding_mode,
)
self.norm2 = MiniMaxH3VideoGroupNorm(norm_num_groups, out_channels, eps=norm_eps, affine=True)
self.conv2 = MiniMaxH3VideoCausalConv3d(
out_channels,
out_channels,
kernel_size=3,
spatial_padding=1,
temporal_padding=2,
spatial_padding_mode=spatial_padding_mode,
)
self.conv_shortcut = None
if in_channels != out_channels:
self.conv_shortcut = MiniMaxH3VideoCausalConv3d(in_channels, out_channels, kernel_size=1)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
residual = hidden_states
hidden_states = F.silu(self.norm1(hidden_states))
hidden_states = self.conv1(hidden_states)
hidden_states = F.silu(self.norm2(hidden_states))
hidden_states = self.conv2(hidden_states)
if self.conv_shortcut is not None:
residual = self.conv_shortcut(residual)
return residual + hidden_states
class MiniMaxH3VideoDownsample3d(nn.Module):
r"""
Strided 3x3x3 downsampling convolution. A spatial stride of 2 is preceded by an asymmetric bottom/right pad of 1
(the convolution itself carries no spatial padding), so the output is exactly `ceil(size / 2)`.
"""
def __init__(
self,
in_channels: int,
out_channels: int,
temporal_stride: int = 1,
spatial_stride: int = 2,
spatial_padding_mode: str = "reflect",
) -> None:
super().__init__()
self.spatial_stride = spatial_stride
self.spatial_padding_mode = spatial_padding_mode
self.conv = MiniMaxH3VideoCausalConv3d(
in_channels,
out_channels,
kernel_size=3,
stride=(temporal_stride, spatial_stride, spatial_stride),
spatial_padding=0,
temporal_padding=2,
spatial_padding_mode=spatial_padding_mode,
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
if self.spatial_stride == 2:
hidden_states = F.pad(hidden_states, (0, 1, 0, 1, 0, 0), mode=self.spatial_padding_mode)
return self.conv(hidden_states)
class MiniMaxH3VideoDownBlock3d(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
num_layers: int,
temporal_downsample_factor: int,
spatial_downsample_factor: int,
norm_num_groups: int = 32,
norm_eps: float = 1e-6,
spatial_padding_mode: str = "reflect",
) -> None:
super().__init__()
self.resnets = nn.ModuleList(
[
MiniMaxH3VideoResnetBlock3d(
in_channels=in_channels if i == 0 else out_channels,
out_channels=out_channels,
norm_num_groups=norm_num_groups,
norm_eps=norm_eps,
spatial_padding_mode=spatial_padding_mode,
)
for i in range(num_layers)
]
)
self.downsamplers = None
if temporal_downsample_factor * spatial_downsample_factor > 1:
self.downsamplers = nn.ModuleList(
[
MiniMaxH3VideoDownsample3d(
out_channels,
out_channels,
temporal_stride=temporal_downsample_factor,
spatial_stride=spatial_downsample_factor,
spatial_padding_mode=spatial_padding_mode,
)
]
)
self.gradient_checkpointing = False
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
for resnet in self.resnets:
if torch.is_grad_enabled() and self.gradient_checkpointing:
hidden_states = self._gradient_checkpointing_func(resnet, hidden_states)
else:
hidden_states = resnet(hidden_states)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states)
return hidden_states
class MiniMaxH3VideoEncoder3d(nn.Module):
r"""
Causal 3D CNN encoder. `block_out_channels` gives the channel count of every level; the per-level
`spatial_downsample_factors` / `temporal_downsample_factors` multiply out to the total compression ratios.
"""
def __init__(
self,
in_channels: int = 3,
out_channels: int = 48,
block_out_channels: tuple[int, ...] = (128, 256, 256, 512, 512, 1024),
layers_per_block: int = 2,
spatial_downsample_factors: tuple[int, ...] = (2, 2, 2, 2, 1, 1),
temporal_downsample_factors: tuple[int, ...] = (1, 2, 2, 1, 1, 1),
norm_num_groups: int = 32,
norm_eps: float = 1e-6,
spatial_padding_mode: str = "reflect",
) -> None:
super().__init__()
self.conv_in = MiniMaxH3VideoCausalConv3d(
in_channels,
block_out_channels[0],
kernel_size=3,
spatial_padding=1,
temporal_padding=2,
spatial_padding_mode=spatial_padding_mode,
)
block_in_channels = (block_out_channels[0],) + tuple(block_out_channels[:-1])
self.down_blocks = nn.ModuleList(
[
MiniMaxH3VideoDownBlock3d(
in_channels=block_in_channels[i],
out_channels=block_out_channels[i],
num_layers=layers_per_block,
temporal_downsample_factor=temporal_downsample_factors[i],
spatial_downsample_factor=spatial_downsample_factors[i],
norm_num_groups=norm_num_groups,
norm_eps=norm_eps,
spatial_padding_mode=spatial_padding_mode,
)
for i in range(len(block_out_channels))
]
)
self.norm_out = MiniMaxH3VideoGroupNorm(norm_num_groups, block_out_channels[-1], eps=norm_eps, affine=True)
self.conv_out = MiniMaxH3VideoCausalConv3d(
block_out_channels[-1],
out_channels,
kernel_size=3,
spatial_padding=1,
temporal_padding=2,
spatial_padding_mode=spatial_padding_mode,
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
hidden_states = self.conv_in(hidden_states)
for down_block in self.down_blocks:
hidden_states = down_block(hidden_states)
hidden_states = F.silu(self.norm_out(hidden_states))
return self.conv_out(hidden_states)
class MiniMaxH3VideoRotaryPosEmbed(nn.Module):
r"""
3-axis rotary embedding for the ViT decoder. Coordinates are length-normalized to `[-1, 1)` per axis and scaled by
`2 * pi`, and the resulting `(t, h, w)` angles are concatenated and then duplicated, so the first
`rope_dim_ratio * attention_head_dim` channels of every head are rotated.
"""
def __init__(self, dim: int, theta: float = 100.0, num_axes: int = 3) -> None:
super().__init__()
if dim % (2 * num_axes) != 0:
raise ValueError(f"`dim` {dim} must be divisible by `2 * num_axes` {2 * num_axes}.")
inv_freq = 1.0 / theta ** torch.arange(0, 1, 2 * num_axes / dim, dtype=torch.float32)
self.register_buffer("inv_freq", inv_freq, persistent=False)
def forward(self, position_ids: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
angles = 2.0 * math.pi * position_ids[:, :, :, None] * self.inv_freq[None, None, None, :]
angles = angles.flatten(2, 3).tile(2).unsqueeze(2)
return angles.cos(), angles.sin()
class MiniMaxH3VideoAttnProcessor:
_attention_backend = None
_parallel_config = None
def __call__(
self,
attn: "MiniMaxH3VideoAttention",
hidden_states: torch.Tensor,
rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None,
) -> torch.Tensor:
query = attn.to_q(hidden_states).unflatten(2, (attn.heads, -1))
key = attn.to_k(hidden_states).unflatten(2, (attn.heads, -1))
value = attn.to_v(hidden_states).unflatten(2, (attn.heads, -1))
# The reference normalizes Q/K in float32 regardless of the compute dtype.
query = attn.norm_q(query.float()).to(query.dtype)
key = attn.norm_k(key.float()).to(key.dtype)
if rotary_emb is not None:
cos, sin = rotary_emb
cos = cos.to(query.dtype)
sin = sin.to(query.dtype)
rotary_dim = cos.shape[-1]
query_rotary, query_pass = query[..., :rotary_dim], query[..., rotary_dim:]
key_rotary, key_pass = key[..., :rotary_dim], key[..., rotary_dim:]
query_first, query_second = query_rotary.chunk(2, dim=-1)
key_first, key_second = key_rotary.chunk(2, dim=-1)
query_rotated = torch.cat([-query_second, query_first], dim=-1)
key_rotated = torch.cat([-key_second, key_first], dim=-1)
query = torch.cat([query_rotary * cos + query_rotated * sin, query_pass], dim=-1)
key = torch.cat([key_rotary * cos + key_rotated * sin, key_pass], dim=-1)
hidden_states = dispatch_attention_fn(
query,
key,
value,
attn_mask=None,
backend=self._attention_backend,
parallel_config=self._parallel_config,
)
hidden_states = hidden_states.flatten(2, 3)
return attn.to_out[0](hidden_states)
class MiniMaxH3VideoAttention(nn.Module, AttentionModuleMixin):
_default_processor_cls = MiniMaxH3VideoAttnProcessor
_available_processors = [MiniMaxH3VideoAttnProcessor]
def __init__(self, dim: int, heads: int, dim_head: int, eps: float = 1e-5, bias: bool = True) -> None:
super().__init__()
self.heads = heads
self.dim_head = dim_head
self.use_bias = bias
inner_dim = heads * dim_head
self.norm_q = nn.RMSNorm(dim_head, eps=eps, elementwise_affine=False)
self.norm_k = nn.RMSNorm(dim_head, eps=eps, elementwise_affine=False)
self.to_q = nn.Linear(dim, inner_dim, bias=bias)
self.to_k = nn.Linear(dim, inner_dim, bias=bias)
self.to_v = nn.Linear(dim, inner_dim, bias=bias)
self.to_out = nn.ModuleList([nn.Linear(inner_dim, dim, bias=bias), nn.Dropout(0.0)])
self.set_processor(MiniMaxH3VideoAttnProcessor())
def forward(
self, hidden_states: torch.Tensor, rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None
) -> torch.Tensor:
return self.processor(self, hidden_states, rotary_emb)
class MiniMaxH3VideoTransformerBlock(nn.Module):
def __init__(
self,
dim: int,
heads: int,
dim_head: int,
ffn_mult: int = 4,
eps: float = 1e-5,
bias: bool = True,
) -> None:
super().__init__()
self.norm1 = nn.RMSNorm(dim, eps=eps, elementwise_affine=True)
self.attn = MiniMaxH3VideoAttention(dim=dim, heads=heads, dim_head=dim_head, eps=eps, bias=bias)
self.scale1 = nn.Parameter(torch.zeros(dim))
self.norm2 = nn.RMSNorm(dim, eps=eps, elementwise_affine=True)
self.ff = FeedForward(dim, mult=ffn_mult, activation_fn="swiglu", bias=bias)
self.scale2 = nn.Parameter(torch.zeros(dim))
def forward(
self, hidden_states: torch.Tensor, rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None
) -> torch.Tensor:
# The reference normalizes in float32 regardless of the compute dtype.
norm_hidden_states = self.norm1(hidden_states.float()).to(hidden_states.dtype)
hidden_states = hidden_states + self.attn(norm_hidden_states, rotary_emb) * self.scale1
norm_hidden_states = self.norm2(hidden_states.float()).to(hidden_states.dtype)
hidden_states = hidden_states + self.ff(norm_hidden_states) * self.scale2
return hidden_states
class MiniMaxH3VideoViTDecoder3d(nn.Module):
r"""
Non-causal ViT decoder. Every latent voxel becomes one token; `num_register_tokens` learned register tokens plus a
single all-zero token are appended (all at position `0`), attended over with full self-attention, and dropped
again before the patch projection expands each token into a `patch_size_t x patch_size x patch_size` pixel block.
"""
def __init__(
self,
in_channels: int = 24,
out_channels: int = 3,
patch_size: int = 16,
patch_size_t: int = 4,
num_layers: int = 36,
num_attention_heads: int = 32,
attention_head_dim: int = 64,
num_register_tokens: int = 4,
ffn_mult: int = 4,
rope_theta: float = 100.0,
rope_dim_ratio: float = 0.75,
norm_eps: float = 1e-5,
) -> None:
super().__init__()
dim = num_attention_heads * attention_head_dim
self.patch_size = patch_size
self.patch_size_t = patch_size_t
self.out_channels = out_channels
self.num_register_tokens = num_register_tokens
self.rope = MiniMaxH3VideoRotaryPosEmbed(int(attention_head_dim * rope_dim_ratio), theta=rope_theta)
self.proj_in = nn.Linear(in_channels, dim)
self.register_tokens = nn.Parameter(torch.zeros(1, num_register_tokens, dim))
self.transformer_blocks = nn.ModuleList(
[
MiniMaxH3VideoTransformerBlock(
dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
ffn_mult=ffn_mult,
eps=norm_eps,
)
for _ in range(num_layers)
]
)
self.norm_out = nn.LayerNorm(dim, elementwise_affine=True, eps=norm_eps)
self.proj_out = nn.Linear(dim, out_channels * patch_size_t * patch_size * patch_size)
self.gradient_checkpointing = False
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
batch_size, num_channels, num_frames, height, width = hidden_states.shape
hidden_states = hidden_states.permute(0, 2, 3, 4, 1).reshape(
batch_size, num_frames * height * width, num_channels
)
hidden_states = self.proj_in(hidden_states)
num_patches = hidden_states.shape[1]
register_tokens = self.register_tokens.expand(batch_size, -1, -1)
cls_token = torch.zeros_like(hidden_states[:, :1, :])
hidden_states = torch.cat([hidden_states, register_tokens, cls_token], dim=1)
grids = [
2.0 * (torch.arange(0.5, size, dtype=torch.float32, device=hidden_states.device) / size) - 1.0
for size in (num_frames, height, width)
]
position_ids = torch.stack(torch.meshgrid(*grids, indexing="ij"), dim=-1).flatten(0, 2)
position_ids = position_ids.unsqueeze(0).expand(batch_size, -1, -1)
suffix_ids = position_ids.new_zeros((batch_size, self.num_register_tokens + 1, 3))
position_ids = torch.cat([position_ids, suffix_ids], dim=1)
rotary_emb = self.rope(position_ids)
for block in self.transformer_blocks:
if torch.is_grad_enabled() and self.gradient_checkpointing:
hidden_states = self._gradient_checkpointing_func(block, hidden_states, rotary_emb)
else:
hidden_states = block(hidden_states, rotary_emb)
hidden_states = self.norm_out(hidden_states)
hidden_states = self.proj_out(hidden_states)
hidden_states = hidden_states[:, :num_patches, :]
patch_size, patch_size_t = self.patch_size, self.patch_size_t
hidden_states = hidden_states.view(
batch_size,
num_frames,
height,
width,
self.out_channels,
patch_size_t,
patch_size,
patch_size,
)
hidden_states = hidden_states.permute(0, 4, 1, 5, 2, 6, 3, 7).contiguous()
return hidden_states.reshape(
batch_size,
self.out_channels,
num_frames * patch_size_t,
height * patch_size,
width * patch_size,
)
class AutoencoderKLMiniMaxH3(ModelMixin, ConfigMixin, AttentionMixin, AutoencoderMixin):
r"""
A VAE model with a causal 3D CNN encoder and a non-causal ViT decoder, used in
[MiniMax-H3](https://huggingface.co/MiniMaxAI).
This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
for all models (such as downloading or saving).
Latents are normalized with per-channel `latents_mean` / `latents_std` rather than a `scaling_factor`; a pipeline
encodes with `(latent - latents_mean) / latents_std` and decodes with `latent * latents_std + latents_mean`.
The pixel convention is ImageNet-normalized RGB over a `[0, 1]` base range, not the usual `[-1, 1]`: `encode`
expects `(pixel - imagenet_mean) / imagenet_std` and `decode` returns values in that same space, so a pipeline has
to apply `sample * imagenet_std + imagenet_mean` (mean `(0.485, 0.456, 0.406)`, std `(0.229, 0.224, 0.225)`) and
clamp to `[0, 1]` before postprocessing.
The temporal geometry is fixed by `clip_length` (17 pixel frames per encoder chunk) and `token_drop` (3 trailing
latent frames dropped per encode): `17 * n + 5` pixel frames map to `5 * n + 2` latent frames.
Unlike most autoencoders in the library, spatial tiling is **on by default**: MiniMax-H3 was released with tiling
enabled for both encoding and decoding, and the released frames are the blended-tile ones, so disabling tiling
changes the output. Use `enable_tiling` to change the tile geometry, `disable_tiling` to turn it off.
"""
_supports_gradient_checkpointing = True
_no_split_modules = ["MiniMaxH3VideoResnetBlock3d", "MiniMaxH3VideoTransformerBlock"]
_repeated_blocks = ["MiniMaxH3VideoTransformerBlock"]
_skip_layerwise_casting_patterns = ["norm"]
# The released checkpoint is float32 and the verified decode recipe is float16 *autocast over float32 weights*
# (see `decode`). A pipeline-level `torch_dtype=torch.bfloat16` must therefore not downcast the weights, so every
# top-level module is pinned, mirroring the transformer's mixed-precision contract.
_keep_in_fp32_modules = ["encoder", "decoder", "quant_conv", "post_quant_conv"]
@register_to_config
def __init__(
self,
in_channels: int = 3,
out_channels: int = 3,
latent_channels: int = 24,
block_out_channels: tuple[int, ...] = (128, 256, 256, 512, 512, 1024),
layers_per_block: int = 2,
spatial_downsample_factors: tuple[int, ...] = (2, 2, 2, 2, 1, 1),
temporal_downsample_factors: tuple[int, ...] = (1, 2, 2, 1, 1, 1),
norm_num_groups: int = 32,
norm_eps: float = 1e-6,
spatial_padding_mode: str = "reflect",
decoder_num_layers: int = 36,
decoder_num_attention_heads: int = 32,
decoder_attention_head_dim: int = 64,
decoder_num_register_tokens: int = 4,
decoder_ffn_mult: int = 4,
decoder_rope_theta: float = 100.0,
decoder_rope_dim_ratio: float = 0.75,
decoder_norm_eps: float = 1e-5,
clip_length: int = 17,
token_drop: int = 3,
latents_mean: tuple[float, ...] = (0.0,) * 24,
latents_std: tuple[float, ...] = (1.0,) * 24,
) -> None:
super().__init__()
self.spatial_compression_ratio = math.prod(spatial_downsample_factors)
self.temporal_compression_ratio = math.prod(temporal_downsample_factors)
self.encoder = MiniMaxH3VideoEncoder3d(
in_channels=in_channels,
out_channels=2 * latent_channels,
block_out_channels=block_out_channels,
layers_per_block=layers_per_block,
spatial_downsample_factors=spatial_downsample_factors,
temporal_downsample_factors=temporal_downsample_factors,
norm_num_groups=norm_num_groups,
norm_eps=norm_eps,
spatial_padding_mode=spatial_padding_mode,
)
self.quant_conv = nn.Conv3d(2 * latent_channels, 2 * latent_channels, kernel_size=1)
self.post_quant_conv = nn.Conv3d(latent_channels, latent_channels, kernel_size=1)
self.decoder = MiniMaxH3VideoViTDecoder3d(
in_channels=latent_channels,
out_channels=out_channels,
patch_size=self.spatial_compression_ratio,
patch_size_t=self.temporal_compression_ratio,
num_layers=decoder_num_layers,
num_attention_heads=decoder_num_attention_heads,
attention_head_dim=decoder_attention_head_dim,
num_register_tokens=decoder_num_register_tokens,
ffn_mult=decoder_ffn_mult,
rope_theta=decoder_rope_theta,
rope_dim_ratio=decoder_rope_dim_ratio,
norm_eps=decoder_norm_eps,
)
# Derived temporal-chunking geometry. `clip_length` pixel frames are encoded at a time; because
# `clip_length` is not a multiple of `temporal_compression_ratio`, the decoder has to re-derive the
# implicit leading pad (`frame_pre_padding`) and the overlap that `token_drop` leaves behind.
self.frame_pre_padding = (-clip_length) % self.temporal_compression_ratio
self.tokens_chunk_size = math.ceil(clip_length / self.temporal_compression_ratio)
self.token_overlap = (-token_drop) % self.tokens_chunk_size
self.frame_overlap = max(self.token_overlap * self.temporal_compression_ratio - self.frame_pre_padding, 0)
# When decoding a batch of video latents at a time, one can save memory by slicing across the batch dimension
# to perform decoding of a single video latent at a time.
self.use_slicing = False
# When encoding/decoding spatially large videos, the memory requirement is very high. By splitting the frames
# into smaller tiles, running the encoder/decoder per tile and blending the overlaps, the memory requirement
# can be lowered. MiniMax-H3 ships with tiling enabled.
self.use_tiling = True
# The tile size in pixel space, and the minimum overlap between two neighbouring tiles. The actual overlaps are
# widened (in multiples of `spatial_compression_ratio`) so that the tiles cover the frame exactly.
self.tile_sample_min_height = 256
self.tile_sample_min_width = 256
self.tile_sample_min_overlap_height = 64
self.tile_sample_min_overlap_width = 64
def enable_tiling(
self,
tile_sample_min_height: int | None = None,
tile_sample_min_width: int | None = None,
tile_sample_min_overlap_height: int | None = None,
tile_sample_min_overlap_width: int | None = None,
) -> None:
r"""
Enable tiled VAE encoding/decoding. When this option is enabled, the VAE splits the frames into tiles, encodes
or decodes each tile separately and linearly blends the overlaps back together. This lowers the memory
requirement and allows processing larger frames.
Args:
tile_sample_min_height (`int`, *optional*):
The tile height in pixel space. Frames taller than this are split along the height dimension.
tile_sample_min_width (`int`, *optional*):
The tile width in pixel space. Frames wider than this are split along the width dimension.
tile_sample_min_overlap_height (`int`, *optional*):
The minimum overlap, in pixels, between two consecutive vertical tiles.
tile_sample_min_overlap_width (`int`, *optional*):
The minimum overlap, in pixels, between two consecutive horizontal tiles.
"""
self.use_tiling = True
self.tile_sample_min_height = tile_sample_min_height or self.tile_sample_min_height
self.tile_sample_min_width = tile_sample_min_width or self.tile_sample_min_width
self.tile_sample_min_overlap_height = tile_sample_min_overlap_height or self.tile_sample_min_overlap_height
self.tile_sample_min_overlap_width = tile_sample_min_overlap_width or self.tile_sample_min_overlap_width
def _split_tiles(self, length: int, tile_size: int, min_overlap: int) -> tuple[list[int], list[int], list[int]]:
r"""
Lay `tile_size`-wide tiles over `length` pixels. The number of tiles is the smallest one whose union can cover
`length` while keeping every overlap at least `min_overlap`; the slack is then distributed round-robin over the
overlaps in whole `spatial_compression_ratio` steps so that every tile boundary stays latent-aligned.
"""
if tile_size >= length:
return [0], [length], []
num_tiles = math.ceil(length / tile_size)
while tile_size * num_tiles - min_overlap * (num_tiles - 1) - length < 0:
num_tiles += 1
overlaps = [min_overlap] * (num_tiles - 1)
remaining = tile_size * num_tiles - sum(overlaps) - length
for i in range(remaining // self.spatial_compression_ratio):
overlaps[i % (num_tiles - 1)] += self.spatial_compression_ratio
tile_start_indices = [0]
for i in range(num_tiles - 1):
tile_start_indices.append(tile_start_indices[-1] + tile_size - overlaps[i])
return tile_start_indices, [tile_size] * num_tiles, overlaps
def _blend(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int, dim: int) -> torch.Tensor:
blend_extent = min(a.shape[dim], b.shape[dim], blend_extent)
positions = torch.arange(blend_extent, device=b.device, dtype=b.dtype)
shape = [1] * a.ndim
shape[dim] = blend_extent
weight_a = (1 - positions / blend_extent).view(shape)
weight_b = (positions / blend_extent).view(shape)
slice_a = [slice(None)] * a.ndim
slice_a[dim] = slice(-blend_extent, None)
slice_b = [slice(None)] * b.ndim
slice_b[dim] = slice(0, blend_extent)
blended = a[tuple(slice_a)] * weight_a + b[tuple(slice_b)] * weight_b
if blend_extent == b.shape[dim]:
return blended
slice_rest = [slice(None)] * b.ndim
slice_rest[dim] = slice(blend_extent, None)
return torch.cat([blended, b[tuple(slice_rest)]], dim=dim)
def _stitch_tiles(
self,
tiles: list[list[torch.Tensor]],
height_overlaps: list[int],
width_overlaps: list[int],
) -> torch.Tensor:
result_rows = []
for i, row in enumerate(tiles):
result_row = []
for j, tile in enumerate(row):
if i > 0:
tile = self._blend(tiles[i - 1][j], tile, height_overlaps[i - 1], dim=-2)
if j > 0:
tile = self._blend(row[j - 1], tile, width_overlaps[j - 1], dim=-1)
if i < len(tiles) - 1:
tile = tile[..., : -height_overlaps[i], :]
if j < len(row) - 1:
tile = tile[..., :, : -width_overlaps[j]]
result_row.append(tile)
result_rows.append(torch.cat(result_row, dim=-1))
return torch.cat(result_rows, dim=-2)
@apply_forward_hook
def _encode_clip(self, x: torch.Tensor) -> torch.Tensor:
r"""
Encode one temporal clip, spatially tiled when tiling is enabled.
MiniMax-H3 encodes a keyframe or an image reference through this method rather than through [`~encode`],
because a single frame must not go through the temporal chunking, so it carries the offload hook too.
"""
if not self.use_tiling:
return self.quant_conv(self.encoder(x))
height, width = x.shape[-2], x.shape[-1]
y_indices, y_lengths, y_overlaps = self._split_tiles(
height, self.tile_sample_min_height, self.tile_sample_min_overlap_height
)
x_indices, x_lengths, x_overlaps = self._split_tiles(
width, self.tile_sample_min_width, self.tile_sample_min_overlap_width
)
rows = []
for i_pos, i_len in zip(y_indices, y_lengths):
row = []
for j_pos, j_len in zip(x_indices, x_lengths):
tile = x[..., i_pos : i_pos + i_len, j_pos : j_pos + j_len]
row.append(self.quant_conv(self.encoder(tile)))
rows.append(row)
latent_y_overlaps = [overlap // self.spatial_compression_ratio for overlap in y_overlaps]
latent_x_overlaps = [overlap // self.spatial_compression_ratio for overlap in x_overlaps]
return self._stitch_tiles(rows, latent_y_overlaps, latent_x_overlaps)
def _decode_clip(self, z: torch.Tensor) -> torch.Tensor:
r"""Decode one temporal clip, spatially tiled when tiling is enabled."""
if not self.use_tiling:
return self.decoder(self.post_quant_conv(z))
# Tiles are laid out in pixel space and then mapped back onto the latent grid.
height = z.shape[-2] * self.spatial_compression_ratio
width = z.shape[-1] * self.spatial_compression_ratio
y_indices, y_lengths, y_overlaps = self._split_tiles(
height, self.tile_sample_min_height, self.tile_sample_min_overlap_height
)
x_indices, x_lengths, x_overlaps = self._split_tiles(
width, self.tile_sample_min_width, self.tile_sample_min_overlap_width
)
ratio = self.spatial_compression_ratio
rows = []
for i_pos, i_len in zip(y_indices, y_lengths):
row = []
for j_pos, j_len in zip(x_indices, x_lengths):
tile = z[
...,
i_pos // ratio : i_pos // ratio + i_len // ratio,
j_pos // ratio : j_pos // ratio + j_len // ratio,
]
row.append(self.decoder(self.post_quant_conv(tile)))
rows.append(row)
return self._stitch_tiles(rows, y_overlaps, x_overlaps)
@apply_forward_hook
def _encode(self, x: torch.Tensor) -> torch.Tensor:
r"""
Encode a video in `clip_length`-frame chunks and drop the `token_drop` trailing latent frames.
MiniMax-H3 encodes a video reference through this method rather than through [`~encode`], because the
posterior is sampled under a fixed generator rather than through the distribution object, so it carries the
offload hook too.
"""
clip_length = self.config.clip_length
num_frames = x.shape[2]
if num_frames % clip_length != 0:
pad_frames = x[:, :, -1:].repeat(1, 1, (-num_frames) % clip_length, 1, 1)
x = torch.cat([x, pad_frames], dim=2)
moments = torch.cat(
[
self._encode_clip(x[:, :, i * clip_length : (i + 1) * clip_length])
for i in range(x.shape[2] // clip_length)
],
dim=2,
)
if self.config.token_drop > 0:
moments = moments[:, :, : -self.config.token_drop]
return moments
def _decode(self, z: torch.Tensor) -> torch.Tensor:
r"""
Decode a latent video, mirroring the chunking that `_encode` applied.
`token_drop` removed the tail of every encoded chunk, so consecutive decoded chunks overlap by
`frame_overlap` pixel frames and are linearly cross-faded. Latent frames are repeated at the end when the
length is not a whole number of chunks; the extra pixel frames are cut off again at the end.
"""
tokens_chunk_size = self.tokens_chunk_size
token_drop = self.config.token_drop
temporal_ratio = self.temporal_compression_ratio
chunk_num_frames = tokens_chunk_size * temporal_ratio
num_tokens = z.shape[2] + token_drop
pad_tokens = (-num_tokens) % tokens_chunk_size
num_chunks = (num_tokens + pad_tokens) // tokens_chunk_size - int(token_drop > 0)
if pad_tokens > 0:
z = torch.cat([z, z[:, :, -1:].repeat(1, 1, pad_tokens, 1, 1)], dim=2)
decoded_chunks = []
overlap = None
for i in range(num_chunks):
start = i * tokens_chunk_size
clip = self._decode_clip(z[:, :, start : start + tokens_chunk_size + self.token_overlap])
for j in range(int(token_drop > 0) + 1):
frame_start = j * chunk_num_frames
chunk = clip[:, :, frame_start : frame_start + chunk_num_frames]
chunk = chunk[:, :, self.frame_pre_padding :]
if j == 0:
if overlap is not None:
chunk = self._blend(overlap, chunk, self.frame_overlap, dim=-3)
decoded_chunks.append(chunk)
else:
overlap = chunk
if overlap is not None:
decoded_chunks.append(overlap)
dec = torch.cat(decoded_chunks, dim=2)
# `pad_tokens` repeated latent frames produced trailing pixel frames that were never requested. A chunk's
# last latent frame only covers `clip_length % temporal_ratio` pixel frames, the others cover `temporal_ratio`.
if pad_tokens > 0:
intra_tail = self.config.clip_length % temporal_ratio
num_tokens_before_pad = z.shape[2] - pad_tokens
pad_frames = sum(
intra_tail if intra_tail and (num_tokens_before_pad + k) % tokens_chunk_size == 0 else temporal_ratio
for k in range(pad_tokens)
)
dec = dec[:, :, :-pad_frames]
return dec
@apply_forward_hook
def encode(self, x: torch.Tensor, return_dict: bool = True) -> AutoencoderKLOutput | tuple[torch.Tensor]:
r"""
Encode a batch of videos into latents.
Args:
x (`torch.Tensor`):
Input batch of videos, shape `(batch_size, in_channels, num_frames, height, width)`.
return_dict (`bool`, *optional*, defaults to `True`):
Whether to return a [`~models.autoencoders.autoencoder_kl.AutoencoderKLOutput`] instead of a plain
tuple.
Returns:
The latent distribution of the encoded videos. Note that MiniMax-H3 normalizes the sampled latents with
`latents_mean` / `latents_std` afterwards.
"""
if self.use_slicing and x.shape[0] > 1:
moments = torch.cat([self._encode(x_slice) for x_slice in x.split(1)])
else:
moments = self._encode(x)
posterior = DiagonalGaussianDistribution(moments)
if not return_dict:
return (posterior,)
return AutoencoderKLOutput(latent_dist=posterior)
@apply_forward_hook
def decode(self, z: torch.Tensor, return_dict: bool = True) -> DecoderOutput | tuple[torch.Tensor]:
r"""
Decode a batch of latent videos.
Args:
z (`torch.Tensor`):
Input batch of latent videos, shape `(batch_size, latent_channels, num_latent_frames, height, width)`.
return_dict (`bool`, *optional*, defaults to `True`):
Whether to return a [`~models.autoencoders.vae.DecoderOutput`] instead of a plain tuple.
Returns:
[`~models.autoencoders.vae.DecoderOutput`] or `tuple`:
The decoded videos, shape `(batch_size, out_channels, num_frames, height, width)`.
"""
if self.use_slicing and z.shape[0] > 1:
decoded = torch.cat([self._decode(z_slice) for z_slice in z.split(1)])
else:
decoded = self._decode(z)
if not return_dict:
return (decoded,)
return DecoderOutput(sample=decoded)
def forward(
self,
sample: torch.Tensor,
sample_posterior: bool = False,
generator: torch.Generator | None = None,
return_dict: bool = True,
) -> DecoderOutput | tuple[torch.Tensor]:
r"""
Encode then decode a batch of videos.
Args:
sample (`torch.Tensor`):
Input batch of videos, shape `(batch_size, in_channels, num_frames, height, width)`.
sample_posterior (`bool`, *optional*, defaults to `False`):
Whether to sample the posterior instead of taking its mode.
generator (`torch.Generator`, *optional*):
Generator used when `sample_posterior=True`.
return_dict (`bool`, *optional*, defaults to `True`):
Whether to return a [`~models.autoencoders.vae.DecoderOutput`] instead of a plain tuple.
Returns:
[`~models.autoencoders.vae.DecoderOutput`] or `tuple`:
The round-tripped videos, shape `(batch_size, out_channels, num_frames, height, width)`.
"""
posterior = self.encode(sample).latent_dist
z = posterior.sample(generator=generator) if sample_posterior else posterior.mode()
return self.decode(z, return_dict=return_dict)