InvokeAI 6.10.0rc1__py3-none-any.whl → 6.11.0__py3-none-any.whl

invokeai/backend/flux/dype/base.py

@@ -0,0 +1,260 @@
+"""DyPE base configuration and utilities."""
+
+import math
+from dataclasses import dataclass
+from typing import Literal
+
+import torch
+from torch import Tensor
+
+
+@dataclass
+class DyPEConfig:
+    """Configuration for Dynamic Position Extrapolation."""
+
+    enable_dype: bool = True
+    base_resolution: int = 1024  # Native training resolution
+    method: Literal["vision_yarn", "yarn", "ntk", "base"] = "vision_yarn"
+    dype_scale: float = 2.0  # Magnitude λs (0.0-8.0)
+    dype_exponent: float = 2.0  # Decay speed λt (0.0-1000.0)
+    dype_start_sigma: float = 1.0  # When DyPE decay starts
+
+
+def get_mscale(scale: float, mscale_factor: float = 1.0) -> float:
+    """Calculate magnitude scaling factor.
+
+    Args:
+        scale: The resolution scaling factor
+        mscale_factor: Adjustment factor for the scaling
+
+    Returns:
+        The magnitude scaling factor
+    """
+    if scale <= 1.0:
+        return 1.0
+    return mscale_factor * math.log(scale) + 1.0
+
+
+def get_timestep_mscale(
+    scale: float,
+    current_sigma: float,
+    dype_scale: float,
+    dype_exponent: float,
+    dype_start_sigma: float,
+) -> float:
+    """Calculate timestep-dependent magnitude scaling.
+
+    The key insight of DyPE: early steps focus on low frequencies (global structure),
+    late steps on high frequencies (details). This function modulates the scaling
+    based on the current timestep/sigma.
+
+    Args:
+        scale: Resolution scaling factor
+        current_sigma: Current noise level (1.0 = full noise, 0.0 = clean)
+        dype_scale: DyPE magnitude (λs)
+        dype_exponent: DyPE decay speed (λt)
+        dype_start_sigma: Sigma threshold to start decay
+
+    Returns:
+        Timestep-modulated scaling factor
+    """
+    if scale <= 1.0:
+        return 1.0
+
+    # Normalize sigma to [0, 1] range relative to start_sigma
+    if current_sigma >= dype_start_sigma:
+        t_normalized = 1.0
+    else:
+        t_normalized = current_sigma / dype_start_sigma
+
+    # Apply exponential decay: stronger extrapolation early, weaker late
+    # decay = exp(-λt * (1 - t))  where t=1 is early (high sigma), t=0 is late
+    decay = math.exp(-dype_exponent * (1.0 - t_normalized))
+
+    # Base mscale from resolution
+    base_mscale = get_mscale(scale)
+
+    # Interpolate between base_mscale and 1.0 based on decay and dype_scale
+    # When decay=1 (early): use scaled value
+    # When decay=0 (late): use base value
+    scaled_mscale = 1.0 + (base_mscale - 1.0) * dype_scale * decay
+
+    return scaled_mscale
+
+
+def compute_vision_yarn_freqs(
+    pos: Tensor,
+    dim: int,
+    theta: int,
+    scale_h: float,
+    scale_w: float,
+    current_sigma: float,
+    dype_config: DyPEConfig,
+) -> tuple[Tensor, Tensor]:
+    """Compute RoPE frequencies using NTK-aware scaling for high-resolution.
+
+    This method extends FLUX's position encoding to handle resolutions beyond
+    the 1024px training resolution by scaling the base frequency (theta).
+
+    The NTK-aware approach smoothly interpolates frequencies to cover larger
+    position ranges without breaking the attention patterns.
+
+    DyPE (Dynamic Position Extrapolation) modulates the NTK scaling based on
+    the current timestep - stronger extrapolation in early steps (global structure),
+    weaker in late steps (fine details).
+
+    Args:
+        pos: Position tensor
+        dim: Embedding dimension
+        theta: RoPE base frequency
+        scale_h: Height scaling factor
+        scale_w: Width scaling factor
+        current_sigma: Current noise level (1.0 = full noise, 0.0 = clean)
+        dype_config: DyPE configuration
+
+    Returns:
+        Tuple of (cos, sin) frequency tensors
+    """
+    assert dim % 2 == 0
+
+    # Use the larger scale for NTK calculation
+    scale = max(scale_h, scale_w)
+
+    device = pos.device
+    dtype = torch.float64 if device.type != "mps" else torch.float32
+
+    # NTK-aware theta scaling: extends position coverage for high-res
+    # Formula: theta_scaled = theta * scale^(dim/(dim-2))
+    # This increases the wavelength of position encodings proportionally
+    if scale > 1.0:
+        ntk_alpha = scale ** (dim / (dim - 2))
+
+        # Apply timestep-dependent DyPE modulation
+        # mscale controls how strongly we apply the NTK extrapolation
+        # Early steps (high sigma): stronger extrapolation for global structure
+        # Late steps (low sigma): weaker extrapolation for fine details
+        mscale = get_timestep_mscale(
+            scale=scale,
+            current_sigma=current_sigma,
+            dype_scale=dype_config.dype_scale,
+            dype_exponent=dype_config.dype_exponent,
+            dype_start_sigma=dype_config.dype_start_sigma,
+        )
+
+        # Modulate NTK alpha by mscale
+        # When mscale > 1: interpolate towards stronger extrapolation
+        # When mscale = 1: use base NTK alpha
+        modulated_alpha = 1.0 + (ntk_alpha - 1.0) * mscale
+        scaled_theta = theta * modulated_alpha
+    else:
+        scaled_theta = theta
+
+    # Standard RoPE frequency computation
+    freq_seq = torch.arange(0, dim, 2, dtype=dtype, device=device) / dim
+    freqs = 1.0 / (scaled_theta**freq_seq)
+
+    # Compute angles = position * frequency
+    angles = torch.einsum("...n,d->...nd", pos.to(dtype), freqs)
+
+    cos = torch.cos(angles)
+    sin = torch.sin(angles)
+
+    return cos.to(pos.dtype), sin.to(pos.dtype)
+
+
+def compute_yarn_freqs(
+    pos: Tensor,
+    dim: int,
+    theta: int,
+    scale: float,
+    current_sigma: float,
+    dype_config: DyPEConfig,
+) -> tuple[Tensor, Tensor]:
+    """Compute RoPE frequencies using YARN/NTK method.
+
+    Uses NTK-aware theta scaling for high-resolution support with
+    timestep-dependent DyPE modulation.
+
+    Args:
+        pos: Position tensor
+        dim: Embedding dimension
+        theta: RoPE base frequency
+        scale: Uniform scaling factor
+        current_sigma: Current noise level (1.0 = full noise, 0.0 = clean)
+        dype_config: DyPE configuration
+
+    Returns:
+        Tuple of (cos, sin) frequency tensors
+    """
+    assert dim % 2 == 0
+
+    device = pos.device
+    dtype = torch.float64 if device.type != "mps" else torch.float32
+
+    # NTK-aware theta scaling with DyPE modulation
+    if scale > 1.0:
+        ntk_alpha = scale ** (dim / (dim - 2))
+
+        # Apply timestep-dependent DyPE modulation
+        mscale = get_timestep_mscale(
+            scale=scale,
+            current_sigma=current_sigma,
+            dype_scale=dype_config.dype_scale,
+            dype_exponent=dype_config.dype_exponent,
+            dype_start_sigma=dype_config.dype_start_sigma,
+        )
+
+        # Modulate NTK alpha by mscale
+        modulated_alpha = 1.0 + (ntk_alpha - 1.0) * mscale
+        scaled_theta = theta * modulated_alpha
+    else:
+        scaled_theta = theta
+
+    freq_seq = torch.arange(0, dim, 2, dtype=dtype, device=device) / dim
+    freqs = 1.0 / (scaled_theta**freq_seq)
+
+    angles = torch.einsum("...n,d->...nd", pos.to(dtype), freqs)
+
+    cos = torch.cos(angles)
+    sin = torch.sin(angles)
+
+    return cos.to(pos.dtype), sin.to(pos.dtype)
+
+
+def compute_ntk_freqs(
+    pos: Tensor,
+    dim: int,
+    theta: int,
+    scale: float,
+) -> tuple[Tensor, Tensor]:
+    """Compute RoPE frequencies using NTK method.
+
+    Neural Tangent Kernel approach - continuous frequency scaling without
+    timestep dependency.
+
+    Args:
+        pos: Position tensor
+        dim: Embedding dimension
+        theta: RoPE base frequency
+        scale: Scaling factor
+
+    Returns:
+        Tuple of (cos, sin) frequency tensors
+    """
+    assert dim % 2 == 0
+
+    device = pos.device
+    dtype = torch.float64 if device.type != "mps" else torch.float32
+
+    # NTK scaling
+    scaled_theta = theta * (scale ** (dim / (dim - 2)))
+
+    freq_seq = torch.arange(0, dim, 2, dtype=dtype, device=device) / dim
+    freqs = 1.0 / (scaled_theta**freq_seq)
+
+    angles = torch.einsum("...n,d->...nd", pos.to(dtype), freqs)
+
+    cos = torch.cos(angles)
+    sin = torch.sin(angles)
+
+    return cos.to(pos.dtype), sin.to(pos.dtype)

invokeai/backend/flux/dype/embed.py

@@ -0,0 +1,116 @@
+"""DyPE-enhanced position embedding module."""
+
+import torch
+from torch import Tensor, nn
+
+from invokeai.backend.flux.dype.base import DyPEConfig
+from invokeai.backend.flux.dype.rope import rope_dype
+
+
+class DyPEEmbedND(nn.Module):
+    """N-dimensional position embedding with DyPE support.
+
+    This class replaces the standard EmbedND from FLUX with a DyPE-aware version
+    that dynamically scales position embeddings based on resolution and timestep.
+
+    The key difference from EmbedND:
+    - Maintains step state (current_sigma, target dimensions)
+    - Uses rope_dype() instead of rope() for frequency computation
+    - Applies timestep-dependent scaling for better high-resolution generation
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        theta: int,
+        axes_dim: list[int],
+        dype_config: DyPEConfig,
+    ):
+        """Initialize DyPE position embedder.
+
+        Args:
+            dim: Total embedding dimension (sum of axes_dim)
+            theta: RoPE base frequency
+            axes_dim: Dimension allocation per axis (e.g., [16, 56, 56] for FLUX)
+            dype_config: DyPE configuration
+        """
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        self.axes_dim = axes_dim
+        self.dype_config = dype_config
+
+        # Step state - updated before each denoising step
+        self._current_sigma: float = 1.0
+        self._target_height: int = 1024
+        self._target_width: int = 1024
+
+    def set_step_state(self, sigma: float, height: int, width: int) -> None:
+        """Update the step state before each denoising step.
+
+        This method should be called by the DyPE extension before each step
+        to update the current noise level and target dimensions.
+
+        Args:
+            sigma: Current noise level (timestep value, 1.0 = full noise)
+            height: Target image height in pixels
+            width: Target image width in pixels
+        """
+        self._current_sigma = sigma
+        self._target_height = height
+        self._target_width = width
+
+    def forward(self, ids: Tensor) -> Tensor:
+        """Compute position embeddings with DyPE scaling.
+
+        Args:
+            ids: Position indices tensor with shape (batch, seq_len, n_axes)
+                 For FLUX: n_axes=3 (time/channel, height, width)
+
+        Returns:
+            Position embedding tensor with shape (batch, 1, seq_len, dim)
+        """
+        n_axes = ids.shape[-1]
+
+        # Compute RoPE for each axis with DyPE scaling
+        embeddings = []
+        for i in range(n_axes):
+            axis_emb = rope_dype(
+                pos=ids[..., i],
+                dim=self.axes_dim[i],
+                theta=self.theta,
+                current_sigma=self._current_sigma,
+                target_height=self._target_height,
+                target_width=self._target_width,
+                dype_config=self.dype_config,
+            )
+            embeddings.append(axis_emb)
+
+        # Concatenate embeddings from all axes
+        emb = torch.cat(embeddings, dim=-3)
+
+        return emb.unsqueeze(1)
+
+    @classmethod
+    def from_embednd(
+        cls,
+        embed_nd: nn.Module,
+        dype_config: DyPEConfig,
+    ) -> "DyPEEmbedND":
+        """Create a DyPEEmbedND from an existing EmbedND.
+
+        This is a convenience method for patching an existing FLUX model.
+
+        Args:
+            embed_nd: Original EmbedND module from FLUX
+            dype_config: DyPE configuration
+
+        Returns:
+            New DyPEEmbedND with same parameters
+        """
+        return cls(
+            dim=embed_nd.dim,
+            theta=embed_nd.theta,
+            axes_dim=embed_nd.axes_dim,
+            dype_config=dype_config,
+        )

invokeai/backend/flux/dype/presets.py

@@ -0,0 +1,148 @@
+"""DyPE presets and automatic configuration."""
+
+from dataclasses import dataclass
+from typing import Literal
+
+from invokeai.backend.flux.dype.base import DyPEConfig
+
+# DyPE preset type - using Literal for proper frontend dropdown support
+DyPEPreset = Literal["off", "manual", "auto", "4k"]
+
+# Constants for preset values
+DYPE_PRESET_OFF: DyPEPreset = "off"
+DYPE_PRESET_MANUAL: DyPEPreset = "manual"
+DYPE_PRESET_AUTO: DyPEPreset = "auto"
+DYPE_PRESET_4K: DyPEPreset = "4k"
+
+# Human-readable labels for the UI
+DYPE_PRESET_LABELS: dict[str, str] = {
+    "off": "Off",
+    "manual": "Manual",
+    "auto": "Auto (>1536px)",
+    "4k": "4K Optimized",
+}
+
+
+@dataclass
+class DyPEPresetConfig:
+    """Preset configuration values."""
+
+    base_resolution: int
+    method: str
+    dype_scale: float
+    dype_exponent: float
+    dype_start_sigma: float
+
+
+# Predefined preset configurations
+DYPE_PRESETS: dict[DyPEPreset, DyPEPresetConfig] = {
+    DYPE_PRESET_4K: DyPEPresetConfig(
+        base_resolution=1024,
+        method="vision_yarn",
+        dype_scale=2.0,
+        dype_exponent=2.0,
+        dype_start_sigma=1.0,
+    ),
+}
+
+
+def get_dype_config_for_resolution(
+    width: int,
+    height: int,
+    base_resolution: int = 1024,
+    activation_threshold: int = 1536,
+) -> DyPEConfig | None:
+    """Automatically determine DyPE config based on target resolution.
+
+    FLUX can handle resolutions up to ~1.5x natively without significant artifacts.
+    DyPE is only activated when the resolution exceeds the activation threshold.
+
+    Args:
+        width: Target image width in pixels
+        height: Target image height in pixels
+        base_resolution: Native training resolution of the model (for scale calculation)
+        activation_threshold: Resolution threshold above which DyPE is activated
+
+    Returns:
+        DyPEConfig if DyPE should be enabled, None otherwise
+    """
+    max_dim = max(width, height)
+
+    if max_dim <= activation_threshold:
+        return None  # FLUX can handle this natively
+
+    # Calculate scaling factor based on base_resolution
+    scale = max_dim / base_resolution
+
+    # Dynamic parameters based on scaling
+    # Higher resolution = higher dype_scale, capped at 8.0
+    dynamic_dype_scale = min(2.0 * scale, 8.0)
+
+    return DyPEConfig(
+        enable_dype=True,
+        base_resolution=base_resolution,
+        method="vision_yarn",
+        dype_scale=dynamic_dype_scale,
+        dype_exponent=2.0,
+        dype_start_sigma=1.0,
+    )
+
+
+def get_dype_config_from_preset(
+    preset: DyPEPreset,
+    width: int,
+    height: int,
+    custom_scale: float | None = None,
+    custom_exponent: float | None = None,
+) -> DyPEConfig | None:
+    """Get DyPE configuration from a preset or custom values.
+
+    Args:
+        preset: The DyPE preset to use
+        width: Target image width
+        height: Target image height
+        custom_scale: Optional custom dype_scale (only used with 'manual' preset)
+        custom_exponent: Optional custom dype_exponent (only used with 'manual' preset)
+
+    Returns:
+        DyPEConfig if DyPE should be enabled, None otherwise
+    """
+    if preset == DYPE_PRESET_OFF:
+        return None
+
+    if preset == DYPE_PRESET_MANUAL:
+        # Manual mode - custom values can override defaults
+        max_dim = max(width, height)
+        scale = max_dim / 1024
+        dynamic_dype_scale = min(2.0 * scale, 8.0)
+        return DyPEConfig(
+            enable_dype=True,
+            base_resolution=1024,
+            method="vision_yarn",
+            dype_scale=custom_scale if custom_scale is not None else dynamic_dype_scale,
+            dype_exponent=custom_exponent if custom_exponent is not None else 2.0,
+            dype_start_sigma=1.0,
+        )
+
+    if preset == DYPE_PRESET_AUTO:
+        # Auto preset - custom values are ignored
+        return get_dype_config_for_resolution(
+            width=width,
+            height=height,
+            base_resolution=1024,
+            activation_threshold=1536,
+        )
+
+    # Use preset configuration (4K etc.) - custom values are ignored
+    preset_config = DYPE_PRESETS.get(preset)
+    if preset_config is None:
+        return None
+
+    return DyPEConfig(
+        enable_dype=True,
+        base_resolution=preset_config.base_resolution,
+        method=preset_config.method,
+        dype_scale=preset_config.dype_scale,
+        dype_exponent=preset_config.dype_exponent,
+        dype_start_sigma=preset_config.dype_start_sigma,
+    )

invokeai/backend/flux/dype/rope.py

@@ -0,0 +1,110 @@
+"""DyPE-enhanced RoPE (Rotary Position Embedding) functions."""
+
+import torch
+from einops import rearrange
+from torch import Tensor
+
+from invokeai.backend.flux.dype.base import (
+    DyPEConfig,
+    compute_ntk_freqs,
+    compute_vision_yarn_freqs,
+    compute_yarn_freqs,
+)
+
+
+def rope_dype(
+    pos: Tensor,
+    dim: int,
+    theta: int,
+    current_sigma: float,
+    target_height: int,
+    target_width: int,
+    dype_config: DyPEConfig,
+) -> Tensor:
+    """Compute RoPE with Dynamic Position Extrapolation.
+
+    This is the core DyPE function that replaces the standard rope() function.
+    It applies resolution-aware and timestep-aware scaling to position embeddings.
+
+    Args:
+        pos: Position indices tensor
+        dim: Embedding dimension per axis
+        theta: RoPE base frequency (typically 10000)
+        current_sigma: Current noise level (1.0 = full noise, 0.0 = clean)
+        target_height: Target image height in pixels
+        target_width: Target image width in pixels
+        dype_config: DyPE configuration
+
+    Returns:
+        Rotary position embedding tensor with shape suitable for FLUX attention
+    """
+    assert dim % 2 == 0
+
+    # Calculate scaling factors
+    base_res = dype_config.base_resolution
+    scale_h = target_height / base_res
+    scale_w = target_width / base_res
+    scale = max(scale_h, scale_w)
+
+    # If no scaling needed and DyPE disabled, use base method
+    if not dype_config.enable_dype or scale <= 1.0:
+        return _rope_base(pos, dim, theta)
+
+    # Select method and compute frequencies
+    method = dype_config.method
+
+    if method == "vision_yarn":
+        cos, sin = compute_vision_yarn_freqs(
+            pos=pos,
+            dim=dim,
+            theta=theta,
+            scale_h=scale_h,
+            scale_w=scale_w,
+            current_sigma=current_sigma,
+            dype_config=dype_config,
+        )
+    elif method == "yarn":
+        cos, sin = compute_yarn_freqs(
+            pos=pos,
+            dim=dim,
+            theta=theta,
+            scale=scale,
+            current_sigma=current_sigma,
+            dype_config=dype_config,
+        )
+    elif method == "ntk":
+        cos, sin = compute_ntk_freqs(
+            pos=pos,
+            dim=dim,
+            theta=theta,
+            scale=scale,
+        )
+    else:  # "base"
+        return _rope_base(pos, dim, theta)
+
+    # Construct rotation matrix from cos/sin
+    # Output shape: (batch, seq_len, dim/2, 2, 2)
+    out = torch.stack([cos, -sin, sin, cos], dim=-1)
+    out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
+
+    return out.to(dtype=pos.dtype, device=pos.device)
+
+
+def _rope_base(pos: Tensor, dim: int, theta: int) -> Tensor:
+    """Standard RoPE without DyPE scaling.
+
+    This matches the original rope() function from invokeai.backend.flux.math.
+    """
+    assert dim % 2 == 0
+
+    device = pos.device
+    dtype = torch.float64 if device.type != "mps" else torch.float32
+
+    scale = torch.arange(0, dim, 2, dtype=dtype, device=device) / dim
+    omega = 1.0 / (theta**scale)
+
+    out = torch.einsum("...n,d->...nd", pos.to(dtype), omega)
+    out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
+    out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
+
+    return out.to(dtype=pos.dtype, device=pos.device)