Add SeedVR2 support (CORE-6) (#14110)

comfy/ldm/modules/attention.py

@@ -735,7 +735,86 @@ def attention_flash(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
         )
     return out
 
+def _var_attention_qkv(q, k, v, heads, skip_reshape):
+    if skip_reshape:
+        return q, k, v, q.shape[-1]
+    total_tokens, embed_dim = q.shape
+    head_dim = embed_dim // heads
+    return (
+        q.view(total_tokens, heads, head_dim),
+        k.view(k.shape[0], heads, head_dim),
+        v.view(v.shape[0], heads, head_dim),
+        head_dim,
+    )
 
+
+def _var_attention_output(out, heads, head_dim, skip_output_reshape):
+    if skip_output_reshape:
+        return out
+    return out.reshape(-1, heads * head_dim)
+
+
+def _use_blackwell_attention():
+    device = model_management.get_torch_device()
+    if device.type != "cuda":
+        return False
+    major, minor = torch.cuda.get_device_capability(device)
+    return (major, minor) >= (12, 0)
+
+
+def _validate_split_cu_seqlens(name, cu_seqlens, token_count):
+    if cu_seqlens.dtype not in (torch.int32, torch.int64):
+        raise ValueError(f"{name} must use an integer dtype")
+    if cu_seqlens.ndim != 1 or cu_seqlens.numel() < 2:
+        raise ValueError(f"{name} must be a 1D tensor with at least two offsets")
+    if cu_seqlens[0].item() != 0:
+        raise ValueError(f"{name} must start at 0")
+    if (cu_seqlens[1:] <= cu_seqlens[:-1]).any().item():
+        raise ValueError(f"{name} must be strictly increasing")
+    if cu_seqlens[-1].item() != token_count:
+        raise ValueError(f"{name} does not match token count")
+
+
+def _split_indices(cu_seqlens):
+    return cu_seqlens[1:-1].to(device="cpu", dtype=torch.long)
+
+
+def var_attention_optimized_split(q, k, v, heads, cu_seqlens_q, cu_seqlens_k, *args, skip_reshape=False, skip_output_reshape=False, **kwargs):
+    q, k, v, head_dim = _var_attention_qkv(q, k, v, heads, skip_reshape)
+
+    _validate_split_cu_seqlens("cu_seqlens_q", cu_seqlens_q, q.shape[0])
+    _validate_split_cu_seqlens("cu_seqlens_k", cu_seqlens_k, k.shape[0])
+    if cu_seqlens_k[-1].item() != v.shape[0]:
+        raise ValueError("cu_seqlens_k does not match v token count")
+
+    q_split_indices = _split_indices(cu_seqlens_q)
+    k_split_indices = _split_indices(cu_seqlens_k)
+    q_splits = torch.tensor_split(q, q_split_indices, dim=0)
+    k_splits = torch.tensor_split(k, k_split_indices, dim=0)
+    v_splits = torch.tensor_split(v, k_split_indices, dim=0)
+    if len(q_splits) != len(k_splits) or len(q_splits) != len(v_splits):
+        raise ValueError("cu_seqlens_q and cu_seqlens_k must describe the same sequence count")
+
+    out = []
+    for q_i, k_i, v_i in zip(q_splits, k_splits, v_splits):
+        q_i = q_i.permute(1, 0, 2).unsqueeze(0)
+        k_i = k_i.permute(1, 0, 2).unsqueeze(0)
+        v_i = v_i.permute(1, 0, 2).unsqueeze(0)
+        out_dtype = q_i.dtype
+        if optimized_attention is attention_sage and q_i.dtype not in (torch.float16, torch.bfloat16):
+            q_i = q_i.to(torch.bfloat16)
+            k_i = k_i.to(torch.bfloat16)
+            v_i = v_i.to(torch.bfloat16)
+        out_i = optimized_attention(q_i, k_i, v_i, heads, skip_reshape=True, skip_output_reshape=True)
+        if out_i.dtype != out_dtype:
+            out_i = out_i.to(out_dtype)
+        out.append(out_i.squeeze(0).permute(1, 0, 2))
+
+    out = torch.cat(out, dim=0)
+    return _var_attention_output(out, heads, head_dim, skip_output_reshape)
+
+
+optimized_var_attention = var_attention_optimized_split
 optimized_attention = attention_basic
 
 if model_management.sage_attention_enabled():
@@ -758,6 +837,8 @@ else:
         logging.info("Using sub quadratic optimization for attention, if you have memory or speed issues try using: --use-split-cross-attention")
         optimized_attention = attention_sub_quad
 
+logging.info("Using optimized_attention split-loop for variable-length attention")
+
 optimized_attention_masked = optimized_attention
 
 
@@ -773,6 +854,7 @@ if model_management.xformers_enabled():
 register_attention_function("pytorch", attention_pytorch)
 register_attention_function("sub_quad", attention_sub_quad)
 register_attention_function("split", attention_split)
+register_attention_function("var_attention_optimized_split", var_attention_optimized_split)
 
 
 def optimized_attention_for_device(device, mask=False, small_input=False):
@@ -1209,5 +1291,3 @@ class SpatialVideoTransformer(SpatialTransformer):
             x = self.proj_out(x)
         out = x + x_in
         return out
-
-

comfy/ldm/modules/diffusionmodules/model.py

@@ -13,6 +13,7 @@ if model_management.xformers_enabled_vae():
     import xformers
     import xformers.ops
 
+
 def torch_cat_if_needed(xl, dim):
     xl = [x for x in xl if x is not None and x.shape[dim] > 0]
     if len(xl) > 1:
@@ -22,7 +23,8 @@ def torch_cat_if_needed(xl, dim):
     else:
         return None
 
-def get_timestep_embedding(timesteps, embedding_dim):
+
+def get_timestep_embedding(timesteps, embedding_dim, flip_sin_to_cos=False, downscale_freq_shift=1):
     """
     This matches the implementation in Denoising Diffusion Probabilistic Models:
     From Fairseq.
@@ -33,11 +35,13 @@ def get_timestep_embedding(timesteps, embedding_dim):
     assert len(timesteps.shape) == 1
 
     half_dim = embedding_dim // 2
-    emb = math.log(10000) / (half_dim - 1)
+    emb = math.log(10000) / (half_dim - downscale_freq_shift)
     emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
     emb = emb.to(device=timesteps.device)
     emb = timesteps.float()[:, None] * emb[None, :]
     emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if flip_sin_to_cos:
+        emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)
     if embedding_dim % 2 == 1:  # zero pad
         emb = torch.nn.functional.pad(emb, (0,1,0,0))
     return emb

comfy/ldm/seedvr/color_fix.py

@@ -0,0 +1,340 @@
+import torch
+import torch.nn.functional as F
+from torch import Tensor
+
+from comfy.ldm.seedvr.model import safe_pad_operation
+from comfy.ldm.seedvr.vae import safe_interpolate_operation
+from comfy.ldm.seedvr.constants import (
+    CIELAB_DELTA,
+    CIELAB_KAPPA,
+    D65_WHITE_X,
+    D65_WHITE_Z,
+    WAVELET_DECOMP_LEVELS,
+)
+
+
+def wavelet_blur(image: Tensor, radius):
+    max_safe_radius = max(1, min(image.shape[-2:]) // 8)
+    if radius > max_safe_radius:
+        radius = max_safe_radius
+
+    num_channels = image.shape[1]
+
+    kernel_vals = [
+        [0.0625, 0.125, 0.0625],
+        [0.125,  0.25,  0.125],
+        [0.0625, 0.125, 0.0625],
+    ]
+    kernel = torch.tensor(kernel_vals, dtype=image.dtype, device=image.device)
+    kernel = kernel[None, None].repeat(num_channels, 1, 1, 1)
+
+    image = safe_pad_operation(image, (radius, radius, radius, radius), mode='replicate')
+    output = F.conv2d(image, kernel, groups=num_channels, dilation=radius)
+
+    return output
+
+def wavelet_decomposition(image: Tensor, levels: int = WAVELET_DECOMP_LEVELS):
+    high_freq = torch.zeros_like(image)
+
+    for i in range(levels):
+        radius = 2 ** i
+        low_freq = wavelet_blur(image, radius)
+        high_freq.add_(image).sub_(low_freq)
+        image = low_freq
+
+    return high_freq, low_freq
+
+def wavelet_reconstruction(content_feat: Tensor, style_feat: Tensor) -> Tensor:
+
+    if content_feat.shape != style_feat.shape:
+        # Resize style to match content spatial dimensions
+        if len(content_feat.shape) >= 3:
+            # safe_interpolate_operation handles FP16 conversion automatically
+            style_feat = safe_interpolate_operation(
+                style_feat,
+                size=content_feat.shape[-2:],
+                mode='bilinear',
+                align_corners=False
+            )
+
+    # Decompose both features into frequency components
+    content_high_freq, content_low_freq = wavelet_decomposition(content_feat)
+    del content_low_freq  # Free memory immediately
+
+    style_high_freq, style_low_freq = wavelet_decomposition(style_feat)
+    del style_high_freq  # Free memory immediately
+
+    if content_high_freq.shape != style_low_freq.shape:
+        style_low_freq = safe_interpolate_operation(
+            style_low_freq,
+            size=content_high_freq.shape[-2:],
+            mode='bilinear',
+            align_corners=False
+        )
+
+    content_high_freq.add_(style_low_freq)
+
+    return content_high_freq.clamp_(-1.0, 1.0)
+
+def _histogram_matching_channel(source: Tensor, reference: Tensor, device: torch.device) -> Tensor:
+    original_shape = source.shape
+
+    # Flatten
+    source_flat = source.flatten()
+    reference_flat = reference.flatten()
+
+    # Sort both arrays
+    source_sorted, source_indices = torch.sort(source_flat)
+    reference_sorted, _ = torch.sort(reference_flat)
+    del reference_flat
+
+    # Quantile mapping
+    n_source = len(source_sorted)
+    n_reference = len(reference_sorted)
+
+    if n_source == n_reference:
+        matched_sorted = reference_sorted
+    else:
+        # Interpolate reference to match source quantiles
+        source_quantiles = torch.linspace(0, 1, n_source, device=device)
+        ref_indices = (source_quantiles * (n_reference - 1)).long()
+        ref_indices.clamp_(0, n_reference - 1)
+        matched_sorted = reference_sorted[ref_indices]
+        del source_quantiles, ref_indices, reference_sorted
+
+    del source_sorted, source_flat
+
+    # Reconstruct using argsort (portable across CUDA/ROCm/MPS)
+    inverse_indices = torch.argsort(source_indices)
+    del source_indices
+    matched_flat = matched_sorted[inverse_indices]
+    del matched_sorted, inverse_indices
+
+    return matched_flat.reshape(original_shape)
+
+def _lab_to_rgb_batch(lab: Tensor, device: torch.device, matrix_inv: Tensor, epsilon: float, kappa: float) -> Tensor:
+    """Convert batch of CIELAB images to RGB color space."""
+    L, a, b = lab[:, 0], lab[:, 1], lab[:, 2]
+
+    # LAB to XYZ
+    fy = (L + 16.0) / 116.0
+    fx = a.div(500.0).add_(fy)
+    fz = fy - b / 200.0
+    del L, a, b
+
+    # XYZ transformation
+    x = torch.where(
+        fx > epsilon,
+        torch.pow(fx, 3.0),
+        fx.mul(116.0).sub_(16.0).div_(kappa)
+    )
+    y = torch.where(
+        fy > epsilon,
+        torch.pow(fy, 3.0),
+        fy.mul(116.0).sub_(16.0).div_(kappa)
+    )
+    z = torch.where(
+        fz > epsilon,
+        torch.pow(fz, 3.0),
+        fz.mul(116.0).sub_(16.0).div_(kappa)
+    )
+    del fx, fy, fz
+
+    # Apply D65 white point (in-place)
+    x.mul_(D65_WHITE_X)
+    # y *= 1.00000  # (no-op, skip)
+    z.mul_(D65_WHITE_Z)
+
+    xyz = torch.stack([x, y, z], dim=1)
+    del x, y, z
+
+    # Matrix multiplication: XYZ -> RGB
+    B, C, H, W = xyz.shape
+    xyz_flat = xyz.permute(0, 2, 3, 1).reshape(-1, 3)
+    del xyz
+
+    # Ensure dtype consistency for matrix multiplication
+    xyz_flat = xyz_flat.to(dtype=matrix_inv.dtype)
+    rgb_linear_flat = torch.matmul(xyz_flat, matrix_inv.T)
+    del xyz_flat
+
+    rgb_linear = rgb_linear_flat.reshape(B, H, W, 3).permute(0, 3, 1, 2)
+    del rgb_linear_flat
+
+    # Apply inverse gamma correction (delinearize)
+    mask = rgb_linear > 0.0031308
+    rgb = torch.where(
+        mask,
+        torch.pow(torch.clamp(rgb_linear, min=0.0), 1.0 / 2.4).mul_(1.055).sub_(0.055),
+        rgb_linear * 12.92
+    )
+    del mask, rgb_linear
+
+    return torch.clamp(rgb, 0.0, 1.0)
+
+def _rgb_to_lab_batch(rgb: Tensor, device: torch.device, matrix: Tensor, epsilon: float, kappa: float) -> Tensor:
+    """Convert batch of RGB images to CIELAB color space using D65 illuminant."""
+    # Apply sRGB gamma correction (linearize)
+    mask = rgb > 0.04045
+    rgb_linear = torch.where(
+        mask,
+        torch.pow((rgb + 0.055) / 1.055, 2.4),
+        rgb / 12.92
+    )
+    del mask
+
+    # Matrix multiplication: RGB -> XYZ
+    B, C, H, W = rgb_linear.shape
+    rgb_flat = rgb_linear.permute(0, 2, 3, 1).reshape(-1, 3)
+    del rgb_linear
+
+    # Ensure dtype consistency for matrix multiplication
+    rgb_flat = rgb_flat.to(dtype=matrix.dtype)
+    xyz_flat = torch.matmul(rgb_flat, matrix.T)
+    del rgb_flat
+
+    xyz = xyz_flat.reshape(B, H, W, 3).permute(0, 3, 1, 2)
+    del xyz_flat
+
+    # Normalize by D65 white point (in-place)
+    xyz[:, 0].div_(D65_WHITE_X)  # X
+    # xyz[:, 1] /= 1.00000  # Y (no-op, skip)
+    xyz[:, 2].div_(D65_WHITE_Z)  # Z
+
+    # XYZ to LAB transformation
+    epsilon_cubed = epsilon ** 3
+    mask = xyz > epsilon_cubed
+    f_xyz = torch.where(
+        mask,
+        torch.pow(xyz, 1.0 / 3.0),
+        xyz.mul(kappa).add_(16.0).div_(116.0)
+    )
+    del xyz, mask
+
+    # Extract channels and compute LAB
+    L = f_xyz[:, 1].mul(116.0).sub_(16.0)  # Lightness [0, 100]
+    a = (f_xyz[:, 0] - f_xyz[:, 1]).mul_(500.0)  # Green-Red [-128, 127]
+    b = (f_xyz[:, 1] - f_xyz[:, 2]).mul_(200.0)  # Blue-Yellow [-128, 127]
+    del f_xyz
+
+    return torch.stack([L, a, b], dim=1)
+
+def lab_color_transfer(
+    content_feat: Tensor,
+    style_feat: Tensor,
+    luminance_weight: float = 0.8
+) -> Tensor:
+    content_feat = wavelet_reconstruction(content_feat, style_feat)
+
+    if content_feat.shape != style_feat.shape:
+        style_feat = safe_interpolate_operation(
+            style_feat,
+            size=content_feat.shape[-2:],
+            mode='bilinear',
+            align_corners=False
+        )
+
+    device = content_feat.device
+
+    def ensure_float32_precision(c):
+        orig_dtype = c.dtype
+        c = c.float()
+        return c, orig_dtype
+    content_feat, original_dtype = ensure_float32_precision(content_feat)
+    style_feat, _ = ensure_float32_precision(style_feat)
+
+    rgb_to_xyz_matrix = torch.tensor([
+        [0.4124564, 0.3575761, 0.1804375],
+        [0.2126729, 0.7151522, 0.0721750],
+        [0.0193339, 0.1191920, 0.9503041]
+    ], dtype=torch.float32, device=device)
+
+    xyz_to_rgb_matrix = torch.tensor([
+        [ 3.2404542, -1.5371385, -0.4985314],
+        [-0.9692660,  1.8760108,  0.0415560],
+        [ 0.0556434, -0.2040259,  1.0572252]
+    ], dtype=torch.float32, device=device)
+
+    epsilon = CIELAB_DELTA
+    kappa = CIELAB_KAPPA
+
+    content_feat.add_(1.0).mul_(0.5).clamp_(0.0, 1.0)
+    style_feat.add_(1.0).mul_(0.5).clamp_(0.0, 1.0)
+
+    # Convert to LAB color space
+    content_lab = _rgb_to_lab_batch(content_feat, device, rgb_to_xyz_matrix, epsilon, kappa)
+    del content_feat
+
+    style_lab = _rgb_to_lab_batch(style_feat, device, rgb_to_xyz_matrix, epsilon, kappa)
+    del style_feat, rgb_to_xyz_matrix
+
+    # Match chrominance channels (a*, b*) for accurate color transfer
+    matched_a = _histogram_matching_channel(content_lab[:, 1], style_lab[:, 1], device)
+    matched_b = _histogram_matching_channel(content_lab[:, 2], style_lab[:, 2], device)
+
+    # Handle luminance with weighted blending
+    if luminance_weight < 1.0:
+        # Partially match luminance for better overall color accuracy
+        matched_L = _histogram_matching_channel(content_lab[:, 0], style_lab[:, 0], device)
+        # Blend: preserve some content L* for detail, adopt some style L* for color
+        result_L = content_lab[:, 0].mul(luminance_weight).add_(matched_L.mul(1.0 - luminance_weight))
+        del matched_L
+    else:
+        # Fully preserve content luminance
+        result_L = content_lab[:, 0]
+
+    del content_lab, style_lab
+
+    # Reconstruct LAB with corrected channels
+    result_lab = torch.stack([result_L, matched_a, matched_b], dim=1)
+    del result_L, matched_a, matched_b
+
+    # Convert back to RGB
+    result_rgb = _lab_to_rgb_batch(result_lab, device, xyz_to_rgb_matrix, epsilon, kappa)
+    del result_lab, xyz_to_rgb_matrix
+
+    # Convert back to [-1, 1] range (in-place)
+    result = result_rgb.mul_(2.0).sub_(1.0)
+    del result_rgb
+
+    result = result.to(original_dtype)
+
+    return result
+
+
+def wavelet_color_transfer(content_feat: Tensor, style_feat: Tensor) -> Tensor:
+    return wavelet_reconstruction(content_feat, style_feat)
+
+
+def adain_color_transfer(content_feat: Tensor, style_feat: Tensor, eps: float = 1e-5) -> Tensor:
+    if content_feat.shape != style_feat.shape:
+        style_feat = safe_interpolate_operation(
+            style_feat,
+            size=content_feat.shape[-2:],
+            mode='bilinear',
+            align_corners=False,
+        )
+
+    original_dtype = content_feat.dtype
+    content_feat = content_feat.float()
+    style_feat = style_feat.float()
+
+    b, c = content_feat.shape[:2]
+    content_flat = content_feat.reshape(b, c, -1)
+    style_flat = style_feat.reshape(b, c, -1)
+
+    content_mean = content_flat.mean(dim=2).reshape(b, c, 1, 1)
+    content_std = (content_flat.var(dim=2, correction=0) + eps).sqrt().reshape(b, c, 1, 1)
+    style_mean = style_flat.mean(dim=2).reshape(b, c, 1, 1)
+    style_std = (style_flat.var(dim=2, correction=0) + eps).sqrt().reshape(b, c, 1, 1)
+    del content_flat, style_flat
+
+    normalized = (content_feat - content_mean) / content_std
+    del content_mean, content_std
+    result = normalized * style_std + style_mean
+    del normalized, style_mean, style_std
+
+    result = result.clamp_(-1.0, 1.0)
+    if result.dtype != original_dtype:
+        result = result.to(original_dtype)
+    return result

comfy/ldm/seedvr/constants.py

@@ -0,0 +1,79 @@
+"""Named constants for the SeedVR2 integration, grouped by provenance.
+
+Provenance prefixes:
+- ``SEEDVR2_*``   - introduced by this integration (no external origin); rationale inline.
+- ``BYTEDANCE_*`` - ported from the official ByteDance-Seed/SeedVR release; each cites
+                    the upstream config/source path it was lifted from.
+- unprefixed standards (``ROPE_THETA``, ``CIELAB_*``, ``D65_*``) - published literature /
+                    ISO / CIE values; cite the standard.
+"""
+
+# --------------------------------------------------------------------------------------
+# A. Progressive-sampler chunk-size law  (SEEDVR2 - this integration's VRAM experiment)
+#    n_max(frames/chunk) = SEEDVR2_CHUNK_FRAMES_PER_GB * (free_GB - SEEDVR2_CHUNK_GB_MARGIN)
+#    rounded to the 4n+1 grid. Fit on 22 blocked-5090 cells, validated on a real RTX 4070
+#    (3b and 7b). Resolution-independent (the VAE tiling sets the wall, not the DiT).
+# --------------------------------------------------------------------------------------
+SEEDVR2_CHUNK_GB_MARGIN = 3        # fixed VRAM overhead before chunks scale (GiB)
+SEEDVR2_CHUNK_FRAMES_PER_GB = 4    # empirical slope: pixel frames admitted per free GiB
+
+# --------------------------------------------------------------------------------------
+# B. Fork heuristics  (SEEDVR2 - this integration)
+# --------------------------------------------------------------------------------------
+SEEDVR2_7B_VID_DIM = 3072          # runtime 3b-vs-7b sentinel; tested against vid_dim.
+                                   # (3072 is ByteDance's 7b vid_dim; the sentinel use is ours.)
+SEEDVR2_OOM_BACKOFF_DIVISOR = 2    # auto-chunk OOM retry: halve the chunk and retry.
+SEEDVR2_DTYPE_BYTES_FLOOR = 4      # per-element byte floor for memory math (fp32 worst case).
+SEEDVR2_7B_MLP_CHUNK = 8192        # 7b MLP token-chunk to bound peak VRAM.
+SEEDVR2_ROPE_PARTIAL_CHUNK_TOKENS = 4096  # partial-RoPE application token-chunk.
+SEEDVR2_LATENT_CHANNELS = 16       # SeedVR2 latent channel count (== BYTEDANCE latent_channels).
+SEEDVR2_COND_CHANNELS = 17         # conditioning channels = vid_in_channels(33) - latent(16).
+SEEDVR2_DEFAULT_TEMPORAL_SIZE = 16 # default VAE temporal tile when unset.
+
+# Color-correction memory model (fork tuning; per-frame VRAM estimate for chunk sizing)
+SEEDVR2_COLOR_MEM_HEADROOM = 0.75  # fraction of free VRAM usable per color-correction chunk.
+SEEDVR2_LAB_SCALE_MULTIPLIER = 13  # per-frame byte multiplier, LAB path.
+SEEDVR2_WAVELET_SCALE_MULTIPLIER = 10  # per-frame byte multiplier, wavelet path.
+SEEDVR2_ADAIN_SCALE_MULTIPLIER = 6     # per-frame byte multiplier, AdaIN path.
+
+# --------------------------------------------------------------------------------------
+# C. ByteDance config / source  (BYTEDANCE - cite ByteDance-Seed/SeedVR)
+# --------------------------------------------------------------------------------------
+BYTEDANCE_VAE_SCALING_FACTOR = 0.9152   # configs_3b/main.yaml:57 (scaling_factor); latent denorm.
+BYTEDANCE_VAE_SHIFTING_FACTOR = 0.0     # infer.py (shifting_factor default); latent denorm shift.
+BYTEDANCE_VAE_CONV_MEM_GIB = 0.5        # configs_3b/main.yaml:54 (conv_max_mem).
+BYTEDANCE_VAE_NORM_MEM_GIB = 0.5        # configs_3b/main.yaml:55 (norm_max_mem).
+BYTEDANCE_LOGVAR_CLAMP_MIN = -30.0      # video_vae_v3/modules/types.py:28.
+BYTEDANCE_LOGVAR_CLAMP_MAX = 20.0       # video_vae_v3/modules/types.py:28.
+BYTEDANCE_GN_CHUNKS_FP16 = 4            # causal_inflation_lib.py:351 (GroupNorm chunk count, fp16).
+BYTEDANCE_GN_CHUNKS_FP32 = 2            # causal_inflation_lib.py:351 (GroupNorm chunk count, fp32).
+BYTEDANCE_CONTIGUOUS_BATCH_THRESHOLD = 64  # attn_video_vae.py:308 (force .contiguous() above this b*t).
+BYTEDANCE_BLOCK_OUT_CHANNELS = (128, 256, 512, 512)  # s8_c16_t4_inflation_sd3.yaml:7-11.
+BYTEDANCE_SLICING_SAMPLE_MIN = 4        # s8_c16_t4_inflation_sd3.yaml:22 (slicing_sample_min_size).
+BYTEDANCE_VAE_TEMPORAL_DOWNSAMPLE = 4   # infer.py:230 (temporal_downsample_factor); the 4n+1 factor.
+BYTEDANCE_VAE_SPATIAL_DOWNSAMPLE = 8    # infer.py:231 (spatial_downsample_factor).
+BYTEDANCE_SCHEDULE_T = 1000.0           # configs_3b/main.yaml:65 (schedule.T); timestep range.
+BYTEDANCE_SPATIAL_DIVISOR = 16          # inference_seedvr2_3b.py:241 (DivisibleCrop((16,16))).
+BYTEDANCE_720P_REF_AREA = 45 * 80       # dit_v2/window.py:32 (720p reference area for window scaling).
+BYTEDANCE_MAX_TEMPORAL_WINDOW = 30      # dit_v2/window.py:35 (max temporal window frames).
+BYTEDANCE_ROPE_MAX_FREQ = 256           # dit_v2/rope.py:31 (pixel-RoPE max frequency).
+BYTEDANCE_SINUSOIDAL_DIM = 256          # dit_3b/nadit.py:120 (timestep sinusoidal embed dim).
+# Resolution-dependent timestep-shift linear fits: (x1, y1, x2, y2) for get_lin_function.
+BYTEDANCE_IMG_SHIFT_FIT = (256 * 256, 1.0, 1024 * 1024, 3.2)            # infer.py:242.
+BYTEDANCE_VID_SHIFT_FIT = (256 * 256 * 37, 1.0, 1280 * 720 * 145, 5.0)  # infer.py:243.
+
+# --------------------------------------------------------------------------------------
+# D. Published standards (cite the literature)
+# --------------------------------------------------------------------------------------
+ROPE_THETA = 10000   # RoPE base; Su et al., "RoFormer", arXiv:2104.09864.
+
+# CIELAB f(t) piecewise constants and D65 white point (CIE 15 colorimetry; CIE D65).
+CIELAB_DELTA = 6.0 / 29.0          # CIE 15 (delta).
+CIELAB_KAPPA = (29.0 / 3.0) ** 3   # CIE 15 (kappa).
+D65_WHITE_X = 0.95047              # CIE D65 standard illuminant Xn (Yn = 1).
+D65_WHITE_Z = 1.08883              # CIE D65 standard illuminant Zn.
+WAVELET_DECOMP_LEVELS = 5          # wavelet color-fix decomposition depth (GIMP/Krita; StableSR).
+
+# NOTE: the sRGB<->XYZ D65 3x3 matrices (IEC 61966-2-1) remain inline in the color code and
+# are named (SRGB_TO_XYZ_D65 / XYZ_TO_SRGB_D65) during the color-module extraction, where the
+# exact existing coefficients move verbatim rather than being retyped here.