From d93133ee5350c14e776f0560a975903dab2a60f7 Mon Sep 17 00:00:00 2001 From: qqingzheng <2533221180@qq.com> Date: Sun, 8 Mar 2026 03:44:13 +0800 Subject: [PATCH] Refactor Helios integration and latent processing with new T2V support. --- comfy/latent_formats.py | 8 + comfy/ldm/helios/model.py | 188 ++++++++---- comfy/model_base.py | 41 ++- comfy/model_detection.py | 12 +- comfy/supported_models.py | 2 +- comfy_extras/nodes_helios.py | 536 +++++++++++++++++++++++++++-------- 6 files changed, 611 insertions(+), 176 deletions(-) diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 6a57bca1c..91db60ab5 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -783,3 +783,11 @@ class ZImagePixelSpace(ChromaRadiance): No VAE encoding/decoding — the model operates directly on RGB pixels. """ pass + +class Helios(Wan21): + """Helios video model latent format + + Helios uses the same latent format as Wan21 (same VAE architecture). + Inherits latents_mean, latents_std, and processing methods from Wan21. + """ + pass diff --git a/comfy/ldm/helios/model.py b/comfy/ldm/helios/model.py index 5ffc91129..6fd37b875 100644 --- a/comfy/ldm/helios/model.py +++ b/comfy/ldm/helios/model.py @@ -6,11 +6,12 @@ import torch.nn as nn from comfy.ldm.modules.attention import optimized_attention from comfy.ldm.flux.layers import EmbedND from comfy.ldm.flux.math import apply_rope1 -from comfy.ldm.wan.model import sinusoidal_embedding_1d, repeat_e +from comfy.ldm.wan.model import sinusoidal_embedding_1d import comfy.ldm.common_dit import comfy.patcher_extension + def pad_for_3d_conv(x, kernel_size): b, c, t, h, w = x.shape pt, ph, pw = kernel_size @@ -20,6 +21,10 @@ def pad_for_3d_conv(x, kernel_size): return torch.nn.functional.pad(x, (0, pad_w, 0, pad_h, 0, pad_t), mode="replicate") +def center_down_sample_3d(x, kernel_size): + return torch.nn.functional.avg_pool3d(x, kernel_size, stride=kernel_size) + + class OutputNorm(nn.Module): def __init__(self, dim, eps=1e-6, operation_settings={}): @@ -50,7 +55,8 @@ class OutputNorm(nn.Module): shift = shift.squeeze(2).to(hidden_states.device) scale = scale.squeeze(2).to(hidden_states.device) hidden_states = hidden_states[:, -original_context_length:, :] - hidden_states = self.norm(hidden_states) * (1 + scale) + shift + # Use float32 for numerical stability like diffusers + hidden_states = (self.norm(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states) return hidden_states @@ -272,36 +278,69 @@ class HeliosAttentionBlock(nn.Module): def forward(self, x, context, e, freqs, original_context_length=None, transformer_options={}): if e.ndim == 4: - e = (self.scale_shift_table.unsqueeze(0) + e.float()).chunk(6, dim=2) - e = [v.squeeze(2) for v in e] + shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = ( + self.scale_shift_table.unsqueeze(0).to(e.device) + e.float() + ).chunk(6, dim=2) + shift_msa = shift_msa.squeeze(2) + scale_msa = scale_msa.squeeze(2) + gate_msa = gate_msa.squeeze(2) + c_shift_msa = c_shift_msa.squeeze(2) + c_scale_msa = c_scale_msa.squeeze(2) + c_gate_msa = c_gate_msa.squeeze(2) else: - e = (self.scale_shift_table + e.float()).chunk(6, dim=1) + shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = ( + self.scale_shift_table.to(e.device) + e.float() + ).chunk(6, dim=1) # self-attn + # Use float32 for numerical stability like diffusers + # norm1 has elementwise_affine=False, so we can safely convert to float32 + norm_x = self.norm1(x.float()) + norm_x = (norm_x * (1 + scale_msa) + shift_msa).type_as(x) y = self.attn1( - torch.addcmul(repeat_e(e[0], x), self.norm1(x), 1 + repeat_e(e[1], x)), + norm_x, freqs=freqs, original_context_length=original_context_length, transformer_options=transformer_options, ) - x = torch.addcmul(x, y, repeat_e(e[2], x)) + x = (x.float() + y.float() * gate_msa).type_as(x) # cross-attn if self.guidance_cross_attn and original_context_length is not None: history_seq_len = x.shape[1] - original_context_length history_x, x_main = torch.split(x, [history_seq_len, original_context_length], dim=1) + # norm2 has elementwise_affine=True, manually do FP32LayerNorm behavior + norm_x_main = torch.nn.functional.layer_norm( + x_main.float(), + self.norm2.normalized_shape, + self.norm2.weight.to(x_main.device).float() if self.norm2.weight is not None else None, + self.norm2.bias.to(x_main.device).float() if self.norm2.bias is not None else None, + self.norm2.eps, + ).type_as(x_main) x_main = x_main + self.attn2( - self.norm2(x_main), + norm_x_main, context=context, transformer_options=transformer_options, ) x = torch.cat([history_x, x_main], dim=1) else: - x = x + self.attn2(self.norm2(x), context=context, transformer_options=transformer_options) + # norm2 has elementwise_affine=True, manually do FP32LayerNorm behavior + norm_x = torch.nn.functional.layer_norm( + x.float(), + self.norm2.normalized_shape, + self.norm2.weight.to(x.device).float() if self.norm2.weight is not None else None, + self.norm2.bias.to(x.device).float() if self.norm2.bias is not None else None, + self.norm2.eps, + ).type_as(x) + x = x + self.attn2(norm_x, context=context, transformer_options=transformer_options) # ffn - y = self.ffn(torch.addcmul(repeat_e(e[3], x), self.norm3(x), 1 + repeat_e(e[4], x))) - x = torch.addcmul(x, y, repeat_e(e[5], x)) + # Use float32 for numerical stability like diffusers + # norm3 has elementwise_affine=False, so we can safely convert to float32 + norm_x = self.norm3(x.float()) + norm_x = (norm_x * (1 + c_scale_msa) + c_shift_msa).type_as(x) + y = self.ffn(norm_x) + x = (x.float() + y.float() * c_gate_msa).type_as(x) return x @@ -358,7 +397,7 @@ class HeliosModel(torch.nn.Module): kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), - dtype=torch.float32, + dtype=operation_settings.get("dtype"), ) self.text_embedding = nn.Sequential( operations.Linear( @@ -411,7 +450,7 @@ class HeliosModel(torch.nn.Module): kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), - dtype=torch.float32, + dtype=operation_settings.get("dtype"), ) self.patch_mid = operations.Conv3d( in_channels, @@ -419,7 +458,7 @@ class HeliosModel(torch.nn.Module): kernel_size=tuple(2 * p for p in patch_size), stride=tuple(2 * p for p in patch_size), device=operation_settings.get("device"), - dtype=torch.float32, + dtype=operation_settings.get("dtype"), ) self.patch_long = operations.Conv3d( in_channels, @@ -427,7 +466,7 @@ class HeliosModel(torch.nn.Module): kernel_size=tuple(4 * p for p in patch_size), stride=tuple(4 * p for p in patch_size), device=operation_settings.get("device"), - dtype=torch.float32, + dtype=operation_settings.get("dtype"), ) # blocks @@ -592,7 +631,7 @@ class HeliosModel(torch.nn.Module): p_t, p_h, p_w = self.patch_size # embeddings - hidden_states = self.patch_embedding(hidden_states.float()).to(hidden_states.dtype) + hidden_states = self.patch_embedding(hidden_states) _, _, post_t, post_h, post_w = hidden_states.shape hidden_states = hidden_states.flatten(2).transpose(1, 2) @@ -614,7 +653,7 @@ class HeliosModel(torch.nn.Module): original_context_length = hidden_states.shape[1] if (latents_history_short is not None and indices_latents_history_short is not None and hasattr(self, "patch_short")): - x_short = self.patch_short(latents_history_short.float()).to(hidden_states.dtype) + x_short = self.patch_short(latents_history_short).to(hidden_states.dtype) _, _, ts, hs, ws = x_short.shape x_short = x_short.flatten(2).transpose(1, 2) f_short = self.rope_encode( @@ -633,44 +672,70 @@ class HeliosModel(torch.nn.Module): freqs = torch.cat([f_short, freqs], dim=1) if (latents_history_mid is not None and indices_latents_history_mid is not None and hasattr(self, "patch_mid")): - x_mid = self.patch_mid(pad_for_3d_conv(latents_history_mid, (2, 4, 4)).float()).to(hidden_states.dtype) + x_mid = self.patch_mid(pad_for_3d_conv(latents_history_mid, (2, 4, 4))).to(hidden_states.dtype) _, _, tm, hm, wm = x_mid.shape x_mid = x_mid.flatten(2).transpose(1, 2) + mid_t = indices_latents_history_mid.shape[1] + if ("hs" in locals()) and ("ws" in locals()): + mid_h, mid_w = hs, ws + else: + mid_h, mid_w = hm * 2, wm * 2 f_mid = self.rope_encode( - t=tm * self.patch_size[0], - h=hm * self.patch_size[1], - w=wm * self.patch_size[2], - steps_t=tm, - steps_h=hm, - steps_w=wm, + t=mid_t * self.patch_size[0], + h=mid_h * self.patch_size[1], + w=mid_w * self.patch_size[2], + steps_t=mid_t, + steps_h=mid_h, + steps_w=mid_w, device=x_mid.device, dtype=x_mid.dtype, transformer_options=transformer_options, frame_indices=indices_latents_history_mid, ) + f_mid = self._rope_downsample_3d(f_mid, (mid_t, mid_h, mid_w), (2, 2, 2)) + if f_mid.shape[1] != x_mid.shape[1]: + f_mid = f_mid[:, :x_mid.shape[1]] hidden_states = torch.cat([x_mid, hidden_states], dim=1) freqs = torch.cat([f_mid, freqs], dim=1) if (latents_history_long is not None and indices_latents_history_long is not None and hasattr(self, "patch_long")): - x_long = self.patch_long(pad_for_3d_conv(latents_history_long, (4, 8, 8)).float()).to(hidden_states.dtype) + x_long = self.patch_long(pad_for_3d_conv(latents_history_long, (4, 8, 8))).to(hidden_states.dtype) _, _, tl, hl, wl = x_long.shape x_long = x_long.flatten(2).transpose(1, 2) + long_t = indices_latents_history_long.shape[1] + if ("hs" in locals()) and ("ws" in locals()): + long_h, long_w = hs, ws + else: + long_h, long_w = hl * 4, wl * 4 f_long = self.rope_encode( - t=tl * self.patch_size[0], - h=hl * self.patch_size[1], - w=wl * self.patch_size[2], - steps_t=tl, - steps_h=hl, - steps_w=wl, + t=long_t * self.patch_size[0], + h=long_h * self.patch_size[1], + w=long_w * self.patch_size[2], + steps_t=long_t, + steps_h=long_h, + steps_w=long_w, device=x_long.device, dtype=x_long.dtype, transformer_options=transformer_options, frame_indices=indices_latents_history_long, ) + f_long = self._rope_downsample_3d(f_long, (long_t, long_h, long_w), (4, 4, 4)) + if f_long.shape[1] != x_long.shape[1]: + f_long = f_long[:, :x_long.shape[1]] hidden_states = torch.cat([x_long, hidden_states], dim=1) freqs = torch.cat([f_long, freqs], dim=1) history_context_length = hidden_states.shape[1] - original_context_length + mismatch = hidden_states.shape[1] != freqs.shape[1] + summary_key = ( + int(post_t), + int(post_h), + int(post_w), + int(original_context_length), + int(hidden_states.shape[1]), + int(freqs.shape[1]), + int(history_context_length), + ) if timestep.ndim == 0: timestep = timestep.unsqueeze(0) @@ -682,7 +747,7 @@ class HeliosModel(torch.nn.Module): e = self.time_embedding(sinusoidal_embedding_1d(self.freq_dim, timestep.flatten()).to(dtype=hidden_states.dtype)) e = e.reshape(batch_size, -1, e.shape[-1]) e0 = self.time_projection(e).unflatten(2, (6, self.dim)) - context = self.text_embedding(context) + context = self.text_embedding(context.to(dtype=hidden_states.dtype)) if self.zero_history_timestep and history_context_length > 0: timestep_t0 = torch.zeros((1, ), dtype=timestep.dtype, device=timestep.device) @@ -701,7 +766,7 @@ class HeliosModel(torch.nn.Module): e0 = e0.permute(0, 2, 1, 3) - for block in self.blocks: + for i_b, block in enumerate(self.blocks): hidden_states = block( hidden_states, context, @@ -710,35 +775,46 @@ class HeliosModel(torch.nn.Module): original_context_length=original_context_length, transformer_options=transformer_options, ) - hidden_states = self.norm_out(hidden_states, e, original_context_length) hidden_states = self.proj_out(hidden_states) - return self.unpatchify(hidden_states, (post_t, post_h, post_w)) def unpatchify(self, x, grid_sizes): - c = self.out_dim + """ + Unpatchify the output from proj_out back to video format. + + Args: + x: [batch, num_patches, out_dim * prod(patch_size)] + grid_sizes: (num_frames, height, width) in patch space + + Returns: + [batch, out_dim, num_frames, height, width] in pixel space + """ b = x.shape[0] - u = x[:, :math.prod(grid_sizes)].view(b, *grid_sizes, *self.patch_size, c) - u = torch.einsum("bfhwpqrc->bcfphqwr", u) - u = u.reshape(b, c, *[i * j for i, j in zip(grid_sizes, self.patch_size)]) - return u - - def load_state_dict(self, state_dict, strict=True, assign=False): - # Keep compatibility with reference diffusers key names. - remapped = {} - for k, v in state_dict.items(): - nk = k - nk = nk.replace("condition_embedder.time_embedder.linear_1.", "time_embedding.0.") - nk = nk.replace("condition_embedder.time_embedder.linear_2.", "time_embedding.2.") - nk = nk.replace("condition_embedder.time_proj.", "time_projection.1.") - nk = nk.replace("condition_embedder.text_embedder.linear_1.", "text_embedding.0.") - nk = nk.replace("condition_embedder.text_embedder.linear_2.", "text_embedding.2.") - nk = nk.replace("blocks.", "blocks.") - remapped[nk] = v - - return super().load_state_dict(remapped, strict=strict, assign=assign) - + post_t, post_h, post_w = grid_sizes + p_t, p_h, p_w = self.patch_size + + # Reshape: [B, T*H*W, out_dim*p_t*p_h*p_w] -> [B, T, H, W, p_t, p_h, p_w, out_dim] + # Use -1 to let PyTorch infer the channel dimension (out_dim) + hidden_states = x.reshape(b, post_t, post_h, post_w, p_t, p_h, p_w, -1) + + # Permute: [B, T, H, W, p_t, p_h, p_w, C] -> [B, C, T, p_t, H, p_h, W, p_w] + hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6) + + # Flatten patches: [B, C, T, p_t, H, p_h, W, p_w] -> [B, C, T*p_t, H*p_h, W*p_w] + output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3) + + return output + def _rope_downsample_3d(self, freqs, grid_sizes, kernel_size): + b, _, one, d, i2, j2 = freqs.shape + gt, gh, gw = grid_sizes + c = one * d * i2 * j2 + freqs_3d = freqs.reshape(b, gt, gh, gw, c).permute(0, 4, 1, 2, 3) + freqs_3d = pad_for_3d_conv(freqs_3d, kernel_size) + freqs_3d = center_down_sample_3d(freqs_3d, kernel_size) + dt, dh, dw = freqs_3d.shape[2:] + freqs_3d = freqs_3d.permute(0, 2, 3, 4, 1).reshape(b, dt * dh * dw, one, d, i2, j2) + return freqs_3d # Backward-compatible alias for existing integration points. HeliosTransformer3DModel = HeliosModel diff --git a/comfy/model_base.py b/comfy/model_base.py index 9bee3049a..d2d178b48 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1287,17 +1287,52 @@ class Helios(BaseModel): "latents_history_short", "latents_history_mid", "latents_history_long", + "helios_stage_sigmas", + "helios_stage_timesteps", ) for key in cond_keys: value = kwargs.get(key, None) if value is None: continue - if key.startswith("latents_"): - value = self.process_latent_in(value) - out[key] = comfy.conds.CONDRegular(value) + # Diffusers forwards Helios history latents without latent-format re-normalization. + # Keep raw history tensors to match transformer inputs across frameworks. + if key in ("helios_stage_sigmas", "helios_stage_timesteps"): + out[key] = comfy.conds.CONDConstant(value) + else: + out[key] = comfy.conds.CONDRegular(value) return out + def process_timestep(self, timestep, **kwargs): + stage_sigmas = kwargs.get("helios_stage_sigmas", None) + stage_timesteps = kwargs.get("helios_stage_timesteps", None) + if stage_sigmas is None or stage_timesteps is None: + return timestep + + if stage_sigmas.ndim > 1: + stage_sigmas = stage_sigmas[0] + if stage_timesteps.ndim > 1: + stage_timesteps = stage_timesteps[0] + + if stage_timesteps.numel() == 0 or stage_sigmas.numel() == 0: + return timestep + + if stage_sigmas.numel() == stage_timesteps.numel() + 1: + sigma_candidates = stage_sigmas[:-1] + else: + sigma_candidates = stage_sigmas[: stage_timesteps.numel()] + + if sigma_candidates.numel() == 0: + return timestep + + multiplier = float(getattr(self.model_sampling, "multiplier", 1000.0)) + sigma_in = timestep / multiplier + idx = torch.argmin(torch.abs(sigma_in.unsqueeze(-1) - sigma_candidates.unsqueeze(0)), dim=-1) + mapped = stage_timesteps[idx].to(dtype=timestep.dtype) + if mapped.dtype.is_floating_point: + mapped = torch.floor(mapped) + return mapped + class WAN21(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 7a130c02d..ae4f254ef 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -489,7 +489,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): return dit_config - if '{}condition_embedder.time_proj.weight'.format(key_prefix) in state_dict_keys and '{}patch_embedding.weight'.format(key_prefix) in state_dict_keys: # Helios + helios_required_keys = ( + '{}patch_mid.weight'.format(key_prefix), + '{}patch_long.weight'.format(key_prefix), + ) + if all(k in state_dict_keys for k in helios_required_keys): # Helios dit_config = {} dit_config["image_model"] = "helios" @@ -501,8 +505,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["patch_size"] = patch_size dit_config["in_channels"] = patch_weight.shape[1] dit_config["out_channels"] = out_proj.shape[0] // math.prod(patch_size) - dit_config["text_dim"] = state_dict['{}condition_embedder.text_embedder.linear_1.weight'.format(key_prefix)].shape[1] - dit_config["freq_dim"] = state_dict['{}condition_embedder.time_embedder.linear_1.weight'.format(key_prefix)].shape[1] + text_w = state_dict['{}text_embedding.0.weight'.format(key_prefix)] + time_w = state_dict['{}time_embedding.0.weight'.format(key_prefix)] + dit_config["text_dim"] = text_w.shape[1] + dit_config["freq_dim"] = time_w.shape[1] dit_config["num_layers"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.') dit_config["num_attention_heads"] = inner_dim // 128 dit_config["attention_head_dim"] = 128 diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 2035f25b8..b0fb3ce3d 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1143,7 +1143,7 @@ class Helios(supported_models_base.BASE): } unet_extra_config = {} - latent_format = latent_formats.Wan21 + latent_format = latent_formats.Helios memory_usage_factor = 1.8 supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] diff --git a/comfy_extras/nodes_helios.py b/comfy_extras/nodes_helios.py index 6c1fd7e20..13894082a 100644 --- a/comfy_extras/nodes_helios.py +++ b/comfy_extras/nodes_helios.py @@ -14,6 +14,9 @@ from typing_extensions import override from comfy_api.latest import ComfyExtension, io + + + def _parse_int_list(values, default): if values is None: return default @@ -72,15 +75,73 @@ def _extract_condition_value(conditioning, key): return None +def _process_latent_in_preserve_zero_frames(model, latent, valid_mask=None): + if latent is None or len(latent.shape) != 5: + return latent + if valid_mask is None: + raise ValueError("Helios requires `helios_history_valid_mask` for history latent conversion.") + vm = valid_mask + if not torch.is_tensor(vm): + vm = torch.tensor(vm, device=latent.device) + vm = vm.to(device=latent.device) + if vm.ndim == 2: + nonzero = vm.any(dim=0) + else: + nonzero = vm.reshape(-1) + nonzero = nonzero.bool() + + if nonzero.numel() == 0 or (not torch.any(nonzero)): + return latent + + if nonzero.shape[0] != latent.shape[2]: + # Keep behavior safe when mask length does not match temporal length. + nonzero = torch.zeros((latent.shape[2],), device=latent.device, dtype=torch.bool) + + converted = model.model.process_latent_in(latent) + out = latent.clone() + out[:, :, nonzero, :, :] = converted[:, :, nonzero, :, :] + return out + + def _upsample_latent_5d(latent, scale=2): b, c, t, h, w = latent.shape x = latent.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w) - x = comfy.utils.common_upscale(x, w * scale, h * scale, "nearest-exact", "disabled") + x = comfy.utils.common_upscale(x, w * scale, h * scale, "nearest", "disabled") x = x.reshape(b, t, c, h * scale, w * scale).permute(0, 2, 1, 3, 4) return x -def _sample_block_noise_like(latent, gamma, patch_size=(1, 2, 2)): +def _downsample_latent_5d_bilinear_x2(latent): + b, c, t, h, w = latent.shape + x = latent.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w) + x = comfy.utils.common_upscale(x, max(1, w // 2), max(1, h // 2), "bilinear", "disabled") * 2.0 + x = x.reshape(b, t, c, max(1, h // 2), max(1, w // 2)).permute(0, 2, 1, 3, 4) + return x + + +def _prepare_stage0_latent(batch, channels, frames, height, width, stage_count, add_noise, seed, dtype, layout, device): + """Prepare initial latent for stage 0 with optional noise""" + full_latent = torch.zeros((batch, channels, frames, height, width), dtype=dtype, layout=layout, device=device) + if add_noise: + full_latent = comfy.sample.prepare_noise(full_latent, seed).to(dtype) + + # Downsample to stage 0 resolution + stage_latent = full_latent + for _ in range(max(0, int(stage_count) - 1)): + stage_latent = _downsample_latent_5d_bilinear_x2(stage_latent) + return stage_latent + + +def _downsample_latent_for_stage0(latent, stage_count): + """Downsample latent to stage 0 resolution (like Diffusers does)""" + stage_latent = latent + for _ in range(max(0, int(stage_count) - 1)): + stage_latent = _downsample_latent_5d_bilinear_x2(stage_latent) + return stage_latent + + + +def _sample_block_noise_like(latent, gamma, patch_size=(1, 2, 2), generator=None, seed=None): b, c, t, h, w = latent.shape _, ph, pw = patch_size block_size = ph * pw @@ -88,13 +149,38 @@ def _sample_block_noise_like(latent, gamma, patch_size=(1, 2, 2)): cov = torch.eye(block_size, device=latent.device) * (1.0 + gamma) - torch.ones(block_size, block_size, device=latent.device) * gamma cov += torch.eye(block_size, device=latent.device) * 1e-6 - dist = torch.distributions.MultivariateNormal(torch.zeros(block_size, device=latent.device), covariance_matrix=cov) - block_number = b * c * t * max(1, h // ph) * max(1, w // pw) + h_blocks = h // ph + w_blocks = w // pw + block_number = b * c * t * h_blocks * w_blocks - noise = dist.sample((block_number,)) - noise = noise.view(b, c, t, max(1, h // ph), max(1, w // pw), ph, pw) - noise = noise.permute(0, 1, 2, 3, 5, 4, 6).reshape(b, c, t, max(1, h // ph) * ph, max(1, w // pw) * pw) - noise = noise[:, :, :, :h, :w] + if generator is not None: + # Exact Diffusers sampling path (MultivariateNormal.sample), while consuming + # from an explicit generator by temporarily swapping default RNG state. + with torch.random.fork_rng(devices=[latent.device] if latent.device.type == "cuda" else []): + if latent.device.type == "cuda": + torch.cuda.set_rng_state(generator.get_state(), device=latent.device) + else: + torch.random.set_rng_state(generator.get_state()) + dist = torch.distributions.MultivariateNormal( + torch.zeros(block_size, device=latent.device), + covariance_matrix=cov, + ) + noise = dist.sample((block_number,)) + if latent.device.type == "cuda": + generator.set_state(torch.cuda.get_rng_state(device=latent.device)) + else: + generator.set_state(torch.random.get_rng_state()) + elif seed is None: + dist = torch.distributions.MultivariateNormal(torch.zeros(block_size, device=latent.device), covariance_matrix=cov) + noise = dist.sample((block_number,)) + else: + # Use deterministic RNG when seed is provided (for cross-framework alignment). + with torch.random.fork_rng(devices=[latent.device] if latent.device.type == "cuda" else []): + torch.manual_seed(int(seed)) + dist = torch.distributions.MultivariateNormal(torch.zeros(block_size, device=latent.device), covariance_matrix=cov) + noise = dist.sample((block_number,)) + noise = noise.view(b, c, t, h_blocks, w_blocks, ph, pw) + noise = noise.permute(0, 1, 2, 3, 5, 4, 6).reshape(b, c, t, h, w) return noise @@ -144,8 +230,10 @@ def _helios_stage_tables(stage_count, stage_range, gamma, num_train_timesteps=10 tmax = min(float(sigmas[int(start_ratio * num_train_timesteps)].item() * num_train_timesteps), 999.0) tmin = float(sigmas[min(int(end_ratio * num_train_timesteps), num_train_timesteps - 1)].item() * num_train_timesteps) - timesteps_per_stage[i] = torch.linspace(tmax, tmin, num_train_timesteps) - sigmas_per_stage[i] = torch.linspace(0.999, 0.0, num_train_timesteps) + timesteps_per_stage[i] = torch.linspace(tmax, tmin, num_train_timesteps + 1)[:-1] + # Fixed: Use same sigma range [0.999, 0] for all stages like Diffusers + sigmas_per_stage[i] = torch.linspace(0.999, 0.0, num_train_timesteps + 1)[:-1] + return { "ori_start_sigmas": ori_start_sigmas, @@ -163,7 +251,8 @@ def _helios_stage_sigmas(stage_idx, stage_steps, stage_tables, is_distilled=Fals stage_steps = stage_steps * 2 if (is_amplify_first_stage and stage_idx == 0) else stage_steps stage_sigma_src = stage_tables["sigmas_per_stage"][stage_idx] - sigmas = torch.linspace(float(stage_sigma_src[0].item()), float(stage_sigma_src[-1].item()), stage_steps + 1) + sigmas = torch.linspace(float(stage_sigma_src[0].item()), float(stage_sigma_src[-1].item()), stage_steps) + sigmas = torch.cat([sigmas, torch.zeros(1, dtype=sigmas.dtype, device=sigmas.device)], dim=0) return sigmas @@ -213,23 +302,37 @@ def _build_cfg_zero_star_pre_cfg(stage_idx, zero_steps, use_zero_init): state["i"] += 1 return conds_out - noise_pred_text = conds_out[0] - noise_uncond = conds_out[1] + denoised_text = conds_out[0] # apply_model 返回的 denoised + denoised_uncond = conds_out[1] cfg = float(args.get("cond_scale", 1.0)) + x = args["input"] # 当前的 noisy latent + sigma = args["sigma"] # 当前的 sigma - positive_flat = noise_pred_text.view(noise_pred_text.shape[0], -1) - negative_flat = noise_uncond.view(noise_uncond.shape[0], -1) + # 关键修复：将 denoised 转换为 flow + # denoised = x - flow * sigma => flow = (x - denoised) / sigma + sigma_reshaped = sigma.reshape(sigma.shape[0], *([1] * (denoised_text.ndim - 1))) + sigma_safe = torch.clamp(sigma_reshaped, min=1e-8) + + flow_text = (x - denoised_text) / sigma_safe + flow_uncond = (x - denoised_uncond) / sigma_safe + + # 在 flow 空间做 CFG Zero Star + positive_flat = flow_text.reshape(flow_text.shape[0], -1) + negative_flat = flow_uncond.reshape(flow_uncond.shape[0], -1) alpha = _optimized_scale(positive_flat, negative_flat) - alpha = alpha.view(noise_pred_text.shape[0], *([1] * (noise_pred_text.ndim - 1))).to(noise_pred_text.dtype) + alpha = alpha.reshape(flow_text.shape[0], *([1] * (flow_text.ndim - 1))).to(flow_text.dtype) if stage_idx == 0 and state["i"] <= int(zero_steps) and bool(use_zero_init): - final = noise_pred_text * 0.0 + flow_final = flow_text * 0.0 else: - final = noise_uncond * alpha + cfg * (noise_pred_text - noise_uncond * alpha) + flow_final = flow_uncond * alpha + cfg * (flow_text - flow_uncond * alpha) + + # 将 flow 转回 denoised + denoised_final = x - flow_final * sigma_safe state["i"] += 1 # Return identical cond/uncond so downstream cfg_function keeps `final` unchanged. - return [final, final] + return [denoised_final, denoised_final] return pre_cfg_fn @@ -310,6 +413,8 @@ def _set_helios_history_values(positive, negative, history_latent, history_sizes latent = history_latent if latent is None or len(latent.shape) != 5: return positive, negative + if prefix_latent is not None and (latent.device != prefix_latent.device or latent.dtype != prefix_latent.dtype): + latent = latent.to(device=prefix_latent.device, dtype=prefix_latent.dtype) sizes = list(history_sizes) if len(sizes) != 3: @@ -342,13 +447,15 @@ def _set_helios_history_values(positive, negative, history_latent, history_sizes prefix = latent[:, :, :1] else: prefix = torch.zeros(latent.shape[0], latent.shape[1], 1, latent.shape[3], latent.shape[4], device=latent.device, dtype=latent.dtype) + if prefix.device != latents_history_short_base.device or prefix.dtype != latents_history_short_base.dtype: + prefix = prefix.to(device=latents_history_short_base.device, dtype=latents_history_short_base.dtype) latents_history_short = torch.cat([prefix, latents_history_short_base], dim=2) else: latents_history_short = latents_history_short_base - idx_short = torch.arange(latents_history_short.shape[2], device=latent.device, dtype=latent.dtype).unsqueeze(0).expand(latent.shape[0], -1) - idx_mid = torch.arange(latents_history_mid.shape[2], device=latent.device, dtype=latent.dtype).unsqueeze(0).expand(latent.shape[0], -1) - idx_long = torch.arange(latents_history_long.shape[2], device=latent.device, dtype=latent.dtype).unsqueeze(0).expand(latent.shape[0], -1) + idx_short = torch.arange(latents_history_short.shape[2], device=latent.device, dtype=torch.int64).unsqueeze(0).expand(latent.shape[0], -1) + idx_mid = torch.arange(latents_history_mid.shape[2], device=latent.device, dtype=torch.int64).unsqueeze(0).expand(latent.shape[0], -1) + idx_long = torch.arange(latents_history_long.shape[2], device=latent.device, dtype=torch.int64).unsqueeze(0).expand(latent.shape[0], -1) values = { "latents_history_short": latents_history_short, @@ -364,7 +471,7 @@ def _set_helios_history_values(positive, negative, history_latent, history_sizes return positive, negative -def _build_helios_indices(batch, history_sizes, keep_first_frame, hidden_frames, device, dtype): +def _build_helios_indices(batch, history_sizes, keep_first_frame, hidden_frames, device): sizes = list(history_sizes) if len(sizes) != 3: sizes = [16, 2, 1] @@ -373,13 +480,13 @@ def _build_helios_indices(batch, history_sizes, keep_first_frame, hidden_frames, if keep_first_frame: total = 1 + long_size + mid_size + short_base_size + hidden_frames - indices = torch.arange(total, device=device, dtype=dtype) + indices = torch.arange(total, device=device, dtype=torch.int64) splits = [1, long_size, mid_size, short_base_size, hidden_frames] indices_prefix, idx_long, idx_mid, idx_1x, idx_hidden = torch.split(indices, splits, dim=0) idx_short = torch.cat([indices_prefix, idx_1x], dim=0) else: total = long_size + mid_size + short_base_size + hidden_frames - indices = torch.arange(total, device=device, dtype=dtype) + indices = torch.arange(total, device=device, dtype=torch.int64) splits = [long_size, mid_size, short_base_size, hidden_frames] idx_long, idx_mid, idx_short, idx_hidden = torch.split(indices, splits, dim=0) @@ -450,7 +557,9 @@ class HeliosImageToVideo(io.ComfyNode): sizes = sorted([max(0, int(v)) for v in sizes], reverse=True) hist_len = max(1, sum(sizes)) history_latent = torch.zeros([batch_size, latent_channels, hist_len, latent.shape[-2], latent.shape[-1]], device=latent.device, dtype=latent.dtype) + history_valid_mask = torch.zeros((batch_size, hist_len), device=latent.device, dtype=torch.bool) image_latent_prefix = None + i2v_noise_gen = None if start_image is not None: image = comfy.utils.common_upscale(start_image[:1].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) @@ -459,20 +568,36 @@ class HeliosImageToVideo(io.ComfyNode): image_latent_prefix = img_latent[:, :, :1] if add_noise_to_image_latents: - g = torch.Generator(device=img_latent.device) - g.manual_seed(int(noise_seed)) + i2v_noise_gen = torch.Generator(device=img_latent.device) + i2v_noise_gen.manual_seed(int(noise_seed)) sigma = ( - torch.rand((img_latent.shape[0], 1, 1, 1, 1), device=img_latent.device, generator=g, dtype=img_latent.dtype) + torch.rand((img_latent.shape[0], 1, 1, 1, 1), device=img_latent.device, generator=i2v_noise_gen, dtype=img_latent.dtype) * (float(image_noise_sigma_max) - float(image_noise_sigma_min)) + float(image_noise_sigma_min) ) - image_latent_prefix = sigma * torch.randn_like(image_latent_prefix, generator=g) + (1.0 - sigma) * image_latent_prefix + image_latent_prefix = sigma * torch.randn_like(image_latent_prefix, generator=i2v_noise_gen) + (1.0 - sigma) * image_latent_prefix min_frames = max(1, (int(num_latent_frames_per_chunk) - 1) * 4 + 1) fake_video = image.repeat(min_frames, 1, 1, 1) fake_latents_full = vae.encode(fake_video) fake_latent = comfy.utils.repeat_to_batch_size(fake_latents_full[:, :, -1:], batch_size) + # Diffusers parity for I2V: + # when adding noise to image latents, fake_image_latents used for history are also noised. + if add_noise_to_image_latents: + if i2v_noise_gen is None: + i2v_noise_gen = torch.Generator(device=fake_latent.device) + i2v_noise_gen.manual_seed(int(noise_seed)) + # Keep backward compatibility with existing I2V node inputs: + # this node exposes only image sigma controls, while fake history + # latents follow the video-noise path in Diffusers. + fake_sigma = ( + torch.rand((fake_latent.shape[0], 1, 1, 1, 1), device=fake_latent.device, generator=i2v_noise_gen, dtype=fake_latent.dtype) + * (float(image_noise_sigma_max) - float(image_noise_sigma_min)) + + float(image_noise_sigma_min) + ) + fake_latent = fake_sigma * torch.randn_like(fake_latent, generator=i2v_noise_gen) + (1.0 - fake_sigma) * fake_latent history_latent[:, :, -1:] = fake_latent + history_valid_mask[:, -1] = True positive, negative = _set_helios_history_values(positive, negative, history_latent, sizes, keep_first_frame, prefix_latent=image_latent_prefix) return io.NodeOutput( @@ -482,6 +607,85 @@ class HeliosImageToVideo(io.ComfyNode): "samples": latent, "helios_history_latent": history_latent, "helios_image_latent_prefix": image_latent_prefix, + "helios_history_valid_mask": history_valid_mask, + }, + ) + + +class HeliosTextToVideo(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="HeliosTextToVideo", + category="conditioning/video_models", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae"), + io.Int.Input("width", default=640, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=384, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("length", default=132, min=1, max=nodes.MAX_RESOLUTION, step=4), + io.Int.Input("batch_size", default=1, min=1, max=4096), + io.String.Input("history_sizes", default="16,2,1", advanced=True), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute( + cls, + positive, + negative, + vae, + width, + height, + length, + batch_size, + history_sizes="16,2,1", + ) -> io.NodeOutput: + spacial_scale = vae.spacial_compression_encode() + latent_channels = vae.latent_channels + latent_t = ((length - 1) // 4) + 1 + + # Create zero latent as shape placeholder (noise will be generated in sampler) + latent = torch.zeros( + [batch_size, latent_channels, latent_t, height // spacial_scale, width // spacial_scale], + device=comfy.model_management.intermediate_device(), + ) + + sizes = _parse_int_list(history_sizes, [16, 2, 1]) + if len(sizes) != 3: + sizes = [16, 2, 1] + sizes = sorted([max(0, int(v)) for v in sizes], reverse=True) + hist_len = max(1, sum(sizes)) + # History latent starts as zeros (no history yet) + history_latent = torch.zeros( + [batch_size, latent_channels, hist_len, latent.shape[-2], latent.shape[-1]], + device=latent.device, + dtype=latent.dtype, + ) + history_valid_mask = torch.zeros((batch_size, hist_len), device=latent.device, dtype=torch.bool) + + positive, negative = _set_helios_history_values( + positive, + negative, + history_latent, + sizes, + False, + prefix_latent=None, + ) + return io.NodeOutput( + positive, + negative, + { + "samples": latent, + "helios_history_latent": history_latent, + "helios_image_latent_prefix": None, + "helios_history_valid_mask": history_valid_mask, }, ) @@ -544,6 +748,7 @@ class HeliosVideoToVideo(io.ComfyNode): sizes = sorted([max(0, int(v)) for v in sizes], reverse=True) hist_len = max(1, sum(sizes)) history_latent = torch.zeros([batch_size, latent_channels, hist_len, latent.shape[-2], latent.shape[-1]], device=latent.device, dtype=latent.dtype) + history_valid_mask = torch.zeros((batch_size, hist_len), device=latent.device, dtype=torch.bool) image_latent_prefix = None if video is not None: @@ -559,11 +764,14 @@ class HeliosVideoToVideo(io.ComfyNode): ) vid_latent = frame_sigmas * torch.randn_like(vid_latent, generator=g) + (1.0 - frame_sigmas) * vid_latent vid_latent = vid_latent[:, :, :hist_len] - if vid_latent.shape[2] < hist_len: - pad = vid_latent[:, :, -1:].repeat(1, 1, hist_len - vid_latent.shape[2], 1, 1) - vid_latent = torch.cat([vid_latent, pad], dim=2) vid_latent = comfy.utils.repeat_to_batch_size(vid_latent, batch_size) - history_latent = vid_latent + if vid_latent.shape[2] < hist_len: + keep_frames = hist_len - vid_latent.shape[2] + history_latent = torch.cat([history_latent[:, :, :keep_frames], vid_latent], dim=2) + history_valid_mask[:, keep_frames:] = True + else: + history_latent = vid_latent[:, :, -hist_len:] + history_valid_mask[:] = True image_latent_prefix = history_latent[:, :, :1] positive, negative = _set_helios_history_values(positive, negative, history_latent, sizes, keep_first_frame, prefix_latent=image_latent_prefix) @@ -574,6 +782,7 @@ class HeliosVideoToVideo(io.ComfyNode): "samples": latent, "helios_history_latent": history_latent, "helios_image_latent_prefix": image_latent_prefix, + "helios_history_valid_mask": history_valid_mask, }, ) @@ -625,25 +834,16 @@ class HeliosPyramidSampler(io.ComfyNode): io.Latent.Input("latent_image"), io.String.Input("pyramid_steps", default="10,10,10"), io.String.Input("stage_range", default="0,0.333333,0.666667,1"), - io.Boolean.Input("is_distilled", default=False), - io.Boolean.Input("is_amplify_first_stage", default=False), - io.Combo.Input("scheduler_mode", options=["euler", "unipc_bh2"]), + io.Boolean.Input("distilled", default=False), + io.Boolean.Input("amplify_first_stage", default=False), io.Float.Input("gamma", default=1.0 / 3.0, min=0.0001, max=10.0, step=0.0001, round=False), - io.Float.Input("shift", default=1.0, min=0.001, max=100.0, step=0.001, round=False, advanced=True), - io.Boolean.Input("use_dynamic_shifting", default=False, advanced=True), - io.Combo.Input("time_shift_type", options=["exponential", "linear"], advanced=True), - io.Int.Input("base_image_seq_len", default=256, min=1, max=65536, advanced=True), - io.Int.Input("max_image_seq_len", default=4096, min=1, max=65536, advanced=True), - io.Float.Input("base_shift", default=0.5, min=0.0, max=10.0, step=0.0001, round=False, advanced=True), - io.Float.Input("max_shift", default=1.15, min=0.0, max=10.0, step=0.0001, round=False, advanced=True), - io.Int.Input("num_train_timesteps", default=1000, min=10, max=100000, advanced=True), io.String.Input("history_sizes", default="16,2,1", advanced=True), io.Boolean.Input("keep_first_frame", default=True, advanced=True), io.Int.Input("num_latent_frames_per_chunk", default=9, min=1, max=256, advanced=True), - io.Boolean.Input("is_cfg_zero_star", default=False, advanced=True), + io.Boolean.Input("cfg_zero_star", default=True, advanced=True), io.Boolean.Input("use_zero_init", default=True, advanced=True), io.Int.Input("zero_steps", default=1, min=0, max=10000, advanced=True), - io.Boolean.Input("is_skip_first_chunk", default=False, advanced=True), + io.Boolean.Input("skip_first_chunk", default=False, advanced=True), ], outputs=[ io.Latent.Output(display_name="output"), @@ -663,33 +863,40 @@ class HeliosPyramidSampler(io.ComfyNode): latent_image, pyramid_steps, stage_range, - is_distilled, - is_amplify_first_stage, - scheduler_mode, + distilled, + amplify_first_stage, gamma, - shift, - use_dynamic_shifting, - time_shift_type, - base_image_seq_len, - max_image_seq_len, - base_shift, - max_shift, - num_train_timesteps, history_sizes, keep_first_frame, num_latent_frames_per_chunk, - is_cfg_zero_star, + cfg_zero_star, use_zero_init, zero_steps, - is_skip_first_chunk, + skip_first_chunk, ) -> io.NodeOutput: + # Keep these scheduler knobs internal (not exposed in node UI). + shift = 1.0 + num_train_timesteps = 1000 + # Keep dynamic shifting always on for Helios parity; not exposed in node UI. + use_dynamic_shifting = True + time_shift_type = "exponential" + base_image_seq_len = 256 + max_image_seq_len = 4096 + base_shift = 0.5 + max_shift = 1.15 + latent = latent_image.copy() latent_samples = comfy.sample.fix_empty_latent_channels(model, latent["samples"], latent.get("downscale_ratio_spacial", None)) + if not add_noise: + latent_samples = _process_latent_in_preserve_zero_frames(model, latent_samples) stage_steps = _parse_int_list(pyramid_steps, [10, 10, 10]) stage_steps = [max(1, int(s)) for s in stage_steps] stage_count = len(stage_steps) history_sizes_list = sorted([max(0, int(v)) for v in _parse_int_list(history_sizes, [16, 2, 1])], reverse=True) + # Diffusers parity: if not keeping first frame, fold prefix slot into short history size. + if not keep_first_frame and len(history_sizes_list) > 0: + history_sizes_list[-1] += 1 stage_range_values = _parse_float_list(stage_range, [0.0, 1.0 / 3.0, 2.0 / 3.0, 1.0]) if len(stage_range_values) != stage_count + 1: @@ -706,29 +913,41 @@ class HeliosPyramidSampler(io.ComfyNode): b, c, t, h, w = latent_samples.shape chunk_t = max(1, int(num_latent_frames_per_chunk)) chunk_count = max(1, (t + chunk_t - 1) // chunk_t) - low_scale = 2 ** max(0, stage_count - 1) - low_h = max(1, h // low_scale) - low_w = max(1, w // low_scale) - - base_latent = torch.zeros((b, c, chunk_t, low_h, low_w), dtype=latent_samples.dtype, layout=latent_samples.layout, device=latent_samples.device) - - if add_noise: - stage_latent = comfy.sample.prepare_noise(base_latent, noise_seed) - else: - stage_latent = torch.zeros_like(base_latent, device="cpu") - - stage_latent = stage_latent.to(base_latent.dtype).to(comfy.model_management.intermediate_device()) euler_sampler = comfy.samplers.KSAMPLER(_helios_euler_sample) + target_device = comfy.model_management.get_torch_device() + noise_gen = torch.Generator(device=target_device) + noise_gen.manual_seed(int(noise_seed)) + + image_latent_prefix = latent.get("helios_image_latent_prefix", None) + history_valid_mask = latent.get("helios_history_valid_mask", None) + if history_valid_mask is None: + raise ValueError("Helios sampler requires `helios_history_valid_mask` in latent input.") + history_from_latent_applied = False + if image_latent_prefix is not None: + image_latent_prefix = model.model.process_latent_in(image_latent_prefix) + if "helios_history_latent" in latent: + history_in = _process_latent_in_preserve_zero_frames(model, latent["helios_history_latent"], valid_mask=history_valid_mask) + positive, negative = _set_helios_history_values( + positive, + negative, + history_in, + history_sizes_list, + keep_first_frame, + prefix_latent=image_latent_prefix, + ) + history_from_latent_applied = True latents_history_short = _extract_condition_value(positive, "latents_history_short") latents_history_mid = _extract_condition_value(positive, "latents_history_mid") latents_history_long = _extract_condition_value(positive, "latents_history_long") - image_latent_prefix = latent.get("helios_image_latent_prefix", None) + if (not history_from_latent_applied) and latents_history_short is not None and latents_history_mid is not None and latents_history_long is not None: + raise ValueError("Helios requires `helios_history_latent` + `helios_history_valid_mask`; direct history conditioning is not supported.") if latents_history_short is None and "helios_history_latent" in latent: + history_in = _process_latent_in_preserve_zero_frames(model, latent["helios_history_latent"], valid_mask=history_valid_mask) positive, negative = _set_helios_history_values( positive, negative, - latent["helios_history_latent"], + history_in, history_sizes_list, keep_first_frame, prefix_latent=image_latent_prefix, @@ -740,18 +959,100 @@ class HeliosPyramidSampler(io.ComfyNode): x0_output = {} generated_chunks = [] if latents_history_short is not None and latents_history_mid is not None and latents_history_long is not None: - rolling_history = torch.cat([latents_history_long, latents_history_mid, latents_history_short], dim=2) + # Diffusers parity: `history_latents` storage does NOT include the keep_first_frame prefix slot. + # `latents_history_short` in conditioning may include [prefix + short_base], so strip prefix here. + short_base_size = history_sizes_list[-1] if len(history_sizes_list) > 0 else latents_history_short.shape[2] + if keep_first_frame and latents_history_short.shape[2] > short_base_size: + short_for_history = latents_history_short[:, :, -short_base_size:] + else: + short_for_history = latents_history_short + rolling_history = torch.cat([latents_history_long, latents_history_mid, short_for_history], dim=2) elif "helios_history_latent" in latent: rolling_history = latent["helios_history_latent"] + rolling_history = _process_latent_in_preserve_zero_frames(model, rolling_history, valid_mask=history_valid_mask) else: hist_len = max(1, sum(history_sizes_list)) rolling_history = torch.zeros((b, c, hist_len, h, w), device=latent_samples.device, dtype=latent_samples.dtype) - for chunk_idx in range(chunk_count): - if add_noise: - stage_latent = comfy.sample.prepare_noise(base_latent, noise_seed + chunk_idx).to(base_latent.dtype).to(comfy.model_management.intermediate_device()) + # Align with Diffusers behavior: when initial video latents are provided, seed history buffer + # with those latents before the first denoising chunk. + if not add_noise: + hist_len = max(1, sum(history_sizes_list)) + rolling_history = rolling_history.to(device=latent_samples.device, dtype=latent_samples.dtype) + video_latents = latent_samples + video_frames = video_latents.shape[2] + if video_frames < hist_len: + keep_frames = hist_len - video_frames + rolling_history = torch.cat([rolling_history[:, :, :keep_frames], video_latents], dim=2) else: - stage_latent = torch.zeros_like(base_latent, device=comfy.model_management.intermediate_device()) + rolling_history = video_latents[:, :, -hist_len:] + + # Keep history/prefix on the same device/dtype as denoising latents. + rolling_history = rolling_history.to(device=target_device, dtype=latent_samples.dtype) + if image_latent_prefix is not None: + image_latent_prefix = image_latent_prefix.to(device=target_device, dtype=latent_samples.dtype) + + for chunk_idx in range(chunk_count): + # Extract chunk from input latents + chunk_start = chunk_idx * chunk_t + chunk_end = min(chunk_start + chunk_t, t) + latent_chunk = latent_samples[:, :, chunk_start:chunk_end, :, :] + + # Prepare initial latent for this chunk + if add_noise: + # Diffusers parity: each chunk denoises a fixed latent window size. + # Keep chunk temporal length constant and crop only after all chunks. + noise_shape = ( + latent_samples.shape[0], + latent_samples.shape[1], + chunk_t, + latent_samples.shape[3], + latent_samples.shape[4], + ) + stage_latent = torch.randn(noise_shape, device=target_device, dtype=latent_samples.dtype, generator=noise_gen) + else: + # Use actual input latents; pad final short chunk to fixed size like Diffusers windowing. + stage_latent = latent_chunk.clone() + if stage_latent.shape[2] < chunk_t: + if stage_latent.shape[2] == 0: + stage_latent = torch.zeros( + ( + latent_samples.shape[0], + latent_samples.shape[1], + chunk_t, + latent_samples.shape[3], + latent_samples.shape[4], + ), + device=latent_samples.device, + dtype=latent_samples.dtype, + ) + else: + pad = stage_latent[:, :, -1:].repeat(1, 1, chunk_t - stage_latent.shape[2], 1, 1) + stage_latent = torch.cat([stage_latent, pad], dim=2) + + # Downsample to stage 0 resolution + for _ in range(max(0, int(stage_count) - 1)): + stage_latent = _downsample_latent_5d_bilinear_x2(stage_latent) + + # Keep stage latents on model device for parity with Diffusers scheduler/noise path. + stage_latent = stage_latent.to(target_device) + + # Diffusers parity: + # keep_first_frame=True and no image_latent_prefix on the first chunk + # should use an all-zero prefix frame, not history[:, :, :1]. + chunk_prefix = image_latent_prefix + if keep_first_frame and image_latent_prefix is None and chunk_idx == 0: + chunk_prefix = torch.zeros( + ( + rolling_history.shape[0], + rolling_history.shape[1], + 1, + rolling_history.shape[3], + rolling_history.shape[4], + ), + device=rolling_history.device, + dtype=rolling_history.dtype, + ) positive_chunk, negative_chunk = _set_helios_history_values( positive, @@ -759,37 +1060,28 @@ class HeliosPyramidSampler(io.ComfyNode): rolling_history, history_sizes_list, keep_first_frame, - prefix_latent=image_latent_prefix, + prefix_latent=chunk_prefix, ) latents_history_short = _extract_condition_value(positive_chunk, "latents_history_short") latents_history_mid = _extract_condition_value(positive_chunk, "latents_history_mid") latents_history_long = _extract_condition_value(positive_chunk, "latents_history_long") for stage_idx in range(stage_count): - if stage_idx > 0: - stage_latent = _upsample_latent_5d(stage_latent, scale=2) - - ori_sigma = 1.0 - float(stage_tables["ori_start_sigmas"][stage_idx]) - alpha = 1.0 / (math.sqrt(1.0 + (1.0 / gamma)) * (1.0 - ori_sigma) + ori_sigma) - beta = alpha * (1.0 - ori_sigma) / math.sqrt(gamma) - - noise = _sample_block_noise_like(stage_latent, gamma, patch_size=(1, 2, 2)).to(stage_latent) - stage_latent = alpha * stage_latent + beta * noise - + stage_latent = stage_latent.to(comfy.model_management.get_torch_device()) sigmas = _helios_stage_sigmas( stage_idx=stage_idx, stage_steps=stage_steps[stage_idx], stage_tables=stage_tables, - is_distilled=is_distilled, - is_amplify_first_stage=is_amplify_first_stage and chunk_idx == 0, - ).to(stage_latent.dtype) + is_distilled=distilled, + is_amplify_first_stage=amplify_first_stage and chunk_idx == 0, + ).to(device=stage_latent.device, dtype=torch.float32) timesteps = _helios_stage_timesteps( stage_idx=stage_idx, stage_steps=stage_steps[stage_idx], stage_tables=stage_tables, - is_distilled=is_distilled, - is_amplify_first_stage=is_amplify_first_stage and chunk_idx == 0, - ).to(stage_latent.dtype) + is_distilled=distilled, + is_amplify_first_stage=amplify_first_stage and chunk_idx == 0, + ).to(device=stage_latent.device, dtype=torch.float32) if use_dynamic_shifting: patch_size = (1, 2, 2) image_seq_len = (stage_latent.shape[-1] * stage_latent.shape[-2] * stage_latent.shape[-3]) // (patch_size[0] * patch_size[1] * patch_size[2]) @@ -800,10 +1092,24 @@ class HeliosPyramidSampler(io.ComfyNode): base_shift=base_shift, max_shift=max_shift, ) - sigmas = _time_shift(sigmas, mu=mu, sigma=1.0, mode=time_shift_type).to(stage_latent.dtype) + sigmas = _time_shift(sigmas, mu=mu, sigma=1.0, mode=time_shift_type).to(torch.float32) tmin = torch.min(timesteps) tmax = torch.max(timesteps) timesteps = tmin + sigmas[:-1] * (tmax - tmin) + else: + pass + + # Keep parity with Diffusers pipeline order: + # stage timesteps are computed before upsampling/renoise for stage > 0. + if stage_idx > 0: + stage_latent = _upsample_latent_5d(stage_latent, scale=2) + + ori_sigma = 1.0 - float(stage_tables["ori_start_sigmas"][stage_idx]) + alpha = 1.0 / (math.sqrt(1.0 + (1.0 / gamma)) * (1.0 - ori_sigma) + ori_sigma) + beta = alpha * (1.0 - ori_sigma) / math.sqrt(gamma) + + noise = _sample_block_noise_like(stage_latent, gamma, patch_size=(1, 2, 2), generator=noise_gen).to(stage_latent) + stage_latent = alpha * stage_latent + beta * noise indices_hidden_states, idx_short, idx_mid, idx_long = _build_helios_indices( batch=stage_latent.shape[0], @@ -811,7 +1117,6 @@ class HeliosPyramidSampler(io.ComfyNode): keep_first_frame=keep_first_frame, hidden_frames=stage_latent.shape[2], device=stage_latent.device, - dtype=stage_latent.dtype, ) positive_stage = node_helpers.conditioning_set_values(positive_chunk, {"indices_hidden_states": indices_hidden_states}) negative_stage = node_helpers.conditioning_set_values(negative_chunk, {"indices_hidden_states": indices_hidden_states}) @@ -831,19 +1136,22 @@ class HeliosPyramidSampler(io.ComfyNode): positive_stage = node_helpers.conditioning_set_values(positive_stage, values) negative_stage = node_helpers.conditioning_set_values(negative_stage, values) - cfg_use = 1.0 if is_distilled else cfg + stage_time_values = { + "helios_stage_sigmas": sigmas, + "helios_stage_timesteps": timesteps, + } + positive_stage = node_helpers.conditioning_set_values(positive_stage, stage_time_values) + negative_stage = node_helpers.conditioning_set_values(negative_stage, stage_time_values) - if stage_idx == 0 and add_noise: - noise = comfy.sample.prepare_noise(stage_latent, noise_seed + chunk_idx * 100 + stage_idx) - latent_start = torch.zeros_like(stage_latent) - else: - sigma0 = max(float(sigmas[0].item()), 1e-6) - noise = (stage_latent / sigma0).to("cpu") - latent_start = torch.zeros_like(stage_latent) + cfg_use = 1.0 if distilled else cfg + + sigma0 = max(float(sigmas[0].item()), 1e-6) + noise = stage_latent / sigma0 + latent_start = torch.zeros_like(stage_latent) stage_start_for_dmd = stage_latent.clone() - if is_distilled: + if distilled: sampler = comfy.samplers.KSAMPLER( _helios_dmd_sample, extra_options={ @@ -854,14 +1162,11 @@ class HeliosPyramidSampler(io.ComfyNode): }, ) else: - if scheduler_mode == "unipc_bh2": - sampler = comfy.samplers.ksampler("uni_pc_bh2") - else: - sampler = euler_sampler + sampler = euler_sampler callback = latent_preview.prepare_callback(model, sigmas.shape[-1] - 1, x0_output) stage_model = model - if is_cfg_zero_star and not is_distilled: + if cfg_zero_star and not distilled: stage_model = model.clone() stage_model.model_options = comfy.model_patcher.set_model_options_pre_cfg_function( stage_model.model_options, @@ -882,6 +1187,10 @@ class HeliosPyramidSampler(io.ComfyNode): disable_pbar=not comfy.utils.PROGRESS_BAR_ENABLED, seed=noise_seed + chunk_idx * 100 + stage_idx, ) + # sample_custom returns latent_format.process_out(samples); convert back to model-space + # so subsequent pyramid stages and history conditioning stay in the same latent space + # as Diffusers' internal denoising latents. + stage_latent = model.model.process_latent_in(stage_latent) if stage_latent.shape[-2] != h or stage_latent.shape[-1] != w: b2, c2, t2, h2, w2 = stage_latent.shape @@ -891,7 +1200,7 @@ class HeliosPyramidSampler(io.ComfyNode): stage_latent = stage_latent[:, :, :, :h, :w] generated_chunks.append(stage_latent) - if keep_first_frame and ((chunk_idx == 0 and image_latent_prefix is None) or (is_skip_first_chunk and chunk_idx == 1)): + if keep_first_frame and ((chunk_idx == 0 and image_latent_prefix is None) or (skip_first_chunk and chunk_idx == 1)): image_latent_prefix = stage_latent[:, :, :1] rolling_history = torch.cat([rolling_history, stage_latent.to(rolling_history.device, rolling_history.dtype)], dim=2) keep_hist = max(1, sum(history_sizes_list)) @@ -901,7 +1210,7 @@ class HeliosPyramidSampler(io.ComfyNode): out = latent.copy() out.pop("downscale_ratio_spacial", None) - out["samples"] = stage_latent + out["samples"] = model.model.process_latent_out(stage_latent) if "x0" in x0_output: x0_out = model.model.process_latent_out(x0_output["x0"].cpu()) @@ -917,6 +1226,7 @@ class HeliosExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ + HeliosTextToVideo, HeliosImageToVideo, HeliosVideoToVideo, HeliosHistoryConditioning,