From f3d0b070f2f9926eefe81bc5623c3ce20510d8fc Mon Sep 17 00:00:00 2001 From: kijai <40791699+kijai@users.noreply.github.com> Date: Tue, 2 Jun 2026 10:35:58 +0300 Subject: [PATCH] Cleanup --- comfy/ldm/wan/model.py | 10 +++------- comfy/model_base.py | 2 +- comfy_extras/nodes_bernini.py | 15 ++++++++++----- 3 files changed, 14 insertions(+), 13 deletions(-) diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index 394b71d08..8e2116a6c 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -570,7 +570,7 @@ class WanModel(torch.nn.Module): full_ref = self.ref_conv(full_ref).flatten(2).transpose(1, 2) x = torch.concat((full_ref, x), dim=1) - # In-context reference streams (Bernini) + # In-context reference (Bernini) context_latents = kwargs.get("context_latents", None) main_len = x.shape[1] if context_latents is not None: @@ -650,10 +650,7 @@ class WanModel(torch.nn.Module): freqs = self.rope_embedder(img_ids).movedim(1, 2) - # In-context reference conditioning (e.g. Bernini): a non-zero source_id - # composes an extra rotation (over the full head_dim) into the spatial - # rope so streams sharing the same spatial coords stay distinct. source_id - # 0 is identity, so this is a no-op for all normal Wan usage. + # In-context reference: a non-zero source_id composes an extra rotation into the spatial rope if source_id: d = self.dim // self.num_heads pos = torch.tensor([[float(source_id)]], device=freqs.device, dtype=torch.float32) @@ -683,8 +680,7 @@ class WanModel(torch.nn.Module): freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options) - # In-context reference streams: one rope block per stream, each with its - # own source_id (1, 2, ...) so they stay distinct from the target (id 0). + # In-context reference: one rope block per stream, each with it's own source_id (1, 2, ...) to distinguish from the target (id 0). context_latents = kwargs.get("context_latents", None) if context_latents is not None: context_latents = [comfy.ldm.common_dit.pad_to_patch_size(lat, self.patch_size) for lat in context_latents] diff --git a/comfy/model_base.py b/comfy/model_base.py index f5224a840..3742062ce 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1525,7 +1525,7 @@ class WAN21(BaseModel): return out def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]): - # In-context streams slicing (Bernini) + # In-context cond slicing (Bernini) if cond_key == "context_latents" and isinstance(getattr(cond_value, "cond", None), list): dim = window.dim out = [] diff --git a/comfy_extras/nodes_bernini.py b/comfy_extras/nodes_bernini.py index c29af857e..4de3460bb 100644 --- a/comfy_extras/nodes_bernini.py +++ b/comfy_extras/nodes_bernini.py @@ -53,10 +53,15 @@ class BerniniConditioning(io.ComfyNode): io.Int.Input("height", default=480, min=16, max=8192, step=16), io.Int.Input("length", default=81, min=1, max=8192, step=4), io.Int.Input("batch_size", default=1, min=1, max=4096), - io.Image.Input("source_video", optional=True, tooltip="Source video to edit/restyle (task v2v or rv2v). Resized to width/height and trimmed to length. Acts as the edit base / canvas."), - io.Image.Input("reference_video", optional=True, tooltip="Moving content to composite into the source video (video insertion / ads2v), e.g. a clip to play on a screen. Kept at native aspect (long edge capped at ref_max_size), trimmed to length."), - io.Image.Input("reference_images", optional=True, tooltip="Reference image(s) injected as in-context tokens (task r2v or rv2v). Each is kept at its native aspect ratio, long edge capped at ref_max_size."), - io.Int.Input("ref_max_size", default=848, min=16, max=8192, step=16, optional=True), + io.Image.Input("source_video", optional=True, tooltip=( + "Source video to edit/restyle (task v2v or rv2v). Resized to width/height and trimmed to length. Acts as the edit base / canvas.")), + io.Image.Input("reference_video", optional=True, tooltip=( + "Moving content to composite into the source video (video insertion / ads2v)," + "e.g. a clip to play on a screen. Kept at native aspect (long edge capped at ref_max_size), trimmed to length.")), + io.Image.Input("reference_images", optional=True, tooltip=( + "Reference image(s) injected as in-context tokens (task r2v or rv2v). Each is kept at its native aspect ratio, long edge capped at ref_max_size.")), + io.Int.Input("ref_max_size", default=848, min=16, max=8192, step=16, optional=True, tooltip=( + "Max size for the long edge of reference_video and reference_images. Resized with preserved aspect ratio, snapped to 16px, and no upscaling.")), ], outputs=[ io.Conditioning.Output(display_name="positive"), @@ -72,7 +77,7 @@ class BerniniConditioning(io.ComfyNode): device=comfy.model_management.intermediate_device()) # Ordered list of condition streams -> source_id by list order: - # source_video (1), reference_video (2), reference_images (3, 4, ...). + # source_video (1), reference_video (2), reference_images (3, 4, ...). context = [] if source_video is not None: vid = comfy.utils.common_upscale(source_video[:length, :, :, :3].movedim(-1, 1), width, height, "area", "center").movedim(1, -1)