v1.3.0 New Nodes: LTX Director and LTX Director Guide

__init__.py

@@ -3,26 +3,43 @@ from .multi_image_loader import MultiImageLoader
 from .ltx_sequencer import LTXSequencer
 from .speech_length_calculator import SpeechLengthCalculator
 from .load_audio_ui import LoadAudioUI
-from .load_video_ui import LoadVideoUI  # Imported new node
+from .load_video_ui import LoadVideoUI
+from .ltx_director import LTXDirector
+from .ltx_director_guide import LTXDirectorGuide
+from comfy_api.latest import ComfyExtension, io
+from typing_extensions import override
+
+class PromptRelay(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            LTXDirector,
+            LTXDirectorGuide
+        ]
+
+async def comfy_entrypoint() -> PromptRelay:
+    return PromptRelay()
     
-# Register the node classes
 NODE_CLASS_MAPPINGS = {
     "LTXKeyframer": LTXKeyframer,
     "MultiImageLoader": MultiImageLoader,
     "LTXSequencer": LTXSequencer,
     "SpeechLengthCalculator": SpeechLengthCalculator,
     "LoadAudioUI": LoadAudioUI,
-    "LoadVideoUI": LoadVideoUI  # Registered new node
+    "LoadVideoUI": LoadVideoUI,
+    "LTXDirector": LTXDirector,
+    "LTXDirectorGuide": LTXDirectorGuide,
 }
 
-# Provide clean display names for the ComfyUI interface
 NODE_DISPLAY_NAME_MAPPINGS = {
     "LTXKeyframer": "LTX Keyframer",
     "MultiImageLoader": "Multi Image Loader",
     "LTXSequencer": "LTX Sequencer",
     "SpeechLengthCalculator": "Speech Length Calculator",
     "LoadAudioUI": "Load Audio UI",
-    "LoadVideoUI": "Load Video UI"  # Registered display name
+    "LoadVideoUI": "Load Video UI",
+    "LTXDirector": "LTX Director",
+    "LTXDirectorGuide": "LTX Director Guide",
 }
 
 WEB_DIRECTORY = "./js"

js/ltx_director_guide.js

@@ -0,0 +1,17 @@
+import { app } from "../../scripts/app.js";
+
+// LTX Director Guide is a pure pass-through processor node.
+// All configuration (images, insert frames, strengths) comes from
+// the guide_data output of Prompt Relay Encode (Timeline).
+// No dynamic widgets or sync logic needed.
+app.registerExtension({
+    name: "Comfy.LTXDirectorGuide",
+    async nodeCreated(node) {
+        if (node.comfyClass !== "LTXDirectorGuide") return;
+        // Nothing to initialize — the node has no configurable widgets.
+    },
+});
+
+        // Nothing to initialize — the node has no configurable widgets.
+    },
+});

ltx_director.py

@@ -0,0 +1,643 @@
+import logging
+import json
+import base64
+import io as _io
+import math
+
+import numpy as np
+import torch
+import av
+from PIL import Image
+
+import os
+import folder_paths
+import comfy.model_management
+
+from comfy_api.latest import io
+
+from .prompt_relay import (
+    get_raw_tokenizer,
+    map_token_indices,
+    build_segments,
+    create_mask_fn,
+    distribute_segment_lengths,
+)
+
+from .patches import detect_model_type, apply_patches
+
+log = logging.getLogger(__name__)
+
+# Custom socket type shared with LTXSequencer
+GuideData = io.Custom("GUIDE_DATA")
+
+
+def _load_image_tensor(seg: dict) -> torch.Tensor:
+    """Decode an image from the ComfyUI input folder (if imageFile provided) or fallback to base64
+    to a ComfyUI-style image tensor of shape [1, H, W, 3], float32 in [0, 1]."""
+    if seg.get("imageFile"):
+        file_path = os.path.join(folder_paths.get_input_directory(), seg["imageFile"])
+        if os.path.exists(file_path):
+            img = Image.open(file_path).convert("RGB")
+            arr = np.array(img, dtype=np.float32) / 255.0
+            return torch.from_numpy(arr).unsqueeze(0)
+
+    b64_str = seg.get("imageB64", "")
+    if not b64_str or b64_str.startswith("/view?"):
+        return torch.zeros((1, 512, 512, 3), dtype=torch.float32)
+
+    if "," in b64_str:
+        b64_str = b64_str.split(",", 1)[1]
+    
+    try:
+        img_bytes = base64.b64decode(b64_str)
+        img = Image.open(_io.BytesIO(img_bytes)).convert("RGB")
+        arr = np.array(img, dtype=np.float32) / 255.0
+        return torch.from_numpy(arr).unsqueeze(0)
+    except:
+        return torch.zeros((1, 512, 512, 3), dtype=torch.float32)
+
+
+def _resize_image(tensor: torch.Tensor, target_w: int, target_h: int, method: str, divisible_by: int) -> torch.Tensor:
+    """Resize a [1, H, W, 3] float32 tensor to target dimensions using the given method,
+    then snap the final dimensions to be divisible by `divisible_by`."""
+    from PIL import Image as _PilImage
+    import torchvision.transforms.functional as TF
+
+    def snap(val, div):
+        return max(div, (val // div) * div)
+
+    tw = snap(target_w, divisible_by)
+    th = snap(target_h, divisible_by)
+
+    img_np = (tensor[0].cpu().numpy() * 255.0).clip(0, 255).astype(np.uint8)
+    pil = _PilImage.fromarray(img_np)
+    src_w, src_h = pil.size
+
+    if method == "stretch to fit":
+        resized = pil.resize((tw, th), _PilImage.LANCZOS)
+
+    elif method == "maintain aspect ratio":
+        ratio = min(tw / src_w, th / src_h)
+        new_w = int(src_w * ratio)
+        new_h = int(src_h * ratio)
+        new_w = snap(new_w, divisible_by)
+        new_h = snap(new_h, divisible_by)
+        resized = pil.resize((new_w, new_h), _PilImage.LANCZOS)
+
+    elif method == "pad":
+        ratio = min(tw / src_w, th / src_h)
+        new_w = snap(int(src_w * ratio), divisible_by)
+        new_h = snap(int(src_h * ratio), divisible_by)
+        inner = pil.resize((new_w, new_h), _PilImage.LANCZOS)
+        resized = _PilImage.new("RGB", (tw, th), (0, 0, 0))
+        resized.paste(inner, ((tw - new_w) // 2, (th - new_h) // 2))
+
+    elif method == "crop":
+        ratio = max(tw / src_w, th / src_h)
+        new_w = int(src_w * ratio)
+        new_h = int(src_h * ratio)
+        inner = pil.resize((new_w, new_h), _PilImage.LANCZOS)
+        left = (new_w - tw) // 2
+        top = (new_h - th) // 2
+        resized = inner.crop((left, top, left + tw, top + th))
+
+    else:
+        resized = pil.resize((tw, th), _PilImage.LANCZOS)
+
+    arr = np.array(resized, dtype=np.float32) / 255.0
+    return torch.from_numpy(arr).unsqueeze(0)
+
+
+def _compress_image(tensor: torch.Tensor, crf: int) -> torch.Tensor:
+    """Apply H.264 compression artefacts to a [1, H, W, 3] float32 tensor (ComfyUI image format).
+    crf=0 means no compression. Uses PyAV to encode/decode a single frame in-memory."""
+    if crf == 0:
+        return tensor
+    img = tensor[0]  # [H, W, 3]
+    # Dimensions must be even for H.264
+    h = (img.shape[0] // 2) * 2
+    w = (img.shape[1] // 2) * 2
+    img_np = (img[:h, :w] * 255.0).byte().cpu().numpy()  # uint8 [H, W, 3]
+
+    try:
+        buf = _io.BytesIO()
+        container = av.open(buf, mode="w", format="mp4")
+        stream = container.add_stream("libx264", rate=1)
+        stream.width = w
+        stream.height = h
+        stream.pix_fmt = "yuv420p"
+        stream.options = {"crf": str(crf), "preset": "ultrafast"}
+        frame = av.VideoFrame.from_ndarray(img_np, format="rgb24")
+        for pkt in stream.encode(frame):
+            container.mux(pkt)
+        for pkt in stream.encode(None):
+            container.mux(pkt)
+        container.close()
+
+        buf.seek(0)
+        container_r = av.open(buf, mode="r")
+        decoded = None
+        for frame_r in container_r.decode(video=0):
+            decoded = frame_r.to_ndarray(format="rgb24")  # [H, W, 3]
+            break
+        container_r.close()
+
+        if decoded is None:
+            return tensor
+        arr = torch.from_numpy(decoded.astype(np.float32) / 255.0).to(tensor.device, tensor.dtype)
+        # Re-embed into original tensor shape (may have been cropped by even-rounding)
+        out = tensor.clone()
+        out[0, :h, :w] = arr
+        return out
+    except Exception as e:
+        log.warning("[PromptRelay] img_compression encode/decode failed: %s", e)
+        return tensor
+
+
+def _build_combined_audio(timeline_data_str: str, duration_frames: int, frame_rate: float) -> dict:
+    """Parses timeline JSON, loads/trims audio directly from memory using PyAV, 
+    and aligns to a global timeline yielding ComfyUI's format.
+    Output length explicitly mimics the timeline's duration_frames length."""
+    target_sr = 44100
+    total_samples = max(1, int(math.ceil(duration_frames / frame_rate * target_sr)))
+    empty_audio = {"waveform": torch.zeros((1, 2, total_samples), dtype=torch.float32), "sample_rate": target_sr}
+
+    if not timeline_data_str:
+        return empty_audio
+
+    try:
+        data = json.loads(timeline_data_str)
+        audio_segs = data.get("audioSegments", [])
+    except Exception:
+        return empty_audio
+
+    if not audio_segs:
+        return empty_audio
+
+    out_waveform = torch.zeros((2, total_samples), dtype=torch.float32)
+
+    for seg in audio_segs:
+        buffer = None
+        if seg.get("audioFile"):
+            file_path = os.path.join(folder_paths.get_input_directory(), seg["audioFile"])
+            if os.path.exists(file_path):
+                with open(file_path, "rb") as f:
+                    buffer = _io.BytesIO(f.read())
+        
+        if not buffer and seg.get("audioB64"):
+            b64 = seg.get("audioB64")
+            if "," in b64:
+                b64 = b64.split(",", 1)[1]
+            try:
+                audio_bytes = base64.b64decode(b64)
+                buffer = _io.BytesIO(audio_bytes)
+            except:
+                pass
+                
+        if not buffer:
+            continue
+
+        try:
+            clip_frames = []
+            
+            # Use PyAV to decode directly from memory buffer
+            with av.open(buffer) as container:
+                stream = container.streams.audio[0]
+                
+                # Setup resampler to ensure output is 44.1kHz, Stereo, Float32 Planar
+                resampler = av.AudioResampler(
+                    format='fltp',
+                    layout='stereo',
+                    rate=target_sr,
+                )
+                
+                for frame in container.decode(stream):
+                    for resampled_frame in resampler.resample(frame):
+                        # to_ndarray() on fltp gives shape (channels, samples)
+                        arr = resampled_frame.to_ndarray()
+                        clip_frames.append(torch.from_numpy(arr))
+                
+                # Flush the resampler to get any remaining samples
+                for resampled_frame in resampler.resample(None):
+                    arr = resampled_frame.to_ndarray()
+                    clip_frames.append(torch.from_numpy(arr))
+
+            if not clip_frames:
+                continue
+
+            # Concatenate all frame blocks along the samples dimension (dim 1)
+            waveform = torch.cat(clip_frames, dim=1) # Shape: [2, total_clip_samples]
+
+            # Calculate interactive trim boundaries
+            trim_start_frames = float(seg.get("trimStart", 0))
+            length_frames = float(seg.get("length", 1))
+            start_frames = float(seg.get("start", 0))
+
+            start_sample_src = int(trim_start_frames / frame_rate * target_sr)
+            length_samples = int(length_frames / frame_rate * target_sr)
+            end_sample_src = start_sample_src + length_samples
+
+            if start_sample_src < 0: start_sample_src = 0
+            if end_sample_src > waveform.shape[1]:
+                end_sample_src = waveform.shape[1]
+
+            actual_length = end_sample_src - start_sample_src
+            if actual_length <= 0: continue
+
+            # Extract the correct segment of the audio
+            clip_waveform = waveform[:, start_sample_src:end_sample_src]
+
+            # Position onto the timeline
+            start_sample_dst = int(start_frames / frame_rate * target_sr)
+            
+            if start_sample_dst >= out_waveform.shape[1]:
+                continue
+                
+            end_sample_dst = start_sample_dst + actual_length
+
+            # Clip any trailing overflow so we don't index past the timeline bounds
+            if end_sample_dst > out_waveform.shape[1]:
+                actual_length = out_waveform.shape[1] - start_sample_dst
+                clip_waveform = clip_waveform[:, :actual_length]
+                end_sample_dst = start_sample_dst + actual_length
+                
+            if actual_length <= 0:
+                continue
+
+            # Additive composite (allows clips overlapping to sum together naturally)
+            out_waveform[:, start_sample_dst:end_sample_dst] += clip_waveform
+
+        except Exception as e:
+            log.warning("[PromptRelay] Audio process error for segment %s: %s", seg.get("fileName"), e)
+            continue
+
+    return {"waveform": out_waveform.unsqueeze(0), "sample_rate": target_sr}
+
+
+def _convert_to_latent_lengths(pixel_lengths, temporal_stride, latent_frames):
+    """Convert pixel-space segment lengths to integer latent-space lengths using the
+    largest-remainder method. Targets the full `latent_frames` when the pixel sum looks
+    like full coverage (within one stride of latent_frames * stride). Otherwise targets
+    round(total_pixel / temporal_stride) so partial-coverage timelines stay partial.
+    """
+    if not pixel_lengths:
+        return []
+    total_pixel = sum(pixel_lengths)
+    if total_pixel <= 0:
+        return [1] * len(pixel_lengths)
+
+    naive_total = max(1, round(total_pixel / temporal_stride))
+    target_total = min(latent_frames, naive_total)
+    # Within one frame of full → user clearly intended full coverage; pin to latent_frames.
+    if target_total >= latent_frames - 1:
+        target_total = latent_frames
+
+    exact = [p * target_total / total_pixel for p in pixel_lengths]
+    result = [int(e) for e in exact]
+    diff = target_total - sum(result)
+    if diff > 0:
+        order = sorted(range(len(exact)), key=lambda i: -(exact[i] - int(exact[i])))
+        for k in range(diff):
+            result[order[k % len(order)]] += 1
+
+    # Ensure every segment has ≥ 1 latent frame (steal from the largest if needed).
+    for i in range(len(result)):
+        if result[i] < 1:
+            max_idx = max(range(len(result)), key=lambda j: result[j])
+            if result[max_idx] > 1:
+                result[max_idx] -= 1
+                result[i] = 1
+
+    return result
+
+
+def _encode_relay(model, clip, latent, global_prompt, local_prompts, segment_lengths, epsilon):
+    for name, val in (("global_prompt", global_prompt),
+                      ("local_prompts", local_prompts),
+                      ("segment_lengths", segment_lengths)):
+        if val is None:
+            raise ValueError(
+                f"PromptRelay: '{name}' arrived as None. "
+                "Likely causes: a stale workflow JSON saved with null, the timeline "
+                "editor's web extension failing to load, or an upstream node returning None. "
+                "Set the field to an empty string or fix the upstream connection."
+            )
+
+    # Split prompts but do NOT filter out empty ones yet, so we can detect them
+    locals_list = [p.strip() for p in local_prompts.split("|")]
+    
+    # Check if any specific segment is empty
+    for p in locals_list:
+        if not p:
+            raise ValueError("There is a segment on the timeline missing a prompt!")
+
+    if not locals_list or (len(locals_list) == 1 and not locals_list[0]):
+        raise ValueError("At least one local prompt is required.")
+
+    arch, patch_size, temporal_stride = detect_model_type(model)
+
+    samples = latent["samples"]
+    latent_frames = samples.shape[2]
+    tokens_per_frame = (samples.shape[3] // patch_size[1]) * (samples.shape[4] // patch_size[2])
+
+    parsed_lengths = None
+    if segment_lengths.strip():
+        pixel_lengths = [int(float(x.strip())) for x in segment_lengths.split(",") if x.strip()]
+        parsed_lengths = _convert_to_latent_lengths(pixel_lengths, temporal_stride, latent_frames)
+
+    raw_tokenizer = get_raw_tokenizer(clip)
+    full_prompt, token_ranges = map_token_indices(raw_tokenizer, global_prompt, locals_list)
+
+    log.info("[PromptRelay] Global: tokens [0:%d] (%d tokens)", token_ranges[0][0], token_ranges[0][0])
+    for i, (s, e) in enumerate(token_ranges):
+        log.info("[PromptRelay] Segment %d: tokens [%d:%d] (%d tokens)", i, s, e, e - s)
+
+    conditioning = clip.encode_from_tokens_scheduled(clip.tokenize(full_prompt))
+
+    effective_lengths = distribute_segment_lengths(len(locals_list), latent_frames, parsed_lengths)
+
+    log.info(
+        "[PromptRelay] Latent: %d frames, %d tokens/frame, segments: %s",
+        latent_frames, tokens_per_frame, effective_lengths,
+    )
+
+    q_token_idx = build_segments(token_ranges, effective_lengths, epsilon, None)
+    mask_fn = create_mask_fn(q_token_idx, tokens_per_frame, latent_frames)
+
+    patched = model.clone()
+    apply_patches(patched, arch, mask_fn)
+
+    return patched, conditioning
+
+
+class LTXDirector(io.ComfyNode):
+    """WYSIWYG timeline variant — segments and lengths come from a visual editor in the node UI."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LTXDirector",
+            display_name="LTX Director",
+            category="conditioning/prompt_relay",
+            description=(
+                "Same as Prompt Relay Encode, but local prompts and segment lengths are edited "
+                "visually as draggable blocks on a timeline. The duration_frames input only sets the "
+                "timeline scale (pixel space) — actual frame count is still read from the latent."
+            ),
+            inputs=[
+                io.Model.Input("model"),
+                io.Clip.Input("clip"),
+                io.Vae.Input("audio_vae", optional=True, tooltip="Optional. Connect an Audio VAE to generate audio latents."),
+                io.Latent.Input("optional_latent", optional=True, tooltip="Optional. Connect a latent to override the auto-generated one."),
+                io.String.Input(
+                    "global_prompt", multiline=True, default="",
+                    tooltip="Conditions the entire video. Anchors persistent characters, objects, and scene context.",
+                ),
+                io.Int.Input(
+                    "duration_frames", default=120, min=1, max=10000, step=1,
+                    tooltip="Total timeline length in pixel-space frames. Used by the editor for visual scale only.",
+                ),
+                io.Float.Input(
+                    "duration_seconds", default=5, min=0.1, max=1000.0, step=0.01,
+                    tooltip="Total timeline duration in seconds (computed/synced from frames).",
+                ),
+                io.String.Input(
+                    "timeline_data", default="",
+                    tooltip="JSON state of the timeline editor (auto-managed; do not edit by hand).",
+                ),
+                io.Boolean.Input(
+                    "use_custom_audio", default=False, optional=True,
+                    tooltip="Toggle between using timeline audio (ON) and generating audio from scratch (OFF).",
+                ),
+                io.String.Input(
+                    "local_prompts", multiline=True, default="",
+                    tooltip="Auto-populated from the timeline editor.",
+                ),
+                io.String.Input(
+                    "segment_lengths", default="",
+                    tooltip="Auto-populated from the timeline editor (pixel-space frame counts).",
+                ),
+                io.Float.Input(
+                    "epsilon", default=0.001, min=0.0001, max=0.99, step=0.0001,
+                    tooltip="Penalty decay parameter. Values below ~0.1 all produce sharp boundaries (paper default 0.001). For softer transitions, try 0.5 or higher.",
+                ),
+                io.Float.Input(
+                    "frame_rate", default=24, min=1, max=240, step=1, optional=True,
+                    tooltip="Frames per second — only affects how time is displayed in the timeline editor when time_units is set to 'seconds'.",
+                ),
+                io.Combo.Input(
+                    "display_mode", options=["frames", "seconds"], default="seconds", optional=True,
+                    tooltip="Display the ruler, segment ranges, length input, and total in frames or seconds. Internal storage is always pixel-space frames.",
+                ),
+                io.String.Input(
+                    "guide_strength", default="",
+                    tooltip="Auto-populated from the timeline editor (comma-separated guide strengths for image segments).",
+                ),
+                io.Int.Input(
+                    "custom_width", default=0, min=0, max=8192, step=1, optional=True,
+                    tooltip="Target output width for all image segments. Set to 0 to use the original image width.",
+                ),
+                io.Int.Input(
+                    "custom_height", default=0, min=0, max=8192, step=1, optional=True,
+                    tooltip="Target output height for all image segments. Set to 0 to use the original image height.",
+                ),
+                io.Combo.Input(
+                    "resize_method",
+                    options=["maintain aspect ratio", "stretch to fit", "pad", "crop"],
+                    default="maintain aspect ratio",
+                    optional=True,
+                    tooltip="How to resize image segments to fit the target dimensions.",
+                ),
+                io.Int.Input(
+                    "divisible_by", default=32, min=1, max=256, step=1, optional=True,
+                    tooltip="Snap the final output image dimensions to be divisible by this number (e.g. 32 for LTX).",
+                ),
+                io.Int.Input(
+                    "img_compression", default=18, min=0, max=100, step=1, optional=True,
+                    tooltip="H.264 CRF compression to apply to each guide image. 0 = no compression, higher = more artefacts.",
+                ),
+            ],
+            outputs=[
+                io.Model.Output(display_name="model"),
+                io.Conditioning.Output(display_name="positive"),
+                io.Latent.Output(display_name="video_latent", tooltip="Auto-generated LTXV empty latent (only populated when no latent is connected)."),
+                io.Latent.Output(display_name="audio_latent", tooltip="Auto-generated audio latent (uses custom audio if enabled)."),
+                GuideData.Output(display_name="guide_data"),
+                io.Float.Output(display_name="frame_rate", tooltip="The frame rate used for the timeline."),
+                io.Audio.Output(display_name="combined_audio", tooltip="Combined timeline audio layout."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model, clip, global_prompt, duration_frames, duration_seconds,
+                timeline_data, local_prompts, segment_lengths, guide_strength="", epsilon=1e-3,
+                frame_rate=24, display_mode="seconds",
+                custom_width=768, custom_height=512, resize_method="maintain aspect ratio",
+                divisible_by=32, img_compression=0, audio_vae=None, optional_latent=None,
+                use_custom_audio=False) -> io.NodeOutput:
+
+        # --- Build guide_data from image segments FIRST (to derive output dimensions) ---
+        guide_data = {"images": [], "insert_frames": [], "strengths": [], "frame_rate": frame_rate}
+        derived_w, derived_h = custom_width, custom_height
+        try:
+            tdata = json.loads(timeline_data) if timeline_data else {}
+            img_segs = [
+                s for s in tdata.get("segments", [])
+                if s.get("type", "image") == "image"
+                and (s.get("imageFile") or s.get("imageB64"))
+                and int(s.get("start", 0)) < duration_frames  # exclude segments fully outside duration
+            ]
+            img_segs.sort(key=lambda s: s["start"])
+
+            strengths = []
+            if guide_strength.strip():
+                strengths = [float(x.strip()) for x in guide_strength.split(",") if x.strip()]
+
+            for idx, seg in enumerate(img_segs):
+                tensor = _load_image_tensor(seg)
+
+                # Apply resize
+                src_h, src_w = tensor.shape[1], tensor.shape[2]
+
+                def snap(val, div):
+                    return max(div, (val // div) * div)
+
+                if custom_width > 0 and custom_height > 0:
+                    # Both dimensions set — apply selected resize_method (pad, crop, stretch, maintain AR)
+                    tensor = _resize_image(tensor, custom_width, custom_height, resize_method, divisible_by)
+                elif custom_width > 0:
+                    # Width only — scale height from AR, snap both, then resize to exact dimensions
+                    tgt_w = snap(custom_width, divisible_by)
+                    tgt_h = snap(int(src_h * tgt_w / src_w), divisible_by)
+                    tensor = _resize_image(tensor, tgt_w, tgt_h, "stretch to fit", divisible_by)
+                elif custom_height > 0:
+                    # Height only — scale width from AR, snap both, then resize to exact dimensions
+                    tgt_h = snap(custom_height, divisible_by)
+                    tgt_w = snap(int(src_w * tgt_h / src_h), divisible_by)
+                    tensor = _resize_image(tensor, tgt_w, tgt_h, "stretch to fit", divisible_by)
+                else:
+                    # Both zero — keep original dimensions, just snap to divisible_by
+                    tensor = _resize_image(tensor, src_w, src_h, "maintain aspect ratio", divisible_by)
+
+
+                # Apply compression
+                if img_compression > 0:
+                    tensor = _compress_image(tensor, img_compression)
+
+                # Record dimensions of the first processed image for latent generation
+                if idx == 0:
+                    derived_h = tensor.shape[1]
+                    derived_w = tensor.shape[2]
+
+                strength = strengths[idx] if idx < len(strengths) else 1.0
+                guide_data["images"].append(tensor)
+                guide_data["insert_frames"].append(int(seg["start"]))
+                guide_data["strengths"].append(float(strength))
+            
+            # If no images were loaded from the timeline, create a dummy image at strength 0
+            # to prevent artifacts in text-to-video mode.
+            if not guide_data["images"]:
+                w = derived_w if derived_w > 0 else 768
+                h = derived_h if derived_h > 0 else 512
+                w = (w // 32) * 32
+                h = (h // 32) * 32
+                
+                dummy_image = torch.zeros((1, h, w, 3), dtype=torch.float32)
+                guide_data["images"].append(dummy_image)
+                guide_data["insert_frames"].append(0)
+                guide_data["strengths"].append(0.0)
+                
+                derived_w = w
+                derived_h = h
+        except Exception as e:
+            log.warning("[PromptRelay] Could not build guide_data: %s", e)
+
+        # --- Auto-generate LTXV latent if none was provided ---
+        ltxv_length = duration_frames + 1
+        if optional_latent is None:
+            latent_w = max(32, (derived_w // 32) * 32)
+            latent_h = max(32, (derived_h // 32) * 32)
+            # LTXV temporal: ((length - 1) // 8) + 1 latent frames; invert to get pixel frames -> length
+            latent_t = ((ltxv_length - 1) // 8) + 1
+            samples = torch.zeros(
+                [1, 128, latent_t, latent_h // 32, latent_w // 32],
+                device=comfy.model_management.intermediate_device(),
+            )
+            latent = {"samples": samples}
+            log.info(
+                "[PromptRelay] Auto-generated LTXV latent: %dx%d, %d pixel frames (%d latent frames)",
+                latent_w, latent_h, ltxv_length, latent_t,
+            )
+        else:
+            latent = optional_latent
+
+        patched, conditioning = _encode_relay(
+            model, clip, latent, global_prompt, local_prompts, segment_lengths, epsilon,
+        )
+
+        # --- Build Audio Output ---
+        audio_out = _build_combined_audio(timeline_data, ltxv_length, float(frame_rate))
+
+        # --- Audio Latent Generation ---
+        audio_latent = {}
+        
+        if audio_vae is not None:
+            # Helper to generate empty latent
+            def get_empty_latent():
+                z_channels = audio_vae.latent_channels
+                audio_freq = audio_vae.first_stage_model.latent_frequency_bins
+                num_audio_latents = audio_vae.first_stage_model.num_of_latents_from_frames(ltxv_length, float(frame_rate))
+                audio_latents = torch.zeros(
+                    (1, z_channels, num_audio_latents, audio_freq),
+                    device=comfy.model_management.intermediate_device(),
+                )
+                return {"samples": audio_latents, "type": "audio"}
+
+            if use_custom_audio:
+                try:
+                    if audio_out is not None:
+                        # 1. Encode audio waveform into latent space
+                        # VAE expects shape (batch, samples, channels), so we move dim 1 (channels) to the end
+                        latent_samples = audio_vae.encode(audio_out["waveform"].movedim(1, -1))
+                        
+                        if latent_samples.numel() == 0:
+                            raise ValueError("Encoded audio latent is empty (0 elements).")
+                        
+                        # 2. Create solid mask with value 0.0 (0 means keep/use conditioning, 1 means generate noise)
+                        mask = torch.full(
+                            (1, latent_samples.shape[-2], latent_samples.shape[-1]), 
+                            0.0, 
+                            dtype=torch.float32, 
+                            device=comfy.model_management.intermediate_device()
+                        )
+                        
+                        # 3. Set Latent Noise Mask
+                        audio_latent = {
+                            "samples": latent_samples,
+                            "type": "audio",
+                            "noise_mask": mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1]))
+                        }
+                        log.info("[PromptRelay] Generated custom audio latent with noise mask (value=0.0).")
+                    else:
+                        raise ValueError("No audio waveform to encode.")
+                except Exception as e:
+                    log.error("[PromptRelay] Failed to generate custom audio latent: %s", e)
+                    raise e
+            else:
+                # Generate empty latent
+                try:
+                    audio_latent = get_empty_latent()
+                    log.info("[PromptRelay] Auto-generated empty audio latent.")
+                except Exception as e:
+                    log.warning("[PromptRelay] Could not generate empty audio latent: %s", e)
+
+        return io.NodeOutput(patched, conditioning, latent, audio_latent, guide_data, float(frame_rate), audio_out)
+
+
+NODE_CLASS_MAPPINGS = {
+    "LTXDirector": LTXDirector,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "PromptRelayEncodeTimeline": "Prompt Relay Encode (Timeline)",
+}

ltx_director_guide.py

@@ -0,0 +1,90 @@
+from comfy_extras.nodes_lt import LTXVAddGuide
+import torch
+import comfy.utils
+from comfy_api.latest import io
+from .ltx_director import GuideData
+
+
+class LTXDirectorGuide(LTXVAddGuide):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="LTXDirectorGuide",
+            display_name="LTX Director Guide",
+            category="LTXVCustom",
+            description=(
+                "Applies guide images from a Prompt Relay Timeline node at the frame positions "
+                "and strengths defined on the timeline. Connect guide_data from the timeline node."
+            ),
+            inputs=[
+                io.Conditioning.Input("positive", tooltip="Positive conditioning to add guide keyframe info to."),
+                io.Conditioning.Input("negative", tooltip="Negative conditioning to add guide keyframe info to."),
+                io.Vae.Input("vae", tooltip="Video VAE used to encode the guide images."),
+                io.Latent.Input("latent", tooltip="Video latent — guides are inserted into this latent."),
+                GuideData.Input("guide_data", tooltip="Guide data produced by Prompt Relay Encode (Timeline)."),
+                io.Float.Input("scale_by", default=1.0, min=0.01, max=8.0, step=0.01, tooltip="Scale the latent by this factor."),
+                io.Combo.Input("upscale_method", options=["nearest-exact", "bilinear", "area", "bicubic", "bislerp"], default="bicubic", tooltip="Method used to upscale/downscale the latent."),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent", tooltip="Video latent with guide frames applied."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, vae, latent, guide_data, scale_by=1.0, upscale_method="bicubic") -> io.NodeOutput:
+        scale_factors = vae.downscale_index_formula
+
+        # Clone latents to avoid mutating upstream nodes
+        latent_image = latent["samples"].clone()
+
+        if "noise_mask" in latent:
+            noise_mask = latent["noise_mask"].clone()
+        else:
+            batch, _, latent_frames, latent_height, latent_width = latent_image.shape
+            noise_mask = torch.ones(
+                (batch, 1, latent_frames, 1, 1),
+                dtype=torch.float32,
+                device=latent_image.device,
+            )
+
+        # Apply scale factor if not 1.0
+        if scale_by != 1.0:
+            B, C, F, H, W = latent_image.shape
+            width = round(W * scale_by)
+            height = round(H * scale_by)
+            
+            # Reshape to 4D for common_upscale
+            latent_4d = latent_image.permute(0, 2, 1, 3, 4).reshape(B * F, C, H, W)
+            latent_resized_4d = comfy.utils.common_upscale(latent_4d, width, height, upscale_method, "disabled")
+            latent_image = latent_resized_4d.reshape(B, F, C, height, width).permute(0, 2, 1, 3, 4)
+
+            # Also resize noise mask if it's not a broadcasted mask
+            if noise_mask.shape[-1] > 1 or noise_mask.shape[-2] > 1:
+                mask_4d = noise_mask.permute(0, 2, 1, 3, 4).reshape(B * F, 1, H, W)
+                mask_resized_4d = comfy.utils.common_upscale(mask_4d, width, height, upscale_method, "disabled")
+                noise_mask = mask_resized_4d.reshape(B, F, 1, height, width).permute(0, 2, 1, 3, 4)
+
+        _, _, latent_length, latent_height, latent_width = latent_image.shape
+
+        images = guide_data.get("images", [])
+        insert_frames = guide_data.get("insert_frames", [])
+        strengths = guide_data.get("strengths", [])
+
+        for idx, img_tensor in enumerate(images):
+            f_idx = insert_frames[idx] if idx < len(insert_frames) else 0
+            strength = strengths[idx] if idx < len(strengths) else 1.0
+
+            image_1, t = cls.encode(vae, latent_width, latent_height, img_tensor, scale_factors)
+            frame_idx, latent_idx = cls.get_latent_index(positive, latent_length, len(image_1), f_idx, scale_factors)
+
+            assert latent_idx + t.shape[2] <= latent_length, (
+                f"Guide image {idx + 1}: conditioning frames exceed the length of the latent sequence."
+            )
+
+            positive, negative, latent_image, noise_mask = cls.append_keyframe(
+                positive, negative, frame_idx, latent_image, noise_mask, t, strength, scale_factors,
+            )
+
+        return io.NodeOutput(positive, negative, {"samples": latent_image, "noise_mask": noise_mask})

patches.py

@@ -0,0 +1,167 @@
+import types
+import torch
+import comfy.ldm.modules.attention
+
+
+def _masked_attention(q, k, v, heads, mask, transformer_options={}, **kwargs):
+    # Bypass wrap_attn (sage/etc may ignore masks) by calling attention_pytorch directly.
+    return comfy.ldm.modules.attention.attention_pytorch(
+        q, k, v, heads, mask=mask,
+        _inside_attn_wrapper=True,
+        transformer_options=transformer_options,
+        **kwargs,
+    )
+
+
+def _wan_t2v_forward(self, mask_fn, x, context, transformer_options={}, **kwargs):
+    q = self.norm_q(self.q(x))
+    k = self.norm_k(self.k(context))
+    v = self.v(context)
+
+    mask = mask_fn(q, k, transformer_options)
+    if mask is not None:
+        x = _masked_attention(q, k, v, heads=self.num_heads, mask=mask,
+                              transformer_options=transformer_options)
+    else:
+        x = comfy.ldm.modules.attention.optimized_attention(
+            q, k, v, heads=self.num_heads, transformer_options=transformer_options,
+        )
+    return self.o(x)
+
+
+def _wan_i2v_forward(self, mask_fn, x, context, context_img_len, transformer_options={}, **kwargs):
+    context_img = context[:, :context_img_len]
+    context_text = context[:, context_img_len:]
+
+    q = self.norm_q(self.q(x))
+
+    k_img = self.norm_k_img(self.k_img(context_img))
+    v_img = self.v_img(context_img)
+    img_x = comfy.ldm.modules.attention.optimized_attention(
+        q, k_img, v_img, heads=self.num_heads, transformer_options=transformer_options,
+    )
+
+    k = self.norm_k(self.k(context_text))
+    v = self.v(context_text)
+
+    mask = mask_fn(q, k, transformer_options)
+    if mask is not None:
+        x = _masked_attention(q, k, v, heads=self.num_heads, mask=mask,
+                              transformer_options=transformer_options)
+    else:
+        x = comfy.ldm.modules.attention.optimized_attention(
+            q, k, v, heads=self.num_heads, transformer_options=transformer_options,
+        )
+
+    return self.o(x + img_x)
+
+
+def _ltx_forward(self, mask_fn, x, context=None, mask=None, pe=None, k_pe=None, transformer_options={}):
+    from comfy.ldm.lightricks.model import apply_rotary_emb
+
+    is_self_attn = context is None
+    context = x if is_self_attn else context
+
+    q = self.q_norm(self.to_q(x))
+    k = self.k_norm(self.to_k(context))
+    v = self.to_v(context)
+
+    if pe is not None:
+        q = apply_rotary_emb(q, pe)
+        k = apply_rotary_emb(k, pe if k_pe is None else k_pe)
+
+    if not is_self_attn:
+        temporal_mask = mask_fn(q, k, transformer_options)
+        if temporal_mask is not None:
+            mask = temporal_mask if mask is None else mask + temporal_mask
+
+    if mask is None:
+        out = comfy.ldm.modules.attention.optimized_attention(
+            q, k, v, self.heads, attn_precision=self.attn_precision,
+            transformer_options=transformer_options,
+        )
+    else:
+        out = _masked_attention(q, k, v, self.heads, mask=mask,
+                                attn_precision=self.attn_precision,
+                                transformer_options=transformer_options)
+
+    if self.to_gate_logits is not None:
+        gate_logits = self.to_gate_logits(x)
+        b, t, _ = out.shape
+        out = out.view(b, t, self.heads, self.dim_head)
+        out = out * (2.0 * torch.sigmoid(gate_logits)).unsqueeze(-1)
+        out = out.view(b, t, self.heads * self.dim_head)
+
+    return self.to_out(out)
+
+
+class _CrossAttnPatch:
+    """Descriptor that binds (impl, mask_fn) as a method onto a cross-attn module."""
+
+    def __init__(self, impl, mask_fn):
+        self.impl = impl
+        self.mask_fn = mask_fn
+
+    def __get__(self, obj, objtype=None):
+        impl, mask_fn = self.impl, self.mask_fn
+
+        def wrapped(self_module, *args, **kwargs):
+            return impl(self_module, mask_fn, *args, **kwargs)
+
+        return types.MethodType(wrapped, obj)
+
+
+def detect_model_type(model):
+    """Return (arch, patch_size, temporal_stride) for latent geometry.
+
+    temporal_stride is the VAE's pixel→latent temporal compression factor,
+    used to convert user-facing pixel frame counts to latent frames.
+    """
+    diff_model = model.model.diffusion_model
+
+    if hasattr(diff_model, "patch_size") and not hasattr(diff_model, "patchifier"):
+        return "wan", tuple(diff_model.patch_size), 4
+
+    if hasattr(diff_model, "patchifier"):
+        return "ltx", (1, 1, 1), int(diff_model.vae_scale_factors[0])
+
+    raise ValueError(
+        f"Unsupported model type: {type(diff_model).__name__}. "
+        f"Currently supports Wan and LTX models."
+    )
+
+
+def _check_unpatched(model_clone, key):
+    if key in getattr(model_clone, "object_patches", {}):
+        raise RuntimeError(
+            f"PromptRelay: cross-attention forward at '{key}' is already patched by "
+            "another node (e.g. KJNodes NAG). Stacking is not supported — remove the "
+            "conflicting node."
+        )
+
+
+def apply_patches(model_clone, arch, mask_fn):
+    diffusion_model = model_clone.get_model_object("diffusion_model")
+
+    if arch == "wan":
+        from comfy.ldm.wan.model import WanI2VCrossAttention
+        for idx, block in enumerate(diffusion_model.blocks):
+            key = f"diffusion_model.blocks.{idx}.cross_attn.forward"
+            _check_unpatched(model_clone, key)
+            cross_attn = block.cross_attn
+            impl = _wan_i2v_forward if isinstance(cross_attn, WanI2VCrossAttention) else _wan_t2v_forward
+            model_clone.add_object_patch(key, _CrossAttnPatch(impl, mask_fn).__get__(cross_attn, cross_attn.__class__))
+        return
+
+    if arch == "ltx":
+        for idx, block in enumerate(diffusion_model.transformer_blocks):
+            for attr in ("attn2", "audio_attn2"):
+                module = getattr(block, attr, None)
+                if module is None:
+                    continue
+                key = f"diffusion_model.transformer_blocks.{idx}.{attr}.forward"
+                _check_unpatched(model_clone, key)
+                model_clone.add_object_patch(key, _CrossAttnPatch(_ltx_forward, mask_fn).__get__(module, module.__class__))
+        return
+
+    raise ValueError(f"Unknown model arch: {arch}")

prompt_relay.py

@@ -0,0 +1,201 @@
+import logging
+import math
+import torch
+
+log = logging.getLogger(__name__)
+
+
+def build_temporal_cost(q_token_idx, Lq, Lk, device, dtype, tokens_per_frame):
+    """Gaussian penalty matrix [Lq, Lk] for video cross-attention (integer frame indexing)."""
+    offset = torch.zeros(Lq, Lk, device=device, dtype=dtype)
+    query_frames = torch.arange(Lq, device=device, dtype=torch.long) // tokens_per_frame
+
+    for seg in q_token_idx:
+        local = seg["local_token_idx"].to(device=device)
+        d = (query_frames.float()[:, None] - seg["midpoint"]).abs()
+        strength = seg.get("strength", 1.0)
+        cost = strength * (torch.relu(d - seg["window"]) ** 2) / (2 * seg["sigma"] ** 2)
+        offset[:, local] = cost.to(offset.dtype)
+
+    return offset
+
+
+def build_temporal_cost_scaled(q_token_idx, Lq, Lk, device, dtype, latent_frames):
+    """Penalty matrix for queries that don't map to integer frames (e.g. LTXAV audio tokens)."""
+    offset = torch.zeros(Lq, Lk, device=device, dtype=dtype)
+    query_frames = torch.arange(Lq, device=device, dtype=torch.float32) * latent_frames / Lq
+
+    for seg in q_token_idx:
+        local = seg["local_token_idx"].to(device=device)
+        d = (query_frames[:, None] - seg["midpoint"]).abs()
+        sigma_a = seg.get("sigma_audio", seg["sigma"])
+        window_a = seg.get("window_audio", seg["window"])
+        strength_a = seg.get("strength_audio", 1.0)
+        cost = strength_a * (torch.relu(d - window_a) ** 2) / (2 * sigma_a ** 2)
+        offset[:, local] = cost.to(offset.dtype)
+
+    return offset
+
+
+def create_mask_fn(q_token_idx, fallback_tokens_per_frame, latent_frames):
+    """Closure: mask_fn(q, k, transformer_options) -> additive mask or None."""
+    cache = {}
+    max_token_idx = max(int(seg["local_token_idx"].max().item()) for seg in q_token_idx) + 1
+
+    def mask_fn(q, k, transformer_options):
+        Lq, Lk = q.shape[1], k.shape[1]
+
+        if Lq == Lk:
+            return None
+
+        # Only apply on conditional pass — not unconditional (negative prompt)
+        cond_or_uncond = transformer_options.get("cond_or_uncond", [])
+        if 1 in cond_or_uncond and 0 not in cond_or_uncond:
+            return None
+
+        grid_sizes = transformer_options.get("grid_sizes", None)
+        video_tpf = int(grid_sizes[1]) * int(grid_sizes[2]) if grid_sizes is not None else fallback_tokens_per_frame
+        video_lq = latent_frames * video_tpf
+
+        # Skip cross-modal attention — text keys are padded to a fixed length ≥ max_token_idx and != video_lq
+        if Lk == video_lq or Lk < max_token_idx:
+            return None
+
+        mode = "video" if Lq == video_lq else "scaled"
+
+        key = (Lq, Lk, mode, q.device)
+        if key not in cache:
+            if mode == "video":
+                cost = build_temporal_cost(q_token_idx, Lq, Lk, q.device, q.dtype, video_tpf)
+            else:
+                cost = build_temporal_cost_scaled(q_token_idx, Lq, Lk, q.device, q.dtype, latent_frames)
+            log.info(
+                "[PromptRelay] Built penalty matrix (%s): Lq=%d, Lk=%d, nonzero=%d/%d",
+                mode, Lq, Lk, (cost > 0).sum().item(), cost.numel(),
+            )
+            cache[key] = -cost
+
+        return cache[key].to(q.dtype)
+
+    return mask_fn
+
+
+def build_segments(token_ranges, segment_lengths, epsilon=1e-3, relay_options=None):
+    """Per-segment metadata for the temporal penalty.
+
+    relay_options (optional dict) overrides per-stream knobs:
+        video_strength, video_window_scale,
+        audio_epsilon, audio_strength, audio_window_scale
+    Audio knobs only affect architectures whose cross-attention takes the scaled
+    (non-integer-frame) path — currently LTX audio_attn2.
+    """
+    # Paper uses constant sigma = 1/ln(1/epsilon) regardless of segment length
+    sigma = 1.0 / math.log(1.0 / epsilon) if 0 < epsilon < 1 else 0.1448
+
+    opts = relay_options or {}
+    v_strength = opts.get("video_strength", 1.0)
+    v_window_scale = opts.get("video_window_scale", 1.0)
+    a_epsilon = opts.get("audio_epsilon")
+    a_strength = opts.get("audio_strength", 1.0)
+    a_window_scale = opts.get("audio_window_scale", 1.0)
+
+    if a_epsilon is not None and 0 < a_epsilon < 1:
+        sigma_audio = 1.0 / math.log(1.0 / a_epsilon)
+    else:
+        sigma_audio = sigma
+
+    if relay_options:
+        log.info(
+            "[PromptRelay] Advanced options active — video: strength=%.3f window_scale=%.3f | "
+            "audio: epsilon=%s strength=%.3f window_scale=%.3f",
+            v_strength, v_window_scale,
+            f"{a_epsilon:.4f}" if a_epsilon is not None else "inherit",
+            a_strength, a_window_scale,
+        )
+
+    q_token_idx = []
+    frame_cursor = 0
+
+    for (tok_start, tok_end), L in zip(token_ranges, segment_lengths):
+        if L <= 0:
+            frame_cursor += L
+            continue
+        midpoint = (2 * frame_cursor + L) // 2
+        base_window = max(L // 2 - 2, 0)
+        q_token_idx.append({
+            "local_token_idx": torch.arange(tok_start, tok_end),
+            "midpoint": midpoint,
+            "window": max(base_window * v_window_scale, 0.0),
+            "sigma": sigma,
+            "strength": v_strength,
+            "window_audio": max(base_window * a_window_scale, 0.0),
+            "sigma_audio": sigma_audio,
+            "strength_audio": a_strength,
+        })
+        frame_cursor += L
+
+    return q_token_idx
+
+
+def get_raw_tokenizer(clip):
+    """Extract the raw SPiece/HF tokenizer from a ComfyUI CLIP object."""
+    tokenizer_wrapper = clip.tokenizer
+    for attr_name in dir(tokenizer_wrapper):
+        if attr_name.startswith("_"):
+            continue
+        inner = getattr(tokenizer_wrapper, attr_name, None)
+        if inner is not None and hasattr(inner, "tokenizer"):
+            return inner.tokenizer
+
+    raise RuntimeError(
+        f"Could not find raw tokenizer on CLIP object. "
+        f"Known attributes: {[a for a in dir(tokenizer_wrapper) if not a.startswith('_')]}"
+    )
+
+
+def map_token_indices(raw_tokenizer, global_prompt, local_prompts):
+    """Tokenize global + space-prefixed locals; return (full_prompt, per-local token ranges).
+
+    Uses incremental tokenization to avoid SentencePiece context-dependency issues.
+    """
+    prefixed_locals = [" " + lp for lp in local_prompts]
+    full_prompt = global_prompt + "".join(prefixed_locals)
+    has_eos = getattr(raw_tokenizer, "add_eos", False)
+    eos_adj = 1 if has_eos else 0
+
+    prev_len = len(raw_tokenizer(global_prompt)["input_ids"]) - eos_adj
+    token_ranges = []
+    built = global_prompt
+
+    for plp in prefixed_locals:
+        built += plp
+        cur_len = len(raw_tokenizer(built)["input_ids"]) - eos_adj
+        if cur_len <= prev_len:
+            raise ValueError(f"Local prompt produced no tokens: '{plp.strip()}'")
+        token_ranges.append((prev_len, cur_len))
+        prev_len = cur_len
+
+    return full_prompt, token_ranges
+
+
+def distribute_segment_lengths(num_segments, latent_frames, specified_lengths=None):
+    """Validate or auto-distribute segment frame counts, capped to fit within latent_frames."""
+    if specified_lengths:
+        if len(specified_lengths) != num_segments:
+            raise ValueError(
+                f"Number of segment_lengths ({len(specified_lengths)}) "
+                f"must match number of local prompts ({num_segments})"
+            )
+        lengths = specified_lengths
+    else:
+        # ceil division — matches reference implementation
+        step = -(-latent_frames // num_segments)
+        lengths = [step] * num_segments
+
+    effective = []
+    cursor = 0
+    for L in lengths:
+        end = min(cursor + L, latent_frames)
+        effective.append(max(end - cursor, 0))
+        cursor = end
+    return effective

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "WhatDreamsCost-ComfyUI"
 description = "A variety of custom ComfyUI nodes and workflows for creatives."
-version = "1.2.9"
+version = "1.3.0"
 license = {file = "LICENSE"} 
 # classifiers = [
 #     # For OS-independent nodes (works on all operating systems)