Connect reference image(s) to source_clip_2 and set keyframe_positions
to the anchored frame indices — the prep node replaces those frames
before passing to the mask generator, no manual frame swap needed.
Supports 1 frame broadcast to all positions, or one frame per position.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Provides real upscaled pixels as VACE reference (not grey) with white
mask, so VACE refines existing content rather than generating blind.
Reuses keyframe_positions to anchor specific frames exactly (black mask)
— wire the same value from the original Keyframe generation, no rewiring
needed.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ComfyUI saves latents under key 'latent_tensor', remap to 'samples'
so downstream nodes get the expected LATENT dict format.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
LATENT dicts can contain nested tensors or other non-JSON-serializable
objects. Filter them out instead of crashing on json.dumps.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Try pipeline["sd"] first (merged base+VACE weights), then
diffusion_model state dict, then reload from checkpoint file.
Applies LoRA patches from model patcher on top.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Model weights stay on meta device (no data) until ComfyUI's model
management loads them. Use load_models_gpu with force_full_load
to materialize weights before extracting state dict.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- pyproject.toml with package metadata (v1.0.0)
- MIT license
- GitHub Actions workflow to auto-publish on pyproject.toml changes
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Inline SVGs are stripped by GitHub's markdown sanitizer. Moved all 11
diagrams to docs/diagrams/*.svg and referenced them with <img> tags.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace torch.cat with pre-allocated tensor + in-place copy.
Clone only small context slices for blending instead of holding
full source/vace tensors during the blend loop.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Middle Extend: split_index=0 now auto-splits at B//2 instead of empty first half
- Middle Extend/Replace: raise ValueError when split_index >= frame count
- Merge blend loops: add bounds checks to prevent index-out-of-bounds
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add source_clip_2 and original_clip_2 to README input tables, update
Join Extend mode reference with two-clip diagram, update MergeBack
description and behavior section, fix mode output type label.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
SourcePrep gets optional source_clip_2 input — when connected in Join
Extend mode, joins two separate clips instead of splitting one in half.
MergeBack gets optional original_clip_2 to reconstruct the full result
from both originals. Single-clip behavior is unchanged.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove segment_1-4 from both MaskGenerator and SourcePrep outputs
- Define shared VACE_MODES list so SourcePrep's mode output type matches
MaskGenerator's combo input type (fixes inability to connect mode wire)
- Remove dead helpers: _placeholder, _ensure_nonempty, ph(), safe()
- Clean up dead variables (mid_seg, part_1, part_4, image_b, segment_1)
- Remove segment tables from README mode reference sections
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace 3 separate forceInput wires (mode, trim_start, trim_end) plus
mask and blend_frames with a single VACE_PIPE dict carrying mode, trim
bounds, and context frame counts. SourcePrep outputs reduced from 12 to
7 (segments removed), MergeBack inputs reduced from 9 to 5. Blending
now auto-derives from context counts instead of manual blend_frames.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds a new node that reconstructs full-length video by splicing VACE
sampler output back into the original clip at the trim positions. Supports
optical flow, alpha, and hard-cut blending at context/generated seams.
Also adds trim_start/trim_end INT outputs to VACESourcePrep.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New node that handles frame selection/trimming before the mask generator,
with input_left/input_right controls for per-mode windowing. Adds
B > target_frames validation to the mask generator with a helpful error
message. Includes JS extension for smart widget visibility per mode.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Consolidates SaveLatentAbsolute and LoadLatentAbsolute into this project
as latent_node.py. Saves and loads LATENT data to absolute file paths in
safetensors format, preserving device info and non-tensor metadata.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Consolidates the WanVideoSaveMergedModel node into this project as
save_node.py. Saves WanVideo diffusion models (with merged LoRAs) as
safetensors files with metadata for traceability.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New generation mode where source_clip keyframes are placed at specific
positions within a target_frames-length output, with everything between
them marked for generation. Supports auto-spread (even distribution) and
manual placement via optional keyframe_positions string input.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Fix output tooltips that incorrectly assumed target_frames length for all
modes, add per-mode detail to segment outputs, complete the unused-mode
lists on split_index and edge_frames inputs, and add a DESCRIPTION class
attribute for the ComfyUI "?" help button.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New 9th mode that works at the pixel level rather than the frame level.
Accepts an optional MASK input (B,H,W) to mark spatial regions for
regeneration, with single-frame broadcast, spatial dimension validation,
and contiguous output tensors.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three new modes for the VACE Mask Generator node, bringing the total to 8.
Bidirectional generates before and after the clip, Frame Interpolation
inserts frames between each source pair, and Replace/Inpaint regenerates
a region in-place. All reuse existing inputs with mode-specific semantics.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Single ComfyUI node implementing 5 VACE mask generation modes
(End/Pre/Middle/Edge/Join Extend) with zero external dependencies.
Replaces ~149 nodes from the original logic.json workflow.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Ethanfel