English | 中文

Multi-reference image injection and seamless video connection for ComfyUI's WanAnimate pipeline.

ComfyUI-WanAnimatePlus adds four major feature groups to the WanVideo workflow:

• prefix_frames: allows passing 1–5 additional reference images for multi-reference guided generation
• transition_video: allows passing the last 21 frames of the previous video segment for seamless video connection

When used together, canvas layout and frame offsets are automatically coordinated without conflicts.

Supports Bernini models. Allows passing source video, reference video, or reference images as generation conditions. Supports v2v, rv2v, r2v, and t2v tasks.

Use cases:

• Multi-shot video sequence generation
• Video continuation / extension
• Motion transfer with multi-reference control
• Video editing with source video + reference images
• Reference-to-video generation

Adds a wrapper-native WanAnimatePlus SCAIL_2 Embeds node for SCAIL-2 models. It prepares reference image, driving pose, colored pose mask, reference mask, optional prefix/transition hard-freeze latents, and prefix-aligned colored masks for the WanAnimatePlus sampler.

Allows 1–5 additional reference images. Internally expands the canvas pixel space and encodes reference images across the front frames, with automatic frame offset coordination for control signals (pose / face).

• Supports 1–5 reference images (truncated if exceeding 5)
• Auto-resizes reference images to target resolution
• Automatically aligns frame offsets for pose / face / bg / mask signals

Allows passing the last 21 frames of the previous video segment. Writes these pixel frames directly into the front of the generation canvas, with sampled+reversed padding for control signal offsets.

• Automatically coordinates with prefix when both are used

Generates condition latents from source video, reference video, and/or reference images via VAE encoding. Supports v2v, rv2v, r2v, and t2v — task is auto-detected from connected inputs.

• Reference images kept at native aspect ratio
• Compatible with context windows

Provides SCAIL-2 ref / pose / mask conditioning through WanAnimatePlus SCAIL_2 Embeds.

• Encodes ref_image, pose_images, pose_image_mask, prefix_frames, prefix_mask, bg_image, and reference_image_mask
• Aligns SCAIL-2 inputs to 32-pixel multiples before VAE encoding
• Supports animation and replacement modes
• Supports single-frame prefix reference encoding and optional transition_video hard-freeze conditioning
• By default, prefix_frames are encoded as full-resolution reference latents and do not expand the output canvas; disable single_frame_prefix_encoding to use the legacy 37 front pixel-frame prefix layout
• In single-frame prefix mode, prefix_mask follows the same reference-mask path as reference_image_mask
• Supports context-window sampling; non-first windows can see prepended prefix/transition context without fusing those prepended predictions

Place this repository into ComfyUI's custom_nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/wuwukaka/ComfyUI-WanAnimatePlus.git

Restart ComfyUI after installation.

Important: To use prefix_frames, transition_video, Bernini, or SCAIL_2 Embeds, you must replace the full workflow chain with WanAnimatePlus nodes. Mixing WanAnimatePlus nodes with original WanVideoWrapper nodes in the same workflow will result in degraded output.

1. Start ComfyUI and confirm the WanAnimatePlus nodes appear under the WanAnimatePlus category
2. Replace the entire workflow chain with WanAnimatePlus counterparts: ModelLoader, VAELoader, ContextOptions, AnimateEmbeds, Sampler, Decode, and supporting nodes
3. Do not mix original WanVideoWrapper nodes in the same workflow
4. Connect prefix_frames and/or transition_video inputs as needed
5. Example workflows are available in the example_workflows/ directory

WanAnimatePlus exposes a complete workflow chain to avoid cross-package object mixing with the original WanVideoWrapper nodes.

Core nodes:

• WanAnimatePlus ModelLoader
• WanAnimatePlus VAELoader
• WanAnimatePlus TextEncodeCached
• WanAnimatePlus ClipVisionEncode
• WanAnimatePlus ContextOptions
• WanAnimatePlus AnimateEmbeds
• WanAnimatePlus Sampler / WanAnimatePlus Samplerv2
• WanAnimatePlus Scheduler / WanAnimatePlus Schedulerv2
• WanAnimatePlus Decode / WanAnimatePlus Encode
• WanAnimatePlus LoraSelect / WanAnimatePlus LoraSelectMulti / WanAnimatePlus SetLoRAs
• WanAnimatePlus BlockSwap / WanAnimatePlus SetBlockSwap
• WanAnimatePlus TorchCompileSettings
• WanAnimatePlus SamplerExtraArgs
• WanAnimatePlus Uni3C ControlnetLoader / WanAnimatePlus Uni3C Embeds
• WanAnimatePlus Bernini
• WanAnimatePlus SCAIL_2 Embeds

Core node, replaces the original WanVideoAnimateEmbeds.

New inputs:

Other inputs are identical to the original WanVideoAnimateEmbeds: vae, width, height, num_frames, ref_images, pose_images, face_images, bg_images, mask, start_ref_image, clip_embeds, etc.

Generates condition latents from source video, reference video, and/or reference images for Bernini models.

Inputs:

The task (v2v, rv2v, r2v, t2v) is automatically inferred from which inputs are connected.

Creates SCAIL-2 conditioning for WanAnimatePlus sampling. Use this node with SCAIL-2 checkpoints that include the pose and mask streams.

Inputs:

For short generations, context windows are optional. For long generations or low VRAM, context windows are recommended. In context-window mode, single-frame prefix references remain visible through the SCAIL-2 reference stream. Legacy canvas prefixes and transition latents are prepended for model context and removed before overlap fusion.

For SCAIL-2, the default single_frame_prefix_encoding mode does not expand or trim the output for prefix_frames. If transition_video is connected, the front canvas expands by 21 pixel frames and those 21 frames are trimmed after decoding. With single_frame_prefix_encoding disabled, prefix_frames use the legacy 37 front pixel-frame canvas layout, with transition frames placed at frames 17-36 when transition_video is also connected.

When bg_image is connected in animation mode, it consumes the first prefix slot and is handled as the first prefix image. The node internally adds a white mask for that background image only; user-provided prefix_frames and prefix_mask are then limited to four images each. In replacement mode, bg_image is ignored.

The WanAnimatePlus ModelLoader supports mxfp8 weight quantization for reduced VRAM usage.

Hardware requirements: Hardware-accelerated MXFP8 matmul requires an NVIDIA Blackwell GPU (compute capability >= 10.0, e.g., RTX 5090 / B100 / B200). On non-Blackwell GPUs, MXFP8 weights are automatically dequantized to BF16 at load time for normal inference (no VRAM savings, but the model runs correctly).

Auto-detection: When the quantization dropdown is set to disabled, MXFP8 weights are auto-detected from the state dict by scanning for float8_e8m0fnu block scale tensors. No manual selection is needed.

Installation (optional): Hardware acceleration requires comfy-kitchen:

pip install comfy-kitchen==0.2.10

This package is not a hard dependency — it is only needed for Blackwell GPU acceleration. The model works without it via the dequantization fallback.

MXFP8 quantization support is derived from comfy-kitchen (Apache 2.0).

ComfyUI-WanAnimatePlus/
├─ wanvideo/                 # WanVideo core model code
├─ nodes.py                  # Core WanAnimatePlus embeds / encode / decode nodes
├─ nodes_sampler.py          # Core WanAnimatePlus sampler / scheduler nodes
├─ nodes_model_loading.py    # Core WanAnimatePlus model / VAE / LoRA / block swap nodes
├─ context_windows/          # Context-window scheduling
├─ cache_methods/            # Cache acceleration
├─ utils.py                  # Shared utilities
├─ docs/
│  └─ images/                # Documentation images
├─ example_workflows/        # Example workflows
├─ __init__.py               # Node registration entry point
├─ pyproject.toml
├─ requirements.txt
└─ LICENSE

• Verify the repo path is ComfyUI/custom_nodes/ComfyUI-WanAnimatePlus
• Ensure the original ComfyUI-WanVideoWrapper is also installed
• Restart ComfyUI and search for WanAnimatePlus in the node list

No. All node names use the WanAnimatePlus prefix, completely avoiding conflicts with the original WanVideo prefixed nodes. Both can be installed simultaneously.

3 is recommended. Up to 5 are accepted (excess is truncated). The node works with fewer than 3 as well, but the coverage range will be smaller.

Input is automatically cropped to 21 frames (padded with the first frame if insufficient).

Modified from kijai/ComfyUI-WanVideoWrapper. Deep respect to the original author for their tremendous contributions to the WanVideo ecosystem.

• Bilibili: @wuwukasi
• Email: wuwukawayi@gmail.com

If you find this project helpful, consider supporting me! Your support is what keeps this project going and motivates me to continue improving it.

Every contribution, no matter how small, means a lot and helps me dedicate more time to development and new features. Thank you!

This project is an independently maintained fork / derivative project based on kijai/ComfyUI-WanVideoWrapper and is released under the Apache License, Version 2.0. Thanks again to kijai and the original contributors for their work.

Modified portions and newly added code are Copyright (c) 2026 wuwukasi/wuwukaka. See NOTICE for attribution and detailed modification notice requirements. Downstream projects that use or modify the WanAnimatePlus additions should preserve the copyright/modification notices and include a detailed notice in their README, NOTICE file, or equivalent attribution document. That notice should identify the wuwukasi/wuwukaka-derived modules, files, or feature areas and describe any downstream modifications made to those portions.

The wuwukasi/wuwukaka additions include WanAnimatePlus-specific node registration/renaming and integration code, prefix/transition video conditioning, Bernini in-context conditioning, SCAIL-2 embeds support including prefix-mask handling and sampler freeze/prepend integration, EverAnimate embeds support, sampler/context-window changes, cache/inference safeguards, and related custom-op handling.