ComfyUI-FramePackWrapper_PlusOne

日本語

ComfyUI-FramePackWrapper_PlusOne is a fork derived from ComfyUI-FramePackWrapper and ComfyUI-FramePackWrapper_Plus, containing FramePack's single-frame inference node (with kisekaeichi support).

This repository was forked for public release at the request of @tori29umai0123 as requested here.

Features

• 1-Frame Inference: Supports basic single frame inference and the kisekaeichi method. For technical details, please refer to the musubi-tuner documentation.
• F1 Sampler Support: Uses the improved F1 video generation method for higher quality and better temporal coherence
• LoRA Integration: Full support for HunyuanVideo LoRAs with proper weight handling and fusion options
• Timestamped Prompts: Create dynamic videos with changing prompts at specific timestamps
• Flexible Input Options: Works with both reference images and empty latents for complete creative control
• Resolution Control: Automatic bucket finding for optimal video dimensions
• Blend Control: Smooth transitions between different prompts at timestamps

Not yet supported

• 1-Frame Inference: f-mc (one frame multi-control) is not supported yet.

Installation

1. Clone this repository into your ComfyUI custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/xhiroga/ComfyUI-FramePackWrapper_PlusOne.git

2. Install the required dependencies:

pip install -r requirements.txt

3. Download the necessary model files and place them in your models folder:

• FramePackI2V_HY: HuggingFace Link
• FramePack_F1_I2V_HY: HuggingFace Link

Model Files

Main Model Options

• FramePackI2V_HY_fp8_e4m3fn.safetensors - Optimized fp8 version (smaller file size)
• FramePackI2V_HY_bf16.safetensors - BF16 version (better quality)

Required Components

• CLIP Vision: sigclip_vision_384
• Text Encoder and VAE: HunyuanVideo_repackaged

Usage

See example_workflows.

1-Frame / LoRA @tori29umai	1-Frame / LoRA @kohya-ss	Kisekaeichi / LoRA @tori29umai

License

MIT License

Changelog

v2.0.1 - Bug Fixes (2025-08-26)

Bug Fixes

• Fixed generation failure on second and subsequent runs: Resolved a dtype mismatch error (BFloat16 vs Float32) that caused failures when running generation multiple times in succession.

v2.0.0 - Full musubi-tuner Compatibility (2025-08-08)

Achieved complete compatibility with musubi-tuner specifications to improve inference result consistency when using multiple reference images.

Breaking Changes

1. The denoise_strength of workflows created up to v0.0.2 may be reset to 0. After updating the node, please manually reset it to 1.0.

2. Existing kisekaeichi workflows cannot be reused with this update. The reference image input has been changed to a list format to support multiple reference images. Please refer to Oneframe_kisekae_V2.json in the example_workflows folder for the updated workflow structure.

Major Changes

1. Improved Embedding Integration Method

• ❌ Previous: Weighted average integration (70% input image, 30% reference images)
• ✅ New: musubi-tuner compatible processing (using first reference image embedding)

2. Unified Latent Combination Structure

• ❌ Previous: Separate management of input and reference images before combination
• ✅ New: Direct control_latents combination following musubi-tuner specification

  control_latents = [input_image, reference_image1, reference_image2, ..., zero_latent]
  clean_latents = torch.cat(control_latents, dim=2)

3. Optimized Mask Application Timing

• ❌ Previous: Individual application before latent combination
• ✅ New: Mask application after clean_latents generation (musubi-tuner specification)

4. Dynamic Index Setting Processing

• ❌ Previous: Fixed clean_latent_indices configuration
• ✅ New: Dynamic application of control_indices parameters

  # control_index="0;7;8;9;10" → clean_latent_indices = [0, 7, 8, 9, 10]
  while i < len(control_indices_list) and i < clean_latent_indices.shape[1]:
      clean_latent_indices[:, i] = control_indices_list[i]

5. Improved latent_indices Initialization

• ❌ Previous: ComfyUI-specific initialization method
• ✅ New: musubi-tuner specification initialization

  latent_indices = torch.zeros((1, 1), dtype=torch.int64)
  latent_indices[:, 0] = latent_window_size  # default value
  latent_indices[:, 0] = target_index        # parameter application

Expected Benefits

• Improved Inference Consistency: Generate identical results to musubi-tuner with same reference images and parameters
• Stabilized Multi-Reference Processing: More stable quality through accurate index management
• Parameter Compatibility: Correct functionality of musubi-tuner's control_index and target_index parameters

Technical Details

This update ensures the following processing flow matches musubi-tuner completely:

1. Control Image Processing: Sequential processing of multiple images specified by --control_image_path
2. Index Management: Dynamic application of --one_frame_inference="control_index=0;7;8;9;10,target_index=5"
3. Embedding Processing: Implementation simulating section-wise individual processing
4. Mask Application: Unified mask processing after clean_latents construction

Credits

• FramePack: @lllyasviel's original implementation.
• ComfyUI-FramePackWrapper: @kijai's original implementation.
• ComfyUI-FramePackWrapper_Plus: @ShmuelRonen's F1-supported fork.
• ComfyUI-FramePackWrapper_PlusOne: @tori29umai0123's 1-frame inference-supported fork.
• musubi-tuner: @kohya-ss's high-quality FramePack training and inference library