ComfyUI-FL-SeedVR2

FL SeedVR2

Native ComfyUI image restoration and upscaling nodes for the six-block SeedVR2 1.4B model.

This pack uses ComfyUI's built-in SeedVR2 model, VAE, sampler, tiling, device management, and offloading. It does not bundle a second SeedVR2 implementation or patch ComfyUI at import time.

Features

• One-step restoration - Denoise, deblur, sharpen, and restore image detail
• 1x to 8x scaling - 4x is the recommended default
• Automatic model setup - Downloads the pinned transformer and VAE into the standard ComfyUI model folders
• Verified downloads - Resumable transfers with file-size, model-header, and SHA-256 validation
• ComfyUI-native execution - Uses the built-in SeedVR2 model path and memory management
• Batch and alpha support - Processes image batches independently and preserves RGBA alpha
• FL styling - Matches the title and body colors used by other FL node packs

Nodes

| Node | Description | |------|-------------| | FL SeedVR2 1.4B Loader | Loads the pinned 1.4B transformer and VAE, with optional automatic download | | FL SeedVR2 1.4B Upscale | Restores and upscales an image or independent image batch |

Both nodes appear under FL/SeedVR2.

Installation

ComfyUI Manager

Search for ComfyUI-FL-SeedVR2 and install it, then restart ComfyUI.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-SeedVR2.git

Restart ComfyUI after installation. This pack uses the dependencies already provided by ComfyUI.

Quick Start

1. Add FL SeedVR2 1.4B Loader.
2. Add FL SeedVR2 1.4B Upscale.
3. Connect the loader's MODEL and VAE outputs to the upscale node.
4. Connect an IMAGE and queue the prompt.
5. On first use, allow the loader to finish downloading and verifying both model files.

The upscale node outputs a standard ComfyUI IMAGE. An editable four-node workflow is included at examples/seedvr2_1_4b_upscale.json.

Recommended settings:

scale_multiplier: 4
color_correction: none
vae_tile_size:    512

Use a VAE tile size of 256 if encoding or decoding runs out of memory. SeedVR2 is a one-step model, so the node intentionally fixes the sampler settings required by the model.

Models

With download_if_missing enabled, the loader downloads only these pinned artifacts:

| Artifact | Destination | Download size | |----------|-------------|---------------| | seedvr2_distill_6L_1.4B_sharp_fp16_comfyui.safetensors | ComfyUI/models/diffusion_models/ | 2.89 GB | | seedvr2_ema_vae_fp16.safetensors | ComfyUI/models/vae/ | 501 MB |

Downloads begin only when the loader executes. Partial transfers can resume, and completed files are verified before being moved into place. Model paths registered through extra_model_paths.yaml are also supported.

The transformer must be the _comfyui variant. The plain checkpoint does not contain the fixed conditioning tensors required by ComfyUI.

Troubleshooting

unexpected keys, starting with blocks.0.attn.proj_out.txt.bias

This means ComfyUI is using an older SeedVR2 architecture, not that the automatic download is corrupt. Update ComfyUI to 0.28.0 or newer, restart it, and queue the loader again. The downloaded transformer and VAE are valid and do not need to be deleted or downloaded again.

Key Parameters

• download_if_missing - Download both pinned model files when they are not already registered with ComfyUI
• scale_multiplier - Output scale from 1x to 8x; 2x to 4x is the model's intended range
• seed - Noise seed for repeatable output
• color_correction - none, lab, wavelet, or adain
• vae_tile_size - 512 by default, or 256 for lower peak memory use

Requirements

• ComfyUI 0.28.0 or newer with native SeedVR2 support
• Python 3.9 or newer
• About 3.4 GB of model storage
• Network access during the first execution when automatic download is enabled

MPS was verified end to end. CUDA, ROCm, and CPU use the same native ComfyUI path but were not tested for the first release.

Current Scope

Image restoration, RGB and RGBA input, independent image batches, automatic model setup, and native ComfyUI device/offload behavior are included. Video, GGUF, block swapping, custom attention selection, and training are outside the scope of this release.

Development

From the ComfyUI root:

./venv/bin/python -m pytest --import-mode=importlib -q \
  custom_nodes/ComfyUI-FL-SeedVR2/tests
ruff check custom_nodes/ComfyUI-FL-SeedVR2
node --check custom_nodes/ComfyUI-FL-SeedVR2/web/appearance.js

License

Apache-2.0. SeedVR2 was created by ByteDance Seed, and the 1.4B student model was published by lvladikov.