ComfyUI-FlashVSR-Plus-Tiled

FlashVSR+ Tiled for ComfyUI

Spatially tiled video upscaling for NVIDIA GPUs with limited VRAM. Processing overlapping spatial tiles reduces the working size of each inference pass, making larger output resolutions more practical. Memory use still depends on model, tile dimensions, temporal settings, and clip length.

Native ComfyUI VIDEO and IMAGE super-resolution powered by FlashVSR+. This fork adds spatial tiling with overlap blending for larger output frames, alongside audio preservation, incremental video processing, and INT8 ConvRot support. Requires an NVIDIA CUDA GPU.

Models · Example workflows · Validation · Spatial tiling PRD

What this fork adds

• Independent spatial tile width and height in output pixels, with configurable overlap.
• Three overlap blends: cosine, linear, and Gaussian.
• Streaming VIDEO processing with temporary float16 tile segments, audio preservation, and final stitching.
• Console output showing output dimensions and the tile grid, with one progress bar per tile. The bar includes DiT work and completed Tiny decode segments; percentages are an estimate, not a timing guarantee.
• Pixel-based VAE tile controls and a Tiny decode interval expressed in output video frames.
• Dependent spatial controls are disabled when spatial_tiling is off; VAE tile height, width, and overlap are disabled when vae_tiled is off.
• Separate IMAGE and AUDIO outputs on Tiled Upscale Video, in addition to VIDEO and JSON information.
• Example workflows using the tiled node classes, including the workflow embedded in the portrait PNG.

Video outputs

The output order is video, info, images, audio, fps. The fps output reports the output video frame rate. The original video and info port indices are preserved for existing connections.

images contains the full upscaled frame batch, decoded from the saved output video. It is materialized only when that port is connected in the submitted graph. A float32 RGB batch needs approximately 12 bytes per pixel per frame: 192 frames at 3840×2160 alone need about 17.8 GiB of host RAM, with additional decoding overhead. Keep this port disconnected when you only need a video file.

audio contains the soundtrack from the completed output video, or None when there is no audio track. An audio-only connection extracts the waveform without materializing the image batch.

After updating Python code, restart ComfyUI and refresh its browser tab to load the new output ports. Save your current workflow before refreshing.

Installation

All four nodes include Tiled in their searchable names and category. Use the tiled loader, options, and upscale nodes together; the original T8 loader returns an incompatible model handle for spatial VIDEO tiling.

To install manually, run from the ComfyUI directory and use ComfyUI's own Python to install dependencies:

git clone --branch spatial-tiling https://github.com/hdmotionpersonal/ComfyUI-FlashVSR-Plus-Tiled.git custom_nodes/ComfyUI-FlashVSR-Plus-Tiled
python -m pip install -r custom_nodes/ComfyUI-FlashVSR-Plus-Tiled/requirements.txt

The dependency list does not include PyTorch or Transformers and will not replace them. transformers==4.57.1 was verified in an isolated environment. On Windows, Triton must match the installed PyTorch version; the verified combination is PyTorch 2.7.0+cu128 with triton-windows==3.3.1.post21.

Quick start

1. Import the VIDEO workflow or video-frame IMAGE workflow. Copy the sample video into ComfyUI/input. For a still photo, use the single-image workflow.
2. In FlashVSR+ Tiled Model Loader, choose a mode, precision, attention backend, and decoder. For a first run, use turbo, int8_convrot, and tiny; leave auto_download enabled if the model files are not already installed. INT8 requires Comfy Kitchen >=0.2.35.
3. Set the scale or target dimensions and queue the workflow. The VIDEO path preserves audio, frame rate, and upstream trim/crop settings. Use VIDEO for longer clips.

| Node | Purpose | | --- | --- | | FlashVSR+ Tiled Model Loader | Load Turbo, Base, or V1 weights in BF16, FP16, or INT8 ConvRot with Tiny or Wan VAE decoding | | FlashVSR+ Tiled Upscale Video | Upscale native VIDEO input with incremental reading and encoding | | FlashVSR+ Tiled Upscale Frames | Upscale an RGB image or a batch of video frames and return run information | | FlashVSR+ Tiled Options | Configure spatial tiles, overlap blending, temporal processing, cache offloading, decoding, and color correction |

Spatial tiling

Connect FlashVSR+ Tiled Options to either tiled upscale node and enable spatial_tiling. The tile size and overlap are measured in output-canvas pixels. Tile dimensions must be multiples of 128; overlap can be set in 8-pixel steps. The canvas is padded to fit complete tiles and cropped back to the requested output dimensions.

Spatial tile size, overlap, and blend controls are visibly disabled while spatial_tiling is off; their values are preserved when toggling the option.

Choose one of three overlap blend methods:

• cosine — smooth cosine feathering across shared edges; recommended starting point.
• linear — linear feathering across shared edges.
• gaussian — center-weighted feathering with adjustable spatial_gaussian_sigma.

Before tiled inference, the ComfyUI console reports the tile size, final output dimensions, and tile grid (columns × rows). Weights from overlapping tiles are normalized by their sum. Canvas edges retain full weight where there is no neighboring tile. The console shows a separate tqdm progress bar for each tile, labeled with its index and canvas position. Tile dimensions are printed in the spatial plan.

| Option | Default | Notes | | --- | ---: | --- | | spatial_tiling | false | Enable spatial tile inference. | | spatial_tile_height | 512 | Tile height in output pixels; step 128. | | spatial_tile_width | 512 | Tile width in output pixels, beside height and independently adjustable; step 128. 0 follows the height for older workflows. | | spatial_overlap | 128 | Shared output pixels on both axes; step 8, and less than half of both tile dimensions. | | spatial_blend | cosine | cosine, linear, or gaussian. | | spatial_gaussian_sigma | 0.25 | Controls the Gaussian blend profile. | | decode_every_frames | 112 | Tiny decoder interval in output video frames. Its valid step is 4 × temporal_compression; 0 selects full-clip/auto decoding. | | vae_tile_height / vae_tile_width | 240 / 416 | VAE decoder tile dimensions in output pixels; step 8. | | vae_overlap | 120 | VAE decoder overlap in output pixels; step 8 and less than both VAE tile dimensions. |

Spatial VIDEO tiling currently requires the Tiny decoder and a positive decode_every_frames value. The node adjusts the frame step when temporal_compression changes. VAE tile dimensions and overlap are shown in output pixels and converted internally to Wan VAE latent units. Spatial VIDEO tiling with Wan VAE is not supported yet. It writes temporary float16 tile segments to disk; temporary storage grows with clip length and the total tile area. Each tile uses its crop from the global latent-noise coordinates. Tile boundaries can still differ because each tile has a smaller attention context. Color correction in spatial VIDEO mode is applied per tile; disable it if you see color shifts at seams.

A GPU smoke test on an RTX 4070 Ti SUPER used V1 INT8 ConvRot, Tiny, SDPA, 512-pixel tiles, and 128-pixel overlap. The 33-frame demo clip completed at its full duration with cosine blending. Cosine, linear, and Gaussian each completed an 8-frame GPU run. The full-resolution output was 600×336; the padded canvas produced two horizontal tiles with exactly 128 pixels of overlap.

Single-image upscaling

Import the single-image workflow, choose a static PNG or JPEG in Load Image, and run it to save a PNG through Save Image. Copy the sample image into ComfyUI/input to reproduce the example, or load the tested output PNG with its embedded workflow. See the API workflow for programmatic use.

Defaults are turbo / int8_convrot / tiny, 4× scaling, fixed seed 1, and color correction enabled. BF16 and INT8 GPU checks covered a portrait, a 1080p photo, and a text chart. Photo enhancement passed on these samples; small text and flat graphics are not guaranteed to remain exact. Process unrelated images individually and preserve alpha separately. See comparisons and acceptance scope.

Models and acceleration

Models are discovered in ComfyUI/models/flashvsr_plus and the node's ckpts directory. You can also set a custom model root in the loader. See the model repository for the complete file layout. Downloads use pinned revisions and SHA-256 verification.

• INT8 ConvRot: each preconverted DiT is approximately 1.45 GB. It runs with the matching LQ projection and decoder; the original DiT is not required. Transformer-block linear layers use Comfy Kitchen W8A8, while other components remain in floating point.
• Attention: official BSA, Triton, and SDPA backends are supported. auto selects an available backend; Triton is recommended for high-resolution output. SDPA rejects oversized masks before allocation. Native BSA has not been verified on Windows.
• Segmented decoding: Tiny accepts decode_every_frames values in output frames, in multiples of 4 × temporal_compression. Wan VAE uses full-clip decoding and requires decode_every_frames=0.

INT8 primarily reduces weight memory; speed and image quality depend on the content and hardware. IMAGE output requires RAM for the full batch. VIDEO uses a constant-frame-rate timeline; normalize variable-frame-rate sources to CFR first. Audio that cannot be copied into MP4, such as PCM, is encoded as AAC. Model weights return to CPU after each request; keep_loaded retains the CPU copy between requests.

Validation

Earlier validation recorded 48 passing regression tests. This is a historical baseline, not a test result for every subsequent change. The latest IMAGE/AUDIO outputs and decode-progress changes have passed syntax checks but have not yet received an end-to-end GPU validation. Spatial GPU checks covered all three blend methods, plus a full 33-frame cosine-blended video run on an RTX 4070 Ti SUPER. See validation scope and the 1080p validation clip for the existing full-resolution validation records.

Community and resources

Bilibili · YouTube · API · Free gallery · Online AI apps · ComfyUI bundle · Hugging Face

Credits and license

Based on OpenImagingLab/FlashVSR_Plus and FlashVSR V1.1. INT8 linear operations use Comfy Kitchen. The original ComfyUI integration and derived checkpoints are maintained by T8star. This spatial-tiling fork is maintained at hdmotionpersonal/ComfyUI-FlashVSR-Plus-Tiled. Code and model files are licensed under Apache-2.0; see LICENSE and NOTICE.