FlashVSR Video Upscaling Workflow

A ComfyUI workflow for video frame interpolation + AI video super-resolution, mainly tested on anime / illustration-style videos.

⭐ Recommended Workflow

For low-FPS videos, I recommend:

Input Video → Frame Interpolation ×2 → FlashVSR → Video Combine

For example:

16 FPS → 32 FPS → FlashVSR 4× → 32 FPS

⚠️ Interpolate BEFORE FlashVSR

After testing both processing orders, I found that interpolating the video before FlashVSR significantly reduces visible motion ghosting in low-FPS sources.

❌ FlashVSR → VFI

• Faster at 4K

• More noticeable ghosting/trailing around fast-moving objects

• Particularly noticeable around hands and arms

✅ VFI → FlashVSR

• Much less noticeable ghosting in my testing

• Better temporal consistency

• Recommended for low-FPS sources

• Significantly more expensive at 4K

Important: This does not completely eliminate ghosting. FlashVSR itself can still produce subtle temporal artifacts in difficult high-motion scenes. The recommended order simply makes them much less noticeable in my testing.

⚡ Performance Settings

After testing the Advanced FlashVSR settings on an RTX 5080 16GB, I found a much faster configuration than the default settings.

Recommended 16GB VRAM settings

tile_size     = 224
tile_overlap  = 16
sparse_ratio  = 1.8
kv_ratio      = 2.2

Other settings can remain at their defaults.

With these settings, my ~3-second 16 FPS source took approximately 8 minutes for 4× 4K upscaling using:

VFI ×2 → FlashVSR

This is a personal benchmark on an RTX 5080 16GB, so actual performance will vary depending on GPU, VRAM, source resolution, video length and other settings.

⚠️ VRAM / Tile Size

Larger tile_size does not necessarily mean faster processing.

In my testing:

• 224 → ~73% VRAM usage and very good performance

• 256 → ~95–100% VRAM usage, but little/no speed improvement

• 288 → caused severe VRAM pressure and was significantly slower

Once VRAM is exhausted, system RAM starts being used and performance can drop dramatically.

For 16GB GPUs, 224 is currently my recommended starting point.

🎨 Color Fix

I recommend keeping color_fix enabled.

Without color correction, I observed noticeable color shifts in dark scenes, including changes to the appearance of the night sky/background.

👻 About Ghosting

Some subtle ghosting may still remain, especially during very fast motion.

This appears to be a limitation of the current FlashVSR processing rather than something that can be completely solved by changing the VFI model.

The best result I found was:

VFI ×2 → FlashVSR

rather than:

FlashVSR → VFI

The latter is faster at 4K, but produces more obvious motion artifacts in my testing.

🧪 My Test

Source video:

• Anime / illustration style

• 16 FPS

• ~3 seconds

Recommended pipeline:

Load Video
    ↓
VFI ×2
    ↓
FlashVSR 4×
    ↓
Video Combine

Tested on:

RTX 5080 16GB

Optimized 4K processing time:

~8 minutes

Results:

• ✅ No obvious visual artifacts in normal viewing

• ✅ Much less noticeable motion ghosting

• ✅ Good 4K detail

• ✅ Significantly faster than the default Advanced settings in my testing

• ⚠️ Subtle ghosting can still be visible when examining individual frames closely

⚠️ Before You Render

Please test a short clip first.

4K video processing can become extremely expensive, particularly when frame interpolation is performed before super-resolution.

The optimal settings depend on:

• Source FPS

• Source resolution

• Motion speed

• Scene complexity

• GPU / VRAM

• VFI model

For GPUs with limited VRAM, 2K is recommended as a starting point.

Requirements

• ComfyUI

• ComfyUI-Frame-Interpolation

• FlashVSR v1.1

• A compatible VFI model (ATM-VFI, RIFE, GMFSS Fortuna, etc.)

• NVIDIA GPU

GMFSS Fortuna requires additional dependencies such as CuPy.