Free.ai Video Upscaler

Self-hosted Real-ESRGAN x4 video upscaler with temporal smoothing — a free, GPU-backed alternative to Topaz Video Upscaler.

Powers the Free.ai video upscaler tool.

What it does

Takes a video, upscales every frame 4× with Real-ESRGAN, applies a temporal smoothing pass to eliminate frame-to-frame flicker, then re-encodes preserving the source audio. The output is competitive with commercial video upscalers on per-frame detail and temporal coherence.

Why temporal smoothing matters

Frame-by-frame neural upscalers (the cheap way) introduce high-frequency flicker because each frame is upscaled independently — a slight texture detail in pixel (12,34) might be reconstructed differently by the model on two consecutive frames, producing a shimmer. Topaz solves this by training a video-aware model with temporal context.

We solve it cheaper: a 3-tap weighted average over consecutive output frames (0.15 × prev + 0.7 × current + 0.15 × next) kills the high-freq flicker without smearing motion. A scene-cut detector (per-pixel mean-abs-diff > 60) skips smoothing when adjacent frames are too different — preserves crisp edits, scene transitions, and very fast motion.

Install

# 1. Install torch with the right CUDA version for your driver.
#    Common choice for CUDA 12.x systems:
pip install torch==2.5.1 torchvision==0.20.1 \
    --index-url https://download.pytorch.org/whl/cu121

# 2. Install the upscaler:
pip install free-video-upscaler

# 3. Patch basicsr's torchvision import (newer torchvision removed
#    `functional_tensor` — known basicsr issue, no fix released yet):
python -c "import basicsr.data.degradations as _d, re, pathlib; \
    p = pathlib.Path(_d.__file__); \
    p.write_text(p.read_text().replace('torchvision.transforms.functional_tensor', \
                                       'torchvision.transforms.functional'))"

# 4. Download model weights:
mkdir -p ~/.realesrgan/weights && cd ~/.realesrgan/weights
wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth
wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth

CLI

free-upscaler --input my-video.mp4 --output upscaled.mp4 --scale 2

free-upscaler [-h] --input INPUT --output OUTPUT
              [--scale {2,4}] [--model {x4plus,anime}]
              [--tile TILE] [--no-temporal] [--fp32]
              [--weights WEIGHTS] [--quiet] [--version]

Python API

from free_video_upscaler import upscale_video

upscale_video(
    "in.mp4", "out.mp4",
    scale=2,             # 2 or 4
    model="x4plus",      # "x4plus" or "anime"
    temporal_smooth=True,
    progress=lambda i, n: print(f"{i}/{n}"),
)

How it compares

Scaling notes

• Real-ESRGAN x4plus uses ~6 GB VRAM at 1080p with tile=256. Drop tile to 128 for 8 GB cards, or pass tile=0 for no tiling on big GPUs.
• Anime variant runs ~4× faster (6-block vs 23-block RRDBNet) and uses less VRAM. Use it for line-art / cel content.
• For 60 s+ clips on a single GPU, the package extracts + upscales + re- encodes synchronously; expect 10–20 minutes for a 1080p 60 s clip on a consumer GPU. Run it as a background job for anything longer.

License

Apache-2.0. Real-ESRGAN model weights are themselves Apache-2.0 (see Tencent ARC's Real-ESRGAN repo).