archgod94/tellcue-models

Tellcue models

On-device models downloaded by the Tellcue video editor when a feature is first used.

rife-v4.17-lite.onnx

Frame interpolation for smooth slow motion.

• Network and weights: RIFE v4.17 lite from Practical-RIFE (hzwer, MIT). Weights file flownet_v4.17.lite.pkl from the vs-rife model release (HolyWu, MIT).
• Converted to ONNX (opset 17, fp32) with tool/export_rife.py in the Tellcue repository. Changes from the original: the flow grid is built inside the graph, so any size works; the flow scale is an input; the warps of blocks 1 to 3 run at the block size instead of full size, which is about twice as fast with the same result (within 0.07 dB on real frames).
• Inputs: img0, img1 float32 [1, 3, H, W] RGB 0..1 (H and W multiples of 32 / scale), timestep float32 [1] in 0..1, scale float32 [1] (1, 0.5 or 0.25). Output: out float32 [1, 3, H, W].
• SHA-256: c0887c5a56dba4f18a766f2c826ae889d92edb719b40e0732817da1241125889

deepfilternet3.onnx

Speech noise reduction (48 kHz).

• Network and weights: DeepFilterNet3 by Hendrik Schroeter, models/DeepFilterNet3_onnx.tar.gz (MIT or Apache-2.0).
• The three exported graphs (enc, erb_dec, df_dec) joined into one with tool/export_dfn.py in the Tellcue repository; numbers unchanged.
• Inputs: feat_erb float32 [1, 1, S, 32], feat_spec float32 [1, 2, S, 96] (libDF features, 2 frames look-ahead applied). Outputs: m [1, 1, S, 32] ERB gains, coefs [1, S, 96, 10] deep filter taps, lsnr [1, S, 1].
• SHA-256: 9223f1332c520e942b54a65a5e6e728ae56db7624b9da48c11f4585dc1afddce

migan-pipeline-v2.onnx

Object removal (inpainting) for photos.

• MI-GAN (Picsart AI Research, MIT), the authors' pre-built ONNX pipeline migan_pipeline_v2.onnx as published at andraniksargsyan/migan, copied unchanged.
• Inputs: image uint8 [1, 3, H, W] RGB, mask uint8 [1, 1, H, W] (0 = remove, 255 = keep), any size. Output: result uint8 [1, 3, H, W]; only the masked area changes.
• SHA-256: 6f1f3530a1a2324b19752018ce756088b07973cda8d7d890034ace5c8a48c40b

realesr-general-x4v3.onnx

Photo upscaling (x4) with light denoising.

• Real-ESRGAN by Xintao Wang et al. (BSD-3-Clause), weights realesr-general-x4v3.pth from the v0.2.5.0 release (SRVGGNetCompact, 64 features, 32 convolutions).
• Converted to ONNX (opset 17, fp32, any size) with tool/export_esrgan.py in the Tellcue repository; output matches PyTorch within 6e-6.
• Input: input float32 [1, 3, H, W] RGB 0..1. Output: output float32 [1, 3, 4H, 4W].
• SHA-256: e859a54006946df52e810ea79c6ba71ef522ed1b6f8b04ae1646d66edddce293

See LICENSE for the original licences.