Request personal-use access to RestoreOshiV1
Access is reviewed manually. Personal, non-commercial use only. Do not redistribute or rehost the C3000 weights, including modified or converted versions. Read LICENSE.txt and the Availability section before requesting access.
Log in or Sign Up to review the conditions and access this model content.
RestoreOshiV1
Restore your oshi for a cleaner, more satisfying viewing experience on your beautiful screen.
RestoreOshiV1 is an experimental image-restoration model focused on degraded 3D VTuber recordings.
Its goal is to reduce compression artifacts and improve the clarity of characters, costumes, and stage details while preserving the appearance of the original footage.
Current model
The current version, V1, uses an adapted and fine-tuned RealPLKSR model.
- Task: Image restoration
- Scale: 1× — output dimensions match the input
- Processing: Individual frames
For example, a 1920 × 1080 input produces a restored 1920 × 1080 output.
Development
This project builds on an existing architecture and pretrained weights.
Project work includes adapting the model for restoration at the original resolution, preparing concert-specific training pairs and degradation examples, fine-tuning, and evaluating restoration quality.
Training targets are cleaner reference proxies, rather than guaranteed pristine ground truth.
Optional 4K upscaling
Some demonstrations include a separate upscaling stage:
Original frame → RestoreOshiV1 → 2× Upscale Model of your choice
The restoration model itself operates at 1× resolution. The additional Upscaling model produces the 4K output from a restored 1080p frame.
V2 combines restoration and 2× upscaling in a single model, trained from scratch. On a 240-frame RTX 4090 benchmark, the RealPLKSR V2 model runs at 2.2 s/frame versus 4.3 s/frame for V1 plus a separate upscaler, with 56% lower peak GPU memory. Visual evaluation is ongoing, and V2 weights are not yet released.
Evaluation and limitations
On 400 frames from a concert excluded from training, V1 reduced MSE by 42.4% (+2.4 dB PSNR) versus the degraded input. The reference is a cleaner compressed render, not a pristine master.
Before-and-after examples are selected qualitative comparisons and do not prove recovery of the original details.
- Fine facial features, hair, and textures may be altered.
- Severe blur and clipped highlights may remain.
- Results can vary between recordings and scenes.
- Frame-by-frame processing does not guarantee temporal consistency in video.
Availability
The final C3000 EMA inference weights are available in this repository through manually approved gated access.
- Sign in to Hugging Face, agree to the personal-use conditions and request access.
- After approval, download the release files from Files and versions and follow the download and usage instructions.
Personal, non-commercial use only. Redistribution of the model weights is prohibited, including converted or modified versions. Read LICENSE.txt for the full terms. Third-party components retain their original licenses.
The download contains the 1× C3000 restorer, its minimal inference code and usage instructions. Separate upscalers and V2 weights are not included.
For other permissions or questions, please contact me by email.
Acknowledgments
- Architecture: PLKSR and RealPLKSR by their authors and contributors.
- Initialization: pretrained weights from the darktable-org RealPLKSR ONNX conversion.
Example
Original YouTube Source: "Twinkle" from Stellive
Original YouTube Link: https://youtu.be/he2W_4xY5Zo
- Downloads last month
- -
Model tree for ThdPotato/RestoreOshiV1
Base model
darktable-org/upscale-realplksr-onnx