SnJake/Baikal-Swin-Anime

SnJake Anime Upscale

A experimental lightweight upscaler (x2) for anime/illustration images. Designed for clean, sharp results with minimal artifacts. V2 is slightly sharper and removes edge noise artifacts.

Examples

How to use in ComfyUI

The model is designed to work with the SnJake Anime Upscale ComfyUI node.

1. Install the node from GITHUB REPO.
2. Download the weights from this repository.
3. Place the file(s) into ComfyUI/models/anime_upscale/.
4. Select the weights in the node dropdown and run the workflow.

Training Details

V1:

• Dataset: 40,000 images from Danbooru2024: https://huggingface.co/datasets/deepghs/danbooru2024
• Validation: 600 images
• Epochs: 70

V2:

• Slightly sharper output, no edge noise artifacts.
• Epochs: 20
• Dataset: 49,606 images from Danbooru2024: https://huggingface.co/datasets/deepghs/danbooru2024
• Perceptual backbone: Custom SimSiam pre-trained convnextv2_tiny (Experimental)
• Loss schedule: gradual ramp‑in of perceptual/auxiliary losses for stable training.

V2.1:

• Removed Nearest from resample_methods
• Epochs: 30

V2.2:

• Epochs: 40 (For now)

V3 (SwinFIR):

• Epochs: Stage 1 - 20; Stage 2 - 18
• Perceptual backbone: Custom SimSiam pre-trained convnextv2_base (Experimental).
• Dataset: ~50,000 images from Danbooru2024: https://huggingface.co/datasets/deepghs/danbooru2024

V3.1 (Current Best - SwinFIR): 🏆

• Epochs: Stage 1 (Charbonnier) - 20; Stage 2 (GAN Fine-tuning) - 16.
• Perceptual backbone: Reverted to the robust ImageNet-pretrained convnextv2_base.fcmae_ft_in22k_in1k for superior high-frequency feature extraction.
• Dataset: ~50,000 images from Danbooru2024: https://huggingface.co/datasets/deepghs/danbooru2024

Training code is included in training_code/ for reference.

Disclaimer

This project was made purely for curiosity and personal interest. The code was written by GPT-5.2 Codex.