ImprovedSRV5 x4
Inference-only EMA weights for the final 50,000-step ImprovedSRV5 Stage A
model used by the lckly/super-resolution-v5 academic demo.
Identity
- Architecture: canonical RRDBNet x4 (
ImprovedSRV5) - Parameters: 16,697,987
- State keys: 702
- Training step: 50,000
- DIV2K validation PSNR: 28.756685 dB
- DIV2K validation SSIM: 0.827441
- File:
improved_srv5_x4_step_0050000.pth - Size: 67,042,203 bytes
- SHA-256:
8915ab974187524514593e99a1fce040106b6b05858d3cb485306a6dfaf8c3a0
The deployment file was derived byte-for-byte from the final checkpoint's
ema_model state. Its source checkpoint SHA-256 is
623fc873b277aa811d54731077487bb350fdbff98224c2cb313facbe4f7c02aa.
Loading
Load with torch.load(..., map_location="cpu", weights_only=True), read the
model_state_dict field, and use strict=True. Consumers should verify the
file size and SHA-256 before deserialization.
Intended use
This model performs RGB single-image super-resolution at scale x4. It is published for the associated undergraduate project demonstration and reproducible evaluation. It is not intended for medical, forensic, or other high-stakes reconstruction; generated details must not be treated as factual evidence.