2xBHI_small_realplksr_small_pretrain

2xBHI_small_realplksr_small_pretrain

Scale: 2x
Network type: realplksr_small
Author: Philip Hofmann
License: CC-BY-4.0
Release: 21.05.2025
Purpose: 2x realplksr_small pretrain model with l1&mssim loss only.
Training iterations: 100'000
Description: A 2x realplksr_small pretrain model.

Visual Examples

Tensorboard Validation Graphs on BHI100

2xBHI_small_realplksr_small_dysample_pretrain

2xBHI_small_realplksr_small_dysample_pretrain

Scale: 2x
Network type: realplksr_small dysample
Author: Philip Hofmann
License: CC-BY-4.0
Release: 21.05.2025
Purpose: 2x realplksr_small dysample pretrain model with l1&mssim loss only.
Training iterations: 100'000
Description: A 2x realplksr_small dysample pretrain model.

Visual Examples

Tensorboard Validation Graphs on BHI100

2xBHI_small_realplksr_pretrain

2xBHI_small_realplksr_pretrain

Scale: 2x
Network type: realplksr
Author: Philip Hofmann
License: CC-BY-4.0
Release: 21.05.2025
Purpose: 2x realplksr pretrain model with l1&mssim loss only.
Training iterations: 100'000
Description: A 2x realplksr pretrain model.

Visual Examples

Tensorboard Validation Graphs on BHI100

2xBHI_small_realplksr_large_pretrain

2xBHI_small_realplksr_large_pretrain

Scale: 2x
Network type: realplksr_large
Author: Philip Hofmann
License: CC-BY-4.0
Release: 21.05.2025
Purpose: 2x realplksr_large pretrain model with l1&mssim loss only.
Training iterations: 100'000
Description: A 2x realplksr_large pretrain model.

Visual Examples

Tensorboard Validation Graphs on BHI100

2xBHI_small_realplksr_large_dysample_pretrain

2xBHI_small_realplksr_large_dysample_pretrain

Scale: 2x
Network type: realplksr_large dysample
Author: Philip Hofmann
License: CC-BY-4.0
Release: 21.05.2025
Purpose: 2x realplksr_large dysample pretrain model with l1&mssim loss only.
Training iterations: 100'000
Description: A 2x realplksr_large dysample pretrain model.

Visual Examples

Tensorboard Validation Graphs on BHI100

2xBHI_small_realplksr_dysample_pretrain

2xBHI_small_realplksr_dysample_pretrain

Scale: 2x
Network type: realplksr dysample
Author: Philip Hofmann
License: CC-BY-4.0
Release: 21.05.2025
Purpose: 2x realplksr dysample pretrain model with l1&mssim loss only.
Training iterations: 100'000
Description: A 2x realplksr dysample pretrain model.

Visual Examples

Tensorboard Validation Graphs on BHI100

4xBHI_small_hat-l

Scale: 4x
Network type: HAT-L
Author: Philip Hofmann
License: CC-BY-4.0
Purpose: 4x upscaling good quality (handles no degradations) photography
Pretrained Model: HAT-L_SRx4_ImageNet-pretrain.pth
Training iterations: 150'000
Description: 4x hat-l upscaling model for good quality input. This model does not handle any degradations. This model is rather soft, I tried to balance sharpness and faithfulness/non-artifacts. For a bit sharper output, but can generate a bit of artifacts, you can try the 4xBHI_small_hat-l_sharp version, also included in this release, which might still feel soft if you are used to sharper outputs. You can also try the appended 4xBHI_small_hat-l_fdl_150000.pth or 4xBHI_small_hat-l_157084.pth checkpoints.

Visual Examples:
slow.pics

2xBHI_small_span_pretrain

2xBHI_small_span_pretrain

Scale: 2x
Network type: span
Author: Philip Hofmann
License: CC-BY-4.0
Release date: 19.05.2025
Purpose: 2x span pretrain model with l1&mssim loss only.
Training iterations: 100'000
Description: A 2x span pretrain model.

Slowpic Example

https://slow.pics/s/qCWRujAA?image-fit=cover

Visual Examples

Tensorboard Validation Graphs BHI100

2xBHI_small_span_fast_pretrain

2xBHI_small_span_fast_pretrain

Scale: 2x
Network type: span_fast
Author: Philip Hofmann
License: CC-BY-4.0
Release: 19.05.2025
Purpose: 2x span_fast pretrain model with l1&mssim loss only.
Training iterations: 100'000
Description: A 2x span_fast pretrain model.

Network Option

musl added the span_fast option to neosr:

Changes on span_fast are based on this year's NTIRE efficient SR challenge:

• Remove last SPAB block. According to the XiaomiMM team, it decreases computational complexity without impacting performance.,
• Replace the first Conv3XC of the forward with a single large kernel conv, followed by a simple conv to refine it prior. The idea comes from mbga and Rochester teams, respectively.,
• Reduce channels to 32, according to XiaomiMM and mbga teams.,
• Disable biases, since according to the IESR team it only decreases 0.01db in PSNR, while accounting for 15% of runtime.,
• Initialize network with kaiming_normal_, giving faster/smoother convergence on early iters.,
• Use Mish instead of SiLU activation.

Inference speed

Zarxrax ran a inference speed test with this model on the GeForce RXT 2060
DML Span: 31.33
DML Span_fast: 41.93
DML Compact: 35.96
TRT Span: 69.44
TRT Span_fast: 143.74
TRT Compact: 60.63

Slowpic Example

https://slow.pics/s/CwqMiWs6?image-fit=cover

Visual Examples

Tensorboard Validation Graphs BHI100

2xBHI_small_esc_pretrain

2xBHI_small_esc_pretrain

Scale: 2x
Network type: esc
Author: Philip Hofmann
License: CC-BY-4.0
Release: 19.05.2025
Purpose: 2x esc pretrain model with l1&mssim loss only.
Training iterations: 100'000
Description: A 2x esc pretrain model.

Slowpic Example

https://slow.pics/s/ydR6X6lw?image-fit=cover

Visual Examples

Tensorboard Validation Metric Graphs on BHI100