Upload folder using huggingface_hub

README.md

@@ -0,0 +1,50 @@
+---
+license: apache-2.0
+library_name: traiNNer-redux
+tags:
+  - super-resolution
+  - image-restoration
+  - image-upscaling
+pipeline_tag: image-to-image
+---
+
+# NERVE — Neural Efficient Restoration for Various Edge devices
+
+Private backup of the NERVE 4x/2x release checkpoints, ONNX exports and
+training configs. NERVE is a compact norm-free conv super-resolution
+architecture (Conv-ReLU-Conv residual blocks, no attention) designed for
+mobile/edge deployment.
+
+## Architecture
+
+- `stem`: 5×5 conv opens the receptive field
+- body: `n_blocks` Conv-ReLU-Conv residual blocks (default `dim=48`, `n_blocks=16`)
+- head: PixelShuffle upsampler with **ICNR initialization** (checkerboard-free)
+- global residual + bicubic-input residual
+
+Params: ~688K (4x), ~672K (2x). Fully conv — dynamic-shape ONNX export
+(no fused/unfused pairs), NCNN/CoreML friendly.
+
+## Models
+
+| checkpoint | what | eval (Urban100) |
+|---|---|---|
+| `2x_NERVE_release_500k` | clean bicubic 2x | 31.647 PSNR / 0.9236 SSIM @500k |
+| `4x_NERVE_release_485k` | clean bicubic 4x | 24.9524 PSNR / 0.7473 SSIM @485k |
+| `4x_NERVE_OTF_fidelity_135k` | 4x OTF fidelity (Real-ESRGAN degradation) | 21.95 PSNR / 0.6435 SSIM @135k EMA |
+| `4x_NERVE_OTF_gan_35k` | 4x OTF GAN finetune | topiq 0.4065 / lpips 0.3174 @35k EMA |
+
+## ONNX
+
+`onnx/*.onnx` are dynamic-shape (1,3,H,W) fp32 exports, opset 20, verified
+against PyTorch on CPU with max abs diff < 5e-5 at 48×48 / 323×711 / 720×1280.
+
+## Usage
+
+```bash
+# train a stage
+python train.py -opt configs/4x_NERVE_OTF_fidelity_v2.yml --auto_resume
+```
+
+Checkpoint files are plain `net_g` EMA state dicts (no wrapper) — load with
+`traiNNer.archs.nerve_arch.nerve(scale=N, dim=48, n_blocks=16)`.

checkpoints/2x_NERVE_release_500k.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a783adb6f3300a49e7e79449a3a1a04d7351521b27dd9274429eee5df2893ef9
+size 2692592

checkpoints/4x_NERVE_OTF_fidelity_135k.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29ce83a227cb4325e2475a72acf885b81336011e6e704ed9bf5590de67420288
+size 2754800

checkpoints/4x_NERVE_OTF_gan_35k.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2d13329265e7f9b41542cab86ab2a05fb2bc31db3124390f1650e9573073a59
+size 2754800

checkpoints/4x_NERVE_release_485k.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed0e0d8e245abe9fe6100a8b36c4598d9a9f9e5dd774c876f9a7e95ccf5548ba
+size 2754800

configs/2x_NERVE_release.yml

@@ -0,0 +1,92 @@
+# yaml-language-server: $schema=https://raw.githubusercontent.com/the-database/traiNNer-redux/refs/heads/master/schemas/redux-config.schema.json
+#########################################################################################
+# NERVE 2x release training (clean bicubic fidelity)
+#
+# Canonical NERVE: norm-free, dim 48, n_blocks 16, ~672K params (2x head is
+# smaller than 4x). Trained on LUCID CC0 v2 HC-512 (license-clean) paired with
+# chainner CubicCatrom lr_x2. This is also the warm start for the 4x release:
+# the 4x run loads this model's BODY weights (strict_load_g: false resets the
+# scale-specific PixelShuffle head).
+#########################################################################################
+name: 2x_NERVE_release
+scale: 2
+use_amp: true
+amp_bf16: true
+use_channels_last: true
+fast_matmul: false
+use_compile: true
+compile_mode: default
+num_gpu: auto
+manual_seed: 1024
+
+datasets:
+  train:
+    name: LUCID Train
+    type: pairedimagedataset
+    dataroot_gt: [datasets/train/lucid_hc512/hr]
+    dataroot_lq: [datasets/train/lucid_hc512/lr_x2]
+    lq_size: 96
+    use_hflip: true
+    use_rot: true
+    num_worker_per_gpu: 8
+    batch_size_per_gpu: 8
+    accum_iter: 1
+  val:
+    name: Urban100 Val
+    type: pairedimagedataset
+    dataroot_gt: [datasets/val/urban100/hr]
+    dataroot_lq: [datasets/val/urban100/lr_x2]
+    filename_tmpl: "{}x2"
+
+network_g:
+  type: nerve
+  dim: 48
+  n_blocks: 16
+
+path:
+  param_key_g: ~
+  strict_load_g: true
+  resume_state: ~
+
+train:
+  ema_decay: 0.999
+  ema_power: 0.75
+  grad_clip: false
+  grad_clip_max_norm: 100
+  optim_g:
+    type: AdamW
+    lr: !!float 2e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  scheduler:
+    type: MultiStepLR
+    milestones: [200000, 350000, 450000]
+    gamma: 0.5
+  total_iter: 500000
+  warmup_iter: -1
+  losses:
+    - type: charbonnierloss
+      loss_weight: 1.0
+
+val:
+  val_enabled: true
+  val_freq: 5000
+  save_img: false  # disk-safe: metrics only; visuals on demand via inference scripts
+  tile_size: 0
+  tile_overlap: 8
+  metrics_enabled: true
+  metrics:
+    psnr:
+      type: calculate_psnr
+      crop_border: 2
+      test_y_channel: true
+    ssim:
+      type: calculate_ssim
+      crop_border: 2
+      test_y_channel: true
+
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  save_checkpoint_format: safetensors
+  use_tb_logger: true

configs/4x_NERVE_OTF_fidelity.yml

@@ -0,0 +1,158 @@
+# yaml-language-server: $schema=https://raw.githubusercontent.com/the-database/traiNNer-redux/refs/heads/master/schemas/redux-config.schema.json
+#########################################################################################
+# NERVE 4x OTF fidelity pretrain (Real-ESRGAN-style degradation)
+#
+# Degradation + losses mirror HEART's 4x_HEART_OTF_v2 so the NERVE OTF model is
+# directly comparable to the HEART OTF model. Network is canonical NERVE
+# (norm-free, dim 48, n_blocks 16). Charbonnier + MS-SSIM + small perceptual.
+#
+# Start from the clean release weights (see path.pretrain_network_g); if that
+# model is not available yet, set pretrain_network_g to ~ and train from
+# scratch (slower to converge, same recipe).
+#
+# lq_size 96 has no window-alignment meaning for NERVE (no attention) but is
+# kept to match HEART's degradation/eval pipeline.
+#########################################################################################
+name: 4x_NERVE_OTF_fidelity
+scale: 4
+use_amp: true
+amp_bf16: true
+use_channels_last: true
+fast_matmul: false
+use_compile: true
+compile_mode: default
+num_gpu: auto
+manual_seed: 1024
+
+high_order_degradation: true
+high_order_degradations_debug: false
+high_order_degradations_debug_limit: 100
+
+lq_usm: true
+lq_usm_radius_range: [1, 25]
+
+blur_prob: 0.6
+resize_prob: [0.2, 0.7, 0.1]
+resize_mode_list: ['bilinear', 'bicubic', 'nearest-exact', 'lanczos']
+resize_mode_prob: [0.25, 0.25, 0.25, 0.25]
+resize_range: [0.4, 1.5]
+gaussian_noise_prob: 0.4
+noise_range: [1, 15]
+poisson_scale_range: [0.05, 2]
+gray_noise_prob: 0.1
+jpeg_prob: 0.4
+jpeg_range: [40, 95]
+
+blur_prob2: 0.3
+resize_prob2: [0.3, 0.4, 0.3]
+resize_mode_list2: ['bilinear', 'bicubic', 'nearest-exact', 'lanczos']
+resize_mode_prob2: [0.25, 0.25, 0.25, 0.25]
+resize_range2: [0.6, 1.2]
+gaussian_noise_prob2: 0.3
+noise_range2: [1, 10]
+poisson_scale_range2: [0.05, 1.5]
+gray_noise_prob2: 0.05
+jpeg_prob2: 0.3
+jpeg_range2: [40, 95]
+
+resize_mode_list3: ['bilinear', 'bicubic', 'nearest-exact', 'lanczos']
+resize_mode_prob3: [0.25, 0.25, 0.25, 0.25]
+
+queue_size: 120
+
+datasets:
+  train:
+    name: LUCID OTF Train
+    type: realesrgandataset
+    dataroot_gt: [datasets/train/lucid_hc512/hr]
+
+    blur_kernel_size: 12
+    kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    kernel_range: [5, 17]
+    blur_sigma: [0.2, 2]
+    betag_range: [0.5, 4]
+    betap_range: [1, 2]
+
+    blur_kernel_size2: 12
+    kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    kernel_range2: [5, 17]
+    blur_sigma2: [0.2, 1]
+    betag_range2: [0.5, 4]
+    betap_range2: [1, 2]
+
+    final_kernel_range: [5, 17]
+
+    lq_size: 96
+    use_hflip: true
+    use_rot: true
+    num_worker_per_gpu: 8
+    batch_size_per_gpu: 8
+    accum_iter: 1
+
+  val:
+    name: Urban100 OTF Val
+    type: pairedimagedataset
+    dataroot_gt: [datasets/val/urban100/hr]
+    dataroot_lq: [datasets/val/urban100_otf_x4]
+
+network_g:
+  type: nerve
+  dim: 48
+  n_blocks: 16
+
+path:
+  pretrain_network_g: experiments/4x_NERVE_release/models/net_g_ema_latest.safetensors
+  param_key_g: ~
+  strict_load_g: true
+  resume_state: ~
+
+train:
+  ema_decay: 0.999
+  ema_power: 0.75
+  grad_clip: false
+  grad_clip_max_norm: 100
+  optim_g:
+    type: AdamW
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  scheduler:
+    type: MultiStepLR
+    milestones: [100000, 180000]
+    gamma: 0.5
+  total_iter: 200000
+  warmup_iter: -1
+  losses:
+    - type: charbonnierloss
+      loss_weight: 1.0
+    - type: perceptualloss
+      criterion: charbonnier
+      loss_weight: 0.01
+    - type: mssimloss
+      loss_weight: 0.5
+
+val:
+  val_enabled: true
+  val_freq: 5000
+  save_img: false  # disk-safe: metrics only; visuals on demand via inference scripts
+  tile_size: 0
+  tile_overlap: 8
+  metrics_enabled: true
+  metrics:
+    psnr:
+      type: calculate_psnr
+      crop_border: 4
+      test_y_channel: true
+    ssim:
+      type: calculate_ssim
+      crop_border: 4
+      test_y_channel: true
+
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  save_checkpoint_format: safetensors
+  use_tb_logger: true
+

configs/4x_NERVE_OTF_fidelity_v2.yml

@@ -0,0 +1,163 @@
+# yaml-language-server: $schema=https://raw.githubusercontent.com/the-database/traiNNer-redux/refs/heads/master/schemas/redux-config.schema.json
+#########################################################################################
+# NERVE 4x OTF fidelity pretrain v2 (Real-ESRGAN-style degradation)
+#
+# Changes vs 4x_NERVE_OTF_fidelity:
+#   * ICNR init on the PixelShuffle head at build + icnr_reinit_g scrubs the
+#     warm-started head (kills sub-pixel checkerboard learned by v1).
+#   * lq_phase_jitter: random sub-pixel shift of the final LQ per sample to
+#     break the PixelShuffle/degradation phase lock.
+#   * Final resize-back no longer uses nearest-exact (phase-locked + aliasing);
+#     only anti-aliased bilinear/bicubic/lanczos.
+#
+# Warm-started from the clean release weights. Same losses/recipe as v1 so the
+# checkerboard-vs-clean tradeoff can be compared 1:1.
+#########################################################################################
+name: 4x_NERVE_OTF_fidelity_v2
+scale: 4
+use_amp: true
+amp_bf16: true
+use_channels_last: true
+fast_matmul: false
+use_compile: true
+compile_mode: default
+num_gpu: auto
+manual_seed: 1024
+
+high_order_degradation: true
+high_order_degradations_debug: false
+high_order_degradations_debug_limit: 100
+
+lq_usm: true
+lq_usm_radius_range: [1, 25]
+
+lq_phase_jitter: true
+lq_phase_jitter_range: 0.5
+
+blur_prob: 0.6
+resize_prob: [0.2, 0.7, 0.1]
+resize_mode_list: ['bilinear', 'bicubic', 'nearest-exact', 'lanczos']
+resize_mode_prob: [0.25, 0.25, 0.25, 0.25]
+resize_range: [0.4, 1.5]
+gaussian_noise_prob: 0.4
+noise_range: [1, 15]
+poisson_scale_range: [0.05, 2]
+gray_noise_prob: 0.1
+jpeg_prob: 0.4
+jpeg_range: [40, 95]
+
+blur_prob2: 0.3
+resize_prob2: [0.3, 0.4, 0.3]
+resize_mode_list2: ['bilinear', 'bicubic', 'nearest-exact', 'lanczos']
+resize_mode_prob2: [0.25, 0.25, 0.25, 0.25]
+resize_range2: [0.6, 1.2]
+gaussian_noise_prob2: 0.3
+noise_range2: [1, 10]
+poisson_scale_range2: [0.05, 1.5]
+gray_noise_prob2: 0.05
+jpeg_prob2: 0.3
+jpeg_range2: [40, 95]
+
+# anti-aliased final resize-back only (no nearest-exact: phase-locked + aliasing)
+resize_mode_list3: ['bilinear', 'bicubic', 'lanczos']
+resize_mode_prob3: [0.3333, 0.3333, 0.3334]
+
+queue_size: 120
+
+datasets:
+  train:
+    name: LUCID OTF Train
+    type: realesrgandataset
+    dataroot_gt: [datasets/train/lucid_hc512/hr]
+
+    blur_kernel_size: 12
+    kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    kernel_range: [5, 17]
+    blur_sigma: [0.2, 2]
+    betag_range: [0.5, 4]
+    betap_range: [1, 2]
+
+    blur_kernel_size2: 12
+    kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    kernel_range2: [5, 17]
+    blur_sigma2: [0.2, 1]
+    betag_range2: [0.5, 4]
+    betap_range2: [1, 2]
+
+    final_kernel_range: [5, 17]
+
+    lq_size: 96
+    use_hflip: true
+    use_rot: true
+    num_worker_per_gpu: 8
+    batch_size_per_gpu: 8
+    accum_iter: 1
+
+  val:
+    name: Urban100 OTF Val
+    type: pairedimagedataset
+    dataroot_gt: [datasets/val/urban100/hr]
+    dataroot_lq: [datasets/val/urban100_otf_x4]
+
+network_g:
+  type: nerve
+  dim: 48
+  n_blocks: 16
+  icnr: true
+
+path:
+  pretrain_network_g: experiments/4x_NERVE_release/models/net_g_ema_latest.safetensors
+  param_key_g: ~
+  strict_load_g: true
+  icnr_reinit_g: true
+  resume_state: ~
+
+train:
+  ema_decay: 0.999
+  ema_power: 0.75
+  grad_clip: false
+  grad_clip_max_norm: 100
+  optim_g:
+    type: AdamW
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  scheduler:
+    type: MultiStepLR
+    milestones: [100000, 180000]
+    gamma: 0.5
+  total_iter: 200000
+  warmup_iter: -1
+  losses:
+    - type: charbonnierloss
+      loss_weight: 1.0
+    - type: perceptualloss
+      criterion: charbonnier
+      loss_weight: 0.01
+    - type: mssimloss
+      loss_weight: 0.5
+
+val:
+  val_enabled: true
+  val_freq: 5000
+  save_img: false  # disk-safe: metrics only; visuals on demand via inference scripts
+  tile_size: 0
+  tile_overlap: 8
+  metrics_enabled: true
+  metrics:
+    psnr:
+      type: calculate_psnr
+      crop_border: 4
+      test_y_channel: true
+    ssim:
+      type: calculate_ssim
+      crop_border: 4
+      test_y_channel: true
+
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  save_checkpoint_format: safetensors
+  use_tb_logger: true

configs/4x_NERVE_OTF_gan.yml

@@ -0,0 +1,166 @@
+# yaml-language-server: $schema=https://raw.githubusercontent.com/the-database/traiNNer-redux/refs/heads/master/schemas/redux-config.schema.json
+#########################################################################################
+# NERVE 4x OTF GAN finetune (mirrors 4x_HEART_OTF_gan_lq96)
+#
+# Same degradation + losses as the HEART OTF GAN finetune; the generator is
+# canonical NERVE (norm-free, dim 48, n_blocks 16, ~688K params). Discriminator
+# is dunet like HEART. No attention windows, so the lq-96 window-alignment fix
+# that HEART needed does not apply here; lq 96 is kept to match the degradation
+# pipeline and evaluation.
+#
+# Two-stage recipe (like HEART):
+#   1. 4x_NERVE_OTF_fidelity.yml  (OTF fidelity pretrain)
+#   2. this file (GAN finetune on top of it)
+#########################################################################################
+name: 4x_NERVE_OTF_gan
+scale: 4
+use_amp: true
+amp_bf16: true
+use_channels_last: true
+fast_matmul: false
+use_compile: false  # keep default: inductor errors on discriminator step
+compile_mode: default
+num_gpu: auto
+manual_seed: 1024
+
+high_order_degradation: true
+high_order_degradations_debug: false
+high_order_degradations_debug_limit: 100
+
+lq_usm: true
+lq_usm_radius_range: [1, 25]
+
+blur_prob: 0.6
+resize_prob: [0.2, 0.7, 0.1]
+resize_mode_list: ['bilinear', 'bicubic', 'nearest-exact', 'lanczos']
+resize_mode_prob: [0.25, 0.25, 0.25, 0.25]
+resize_range: [0.4, 1.5]
+gaussian_noise_prob: 0.4
+noise_range: [1, 15]
+poisson_scale_range: [0.05, 2]
+gray_noise_prob: 0.1
+jpeg_prob: 0.4
+jpeg_range: [40, 95]
+
+blur_prob2: 0.3
+resize_prob2: [0.3, 0.4, 0.3]
+resize_mode_list2: ['bilinear', 'bicubic', 'nearest-exact', 'lanczos']
+resize_mode_prob2: [0.25, 0.25, 0.25, 0.25]
+resize_range2: [0.6, 1.2]
+gaussian_noise_prob2: 0.3
+noise_range2: [1, 10]
+poisson_scale_range2: [0.05, 1.5]
+gray_noise_prob2: 0.05
+jpeg_prob2: 0.3
+jpeg_range2: [40, 95]
+
+resize_mode_list3: ['bilinear', 'bicubic', 'nearest-exact', 'lanczos']
+resize_mode_prob3: [0.25, 0.25, 0.25, 0.25]
+
+queue_size: 120
+
+datasets:
+  train:
+    name: LUCID OTF Train
+    type: realesrgandataset
+    dataroot_gt: [datasets/train/lucid_hc512/hr]
+
+    blur_kernel_size: 12
+    kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    kernel_range: [5, 17]
+    blur_sigma: [0.2, 2]
+    betag_range: [0.5, 4]
+    betap_range: [1, 2]
+
+    blur_kernel_size2: 12
+    kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    kernel_range2: [5, 17]
+    blur_sigma2: [0.2, 1]
+    betag_range2: [0.5, 4]
+    betap_range2: [1, 2]
+
+    final_kernel_range: [5, 17]
+
+    lq_size: 96
+    use_hflip: true
+    use_rot: true
+    num_worker_per_gpu: 8
+    batch_size_per_gpu: 4   # NERVE generator is small; batch 4 at lq96 fits easily
+    accum_iter: 1
+
+  val:
+    name: Urban100 OTF Val
+    type: pairedimagedataset
+    dataroot_gt: [datasets/val/urban100/hr]
+    dataroot_lq: [datasets/val/urban100_otf_x4]
+
+network_g:
+  type: nerve
+  dim: 48
+  n_blocks: 16
+
+network_d:
+  type: dunet
+
+path:
+  pretrain_network_g: experiments/4x_NERVE_OTF_fidelity/models/net_g_ema_135000.safetensors
+  param_key_g: ~
+  strict_load_g: true
+  resume_state: ~
+
+train:
+  ema_decay: 0.999
+  ema_power: 0.75
+  grad_clip: false
+  grad_clip_max_norm: 100
+  optim_g:
+    type: AdamW
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  optim_d:
+    type: AdamW
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  scheduler:
+    type: MultiStepLR
+    milestones: [75000, 120000]
+    gamma: 0.5
+  total_iter: 150000
+  warmup_iter: -1
+  losses:
+    - type: mssimloss
+      loss_weight: 0.5
+    - type: perceptualloss
+      criterion: charbonnier
+      loss_weight: 0.01
+    - type: hsluvloss
+      criterion: charbonnier
+      loss_weight: 1.0
+    - type: ganloss
+      gan_type: vanilla
+      loss_weight: 0.1
+
+val:
+  val_enabled: true
+  val_freq: 5000
+  save_img: false  # disk-safe: metrics only; visuals on demand via inference scripts
+  tile_size: 0
+  tile_overlap: 8
+  metrics_enabled: true
+  metrics:
+    topiq:
+      type: calculate_topiq
+    lpips:
+      type: calculate_lpips
+      better: lower
+
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  save_checkpoint_format: safetensors
+  use_tb_logger: true
+

configs/4x_NERVE_OTF_gan_v2.yml

@@ -0,0 +1,169 @@
+# yaml-language-server: $schema=https://raw.githubusercontent.com/the-database/traiNNer-redux/refs/heads/master/schemas/redux-config.schema.json
+#########################################################################################
+# NERVE 4x OTF GAN finetune v2 (mirrors 4x_HEART_OTF_gan_lq96)
+#
+# Changes vs 4x_NERVE_OTF_gan:
+#   * Degradation fixes carried from 4x_NERVE_OTF_fidelity_v2:
+#       - lq_phase_jitter: random sub-pixel LQ shift per sample
+#       - final resize-back is anti-aliased only (no nearest-exact)
+#   * Warm-starts from the fixed fidelity v2 model (which has an ICNR-clean
+#     PixelShuffle head), so the head is kept, not re-scrubbed.
+#
+# Same losses/recipe as v1 for 1:1 comparison.
+#########################################################################################
+name: 4x_NERVE_OTF_gan_v2
+scale: 4
+use_amp: true
+amp_bf16: true
+use_channels_last: true
+fast_matmul: false
+use_compile: false  # keep default: inductor errors on discriminator step
+compile_mode: default
+num_gpu: auto
+manual_seed: 1024
+
+high_order_degradation: true
+high_order_degradations_debug: false
+high_order_degradations_debug_limit: 100
+
+lq_usm: true
+lq_usm_radius_range: [1, 25]
+
+lq_phase_jitter: true
+lq_phase_jitter_range: 0.5
+
+blur_prob: 0.6
+resize_prob: [0.2, 0.7, 0.1]
+resize_mode_list: ['bilinear', 'bicubic', 'nearest-exact', 'lanczos']
+resize_mode_prob: [0.25, 0.25, 0.25, 0.25]
+resize_range: [0.4, 1.5]
+gaussian_noise_prob: 0.4
+noise_range: [1, 15]
+poisson_scale_range: [0.05, 2]
+gray_noise_prob: 0.1
+jpeg_prob: 0.4
+jpeg_range: [40, 95]
+
+blur_prob2: 0.3
+resize_prob2: [0.3, 0.4, 0.3]
+resize_mode_list2: ['bilinear', 'bicubic', 'nearest-exact', 'lanczos']
+resize_mode_prob2: [0.25, 0.25, 0.25, 0.25]
+resize_range2: [0.6, 1.2]
+gaussian_noise_prob2: 0.3
+noise_range2: [1, 10]
+poisson_scale_range2: [0.05, 1.5]
+gray_noise_prob2: 0.05
+jpeg_prob2: 0.3
+jpeg_range2: [40, 95]
+
+# anti-aliased final resize-back only (no nearest-exact: phase-locked + aliasing)
+resize_mode_list3: ['bilinear', 'bicubic', 'lanczos']
+resize_mode_prob3: [0.3333, 0.3333, 0.3334]
+
+queue_size: 120
+
+datasets:
+  train:
+    name: LUCID OTF Train
+    type: realesrgandataset
+    dataroot_gt: [datasets/train/lucid_hc512/hr]
+
+    blur_kernel_size: 12
+    kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    kernel_range: [5, 17]
+    blur_sigma: [0.2, 2]
+    betag_range: [0.5, 4]
+    betap_range: [1, 2]
+
+    blur_kernel_size2: 12
+    kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    kernel_range2: [5, 17]
+    blur_sigma2: [0.2, 1]
+    betag_range2: [0.5, 4]
+    betap_range2: [1, 2]
+
+    final_kernel_range: [5, 17]
+
+    lq_size: 96
+    use_hflip: true
+    use_rot: true
+    num_worker_per_gpu: 8
+    batch_size_per_gpu: 4   # NERVE generator is small; batch 4 at lq96 fits easily
+    accum_iter: 1
+
+  val:
+    name: Urban100 OTF Val
+    type: pairedimagedataset
+    dataroot_gt: [datasets/val/urban100/hr]
+    dataroot_lq: [datasets/val/urban100_otf_x4]
+
+network_g:
+  type: nerve
+  dim: 48
+  n_blocks: 16
+  icnr: true
+
+network_d:
+  type: dunet
+
+path:
+  pretrain_network_g: experiments/4x_NERVE_OTF_fidelity_v2/models/net_g_ema_latest.safetensors
+  param_key_g: ~
+  strict_load_g: true
+  resume_state: ~
+
+train:
+  ema_decay: 0.999
+  ema_power: 0.75
+  grad_clip: false
+  grad_clip_max_norm: 100
+  optim_g:
+    type: AdamW
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  optim_d:
+    type: AdamW
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  scheduler:
+    type: MultiStepLR
+    milestones: [75000, 120000]
+    gamma: 0.5
+  total_iter: 150000
+  warmup_iter: -1
+  losses:
+    - type: mssimloss
+      loss_weight: 0.5
+    - type: perceptualloss
+      criterion: charbonnier
+      loss_weight: 0.01
+    - type: hsluvloss
+      criterion: charbonnier
+      loss_weight: 1.0
+    - type: ganloss
+      gan_type: vanilla
+      loss_weight: 0.1
+
+val:
+  val_enabled: true
+  val_freq: 5000
+  save_img: false  # disk-safe: metrics only; visuals on demand via inference scripts
+  tile_size: 0
+  tile_overlap: 8
+  metrics_enabled: true
+  metrics:
+    topiq:
+      type: calculate_topiq
+    lpips:
+      type: calculate_lpips
+      better: lower
+
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  save_checkpoint_format: safetensors
+  use_tb_logger: true

configs/4x_NERVE_release.yml

@@ -0,0 +1,101 @@
+# yaml-language-server: $schema=https://raw.githubusercontent.com/the-database/traiNNer-redux/refs/heads/master/schemas/redux-config.schema.json
+#########################################################################################
+# NERVE 4x release training (clean bicubic fidelity)
+#
+# Canonical NERVE: norm-free, dim 48, n_blocks 16, ~688K params. Pure conv, no
+# attention, no normalization. Trained on LUCID CC0 v2 HC-512 (license-clean)
+# paired with chainner CubicCatrom lr_x4. This is the fidelity tier; the OTF
+# fidelity + GAN configs fine-tune on top of this family of weights.
+#
+# Warm start (EDSR-style multi-scale init) from the 2x release body:
+# train 2x_NERVE_release.yml first, then this file loads its body weights;
+# strict_load_g: false resets the scale-specific PixelShuffle head.
+#
+# No window-alignment constraints (NERVE has no attention windows), so crop
+# size and batch are free to pick for throughput; lq 96 matches the HEART
+# pipeline for comparability.
+#########################################################################################
+name: 4x_NERVE_release
+scale: 4
+use_amp: true
+amp_bf16: true
+use_channels_last: true
+fast_matmul: false
+use_compile: true
+compile_mode: default  # conv-only net: 'default' is stable and export-safe
+num_gpu: auto
+manual_seed: 1024
+
+datasets:
+  train:
+    name: LUCID Train
+    type: pairedimagedataset
+    dataroot_gt: [datasets/train/lucid_hc512/hr]
+    dataroot_lq: [datasets/train/lucid_hc512/lr_x4]
+    lq_size: 96
+    use_hflip: true
+    use_rot: true
+    num_worker_per_gpu: 8
+    batch_size_per_gpu: 8
+    accum_iter: 1
+  val:
+    name: Urban100 Val
+    type: pairedimagedataset
+    dataroot_gt: [datasets/val/urban100/hr]
+    dataroot_lq: [datasets/val/urban100/lr_x4]
+    filename_tmpl: "{}x4"
+
+network_g:
+  type: nerve
+  dim: 48
+  n_blocks: 16
+
+path:
+  pretrain_network_g: experiments/2x_NERVE_release/models/net_g_ema_latest.safetensors
+  param_key_g: ~
+  strict_load_g: false  # loads the body; resets the scale-specific head
+  resume_state: ~
+
+train:
+  ema_decay: 0.999
+  ema_power: 0.75
+  grad_clip: false
+  grad_clip_max_norm: 100
+  optim_g:
+    type: AdamW
+    lr: !!float 2e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  scheduler:
+    type: MultiStepLR
+    milestones: [200000, 350000, 450000]
+    gamma: 0.5
+  total_iter: 500000
+  warmup_iter: -1
+  losses:
+    - type: charbonnierloss
+      loss_weight: 1.0
+
+val:
+  val_enabled: true
+  val_freq: 5000
+  save_img: false  # disk-safe: metrics only; visuals on demand via inference scripts
+  tile_size: 0
+  tile_overlap: 8
+  metrics_enabled: true
+  metrics:
+    psnr:
+      type: calculate_psnr
+      crop_border: 4
+      test_y_channel: true
+    ssim:
+      type: calculate_ssim
+      crop_border: 4
+      test_y_channel: true
+
+logger:
+  print_freq: 100
+  save_checkpoint_freq: 5000
+  save_checkpoint_format: safetensors
+  use_tb_logger: true
+

onnx/2x_NERVE_1x3xHxW_fp32_op20.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54557cbe6b82a31dbbc2a635d06afca67834558c6a01e5814a24deef1747c849
+size 106686

onnx/4x_NERVE_1x3xHxW_fp32_op20.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37338cdbd5916f879cbf0a7a731a785f5ca83bcaf28aef31e0394984ce3dfdbc
+size 106687