NERVE 1.80M release: models, optimized ONNX, ncnn, configs, arch, README

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -1,56 +1,167 @@
 ---
 license: apache-2.0
-library_name: traiNNer-redux
+language:
+- en
 tags:
-  - super-resolution
-  - image-restoration
-  - image-upscaling
+- super-resolution
+- image-restoration
+- pytorch
+- onnx
+- ncnn
+- spandrel
+- chaiNNer
 pipeline_tag: image-to-image
 ---
 
-# NERVE — Neural Efficient Restoration for Various Edge devices
+# NERVE
 
-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.
+**Norm-free Efficient Restoration for Various Edge devices**
 
-## Architecture
+A lightweight super-resolution architecture that is simple to train, fast to run,
+and works everywhere: **traiNNer-redux**, **chaiNNer**, **ONNX Runtime**,
+**ncnn**, **CoreML**.
+
+> Short version: copy `nerve_arch.py` into traiNNer-redux, train with the sample
+> configs, use the released `.safetensors` / `.onnx` / `.ncnn` models anywhere.
+> One architecture, one size, no fused/unfused pairs. Details below.
+
+Part of the **BODY** suite by **Philip Hofmann** — networks built to be
+*real-world usable first*, not benchmark-chasing. See
+[Relation to HEART](#relation-to-heart) for which one to pick.
+
+---
+
+## Quickstart
 
-- `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
+### As a user (just want to upscale / restore images)
 
-Params: ~688K (4x), ~672K (2x). Fully conv — dynamic-shape ONNX export
-(no fused/unfused pairs), NCNN/CoreML friendly.
+Pick a model and run it:
+
+| I want to... | Use |
+|---|---|
+| upscale a clean image 4x | `models/4x_NERVE_release.safetensors` |
+| upscale a clean image 2x | `models/2x_NERVE_release.safetensors` |
+| restore a blurry / noisy / JPEG-y image 4x | `models/4x_NERVE_OTF_gan.safetensors` |
+| run in the browser / on a phone / in a game engine | the `onnx/` or `ncnn/` files |
+
+- **chaiNNer / spandrel**: drop the `.safetensors` in — it auto-detects NERVE.
+- **ONNX Runtime / TensorRT / DirectML**: use `onnx/*.onnx` (dynamic H/W).
+- **ncnn**: use `ncnn/*.ncnn.param` + `*.ncnn.bin` (fp16).
+- Input: RGB, any size. Output: 2x or 4x.
+
+### As a trainer (want to train your own)
+
+```bash
+# 1. Put the architecture in traiNNer-redux
+cp nerve_arch.py <traiNNer-redux>/traiNNer/archs/
+
+# 2. Train (copy a config from configs/ and point the paths at your data)
+python train.py -opt configs/4x_NERVE_release.yml --auto_resume
+
+# 3. Export to dynamic ONNX
+python scripts/export_dynamic.py --dim 64 --n-blocks 24 \
+    --checkpoint <run>/models/net_g_ema_latest.safetensors --scale 4
+```
+
+That's the whole workflow. The net is a single file, the config is a single
+file, and the checkpoint is a single file — no fusion step, no paired
+fused/unfused weights.
+
+**Training tips** (all NERVE models follow these):
+- bf16 AMP is fine; no gradient clipping needed (norm-free net).
+- `--auto_resume` is recommended for long runs.
+- For real-world input, continue from the clean release into the OTF fidelity
+  then the OTF GAN config — that's the chain the released models use.
+
+---
 
 ## Models
 
-| checkpoint | what | eval (Urban100) |
+| File | Role | Notes |
 |---|---|---|
-| `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 |
+| `models/2x_NERVE_release.safetensors` | **2x official pretrain** | clean bicubic, 31.84 dB / 0.9255 SSIM (Urban100) |
+| `models/4x_NERVE_release.safetensors` | **4x official pretrain** | clean bicubic, 25.09 dB / 0.7531 SSIM (Urban100) |
+| `models/4x_NERVE_OTF_fidelity.safetensors` | 4x OTF fidelity | the pretrain for OTF GAN finetunes |
+| `models/4x_NERVE_OTF_gan.safetensors` | 4x OTF GAN | **sharpest real-world output** |
+| `onnx/4x_NERVE_1x3xHxW_fp32_op20.onnx` | 4x release, dynamic ONNX | onnxslim-optimized, fp32 |
+| `onnx/2x_NERVE_1x3xHxW_fp32_op20.onnx` | 2x release, dynamic ONNX | onnxslim-optimized, fp32 |
+| `ncnn/4x_NERVE_fp16.ncnn.param` + `.bin` | 4x release, ncnn | fp16 weights |
+| `ncnn/2x_NERVE_fp16.ncnn.param` + `.bin` | 2x release, ncnn | fp16 weights |
 
-## ONNX / NCNN
+`models/*.safetensors` are plain EMA state dicts (no wrapper) — load them with
+`nerve_arch.nerve(scale=N)`.
 
-`onnx/*.onnx` are self-contained, dynamic-shape (1,3,H,W) fp32 exports, opset 20,
-verified against PyTorch on CPU (max abs diff < 5e-5 at 48×48 / 323×711 /
-720×1280; < 2e-6 on 256×256).
+The release models (2x/4x) are final. The OTF fidelity/GAN models are added
+once their training finishes; the release models are ready to use and to
+warm-start from today.
 
-`onnx/*_fp16.ncnn.param/.bin` are ncnn conversions (pnnx, optlevel 2, fp16
-weights, fixed 256×256 input for mobile deployment) verified against PyTorch
-(max abs diff ~4e-3 on unclamped output). `onnx/*_fp32.ncnn.*` are the fp32
-variants (~3e-6). `onnx/pnnx/` holds the pnnx intermediates for reproducibility.
+**Pretrain chain** (how the models relate):
 
-## Usage
+```
+2x_NERVE_release ──→ 4x_NERVE_release ──→ OTF fidelity ──→ OTF GAN
+```
 
-```bash
-# train a stage
-python train.py -opt configs/4x_NERVE_OTF_fidelity_v2.yml --auto_resume
+The release models are the shared starting point for anyone training NERVE:
+warm-start from them (2x → 4x, or release → OTF fidelity → OTF GAN) to skip the
+expensive early training.
+
+---
+
+## Architecture
+
+```
+5x5 conv stem → 24 × (Conv-ReLU-Conv + residual) → PixelShuffle → + bicubic(x)
 ```
 
-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)`.
+- **~1.8M params**, pure convolution — **no attention, no normalization, no
+  gating, no positional encoding, no reparameterization.**
+- A **bicubic input residual** means the net only learns what bicubic upscaling
+  misses — that's why it stays small and trains fast.
+- **ICNR-initialized** PixelShuffle head (checkerboard-free by construction).
+- The op graph is trivial: `Conv, Add, ReLU, DepthToSpace, Resize`. That's the
+  entire reason it exports to dynamic ONNX and converts to ncnn/CoreML without
+  custom ops or fused/unfused pairs.
+
+**Design philosophy.** NERVE trades the last fraction of a dB of PSNR for
+everything that actually matters in production:
+- **One variant.** No S/M/L, no fused vs unfused, no attention-window caveats.
+- **Clean to train.** Norm-free, bf16-stable, a single loss for the base stage.
+- **Clean to deploy.** Standard ops only; what you train is what you export.
+- **Maintainable.** One ~90-line arch file a maintainer can read in full.
+
+It does **not** chase benchmark leaderboards. It is the model you can hand to
+someone and they can train, export, and ship it the same afternoon.
+
+## Relation to HEART
+
+NERVE and HEART are the two members of the **BODY** suite (by Philip Hofmann):
+same philosophy (real-world usability > benchmark scores), different tiers.
+
+| | **HEART** | **NERVE** |
+|---|---|---|
+| size | ~16.7M params | ~1.8M params |
+| design | attention (FlashAttention-friendly) | pure convolution |
+| use when | you want the **best quality** and have the compute (desktop, server, high-end device) | you want **speed, small size, and dead-simple deployment** (mobile, edge, web, game engines) |
+| cost | heavier to run | a fraction of the cost |
+
+Rule of thumb: **HEART for maximum quality, NERVE for maximum practicality.**
+Both export to dynamic ONNX; NERVE additionally converts to ncnn cleanly
+because it has no attention ops.
+
+---
+
+## Repository layout
+
+```
+nerve_arch.py                    the architecture (single file)
+configs/                         example training configs (release, OTF fidelity, OTF GAN, 2x)
+models/                          pretrained checkpoints (see table above)
+onnx/                            dynamic fp32 ONNX (opset 20, onnxslim-optimized)
+ncnn/                            ncnn fp16 param/bin (fixed 256x256 input)
+scripts/export_dynamic.py        checkpoint → dynamic ONNX exporter (+ ORT verify)
+LICENSE                          Apache-2.0
+```
+
+## License
+
+Apache-2.0 (this repository's code and pretrained weights).

configs/2x_NERVE_release.yml

@@ -23,7 +23,7 @@ datasets:
   train:
     name: LUCID Train
     type: pairedimagedataset
-    dataroot_gt: [datasets/train/lucid_hc512/hr]
+    dataroot_gt: [datasets/train/hr]
     dataroot_lq: [datasets/train/lucid_hc512/lr_x2]
     lq_size: 96
     use_hflip: true
@@ -34,16 +34,17 @@ datasets:
   val:
     name: Urban100 Val
     type: pairedimagedataset
-    dataroot_gt: [datasets/val/urban100/hr]
+    dataroot_gt: [datasets/val/hr]
     dataroot_lq: [datasets/val/urban100/lr_x2]
     filename_tmpl: "{}x2"
 
 network_g:
   type: nerve
-  dim: 48
-  n_blocks: 16
+  dim: 64
+  n_blocks: 24
 
 path:
+  pretrain_network_g: ~
   param_key_g: ~
   strict_load_g: true
   resume_state: ~
@@ -60,9 +61,9 @@ train:
     betas: [0.9, 0.99]
   scheduler:
     type: MultiStepLR
-    milestones: [200000, 350000, 450000]
+    milestones: [100000, 160000]
     gamma: 0.5
-  total_iter: 500000
+  total_iter: 200000
   warmup_iter: -1
   losses:
     - type: charbonnierloss

configs/4x_NERVE_OTF_fidelity.yml

@@ -64,7 +64,7 @@ datasets:
   train:
     name: LUCID OTF Train
     type: realesrgandataset
-    dataroot_gt: [datasets/train/lucid_hc512/hr]
+    dataroot_gt: [datasets/train/hr]
 
     blur_kernel_size: 12
     kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
@@ -94,13 +94,13 @@ datasets:
   val:
     name: Urban100 OTF Val
     type: pairedimagedataset
-    dataroot_gt: [datasets/val/urban100/hr]
+    dataroot_gt: [datasets/val/hr]
     dataroot_lq: [datasets/val/urban100_otf_x4]
 
 network_g:
   type: nerve
-  dim: 48
-  n_blocks: 16
+  dim: 64
+  n_blocks: 24
 
 path:
   pretrain_network_g: experiments/4x_NERVE_release/models/net_g_ema_latest.safetensors

configs/4x_NERVE_OTF_gan.yml

@@ -1,16 +1,19 @@
 # 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)
+# NERVE 4x — OTF GAN (the showcase recipe, real-world degraded input)
 #
-# 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.
+# Real-ESRGAN-style on-the-fly degradation (blur kernels, noise, JPEG) with a
+# U-Net discriminator. This is the recipe the released OTF GAN model uses.
 #
-# Two-stage recipe (like HEART):
-#   1. 4x_NERVE_OTF_fidelity.yml  (OTF fidelity pretrain)
-#   2. this file (GAN finetune on top of it)
+# Two-stage chain (train in this order):
+#   1. 4x_NERVE_release.yml        clean pretrain
+#   2. 4x_NERVE_OTF_fidelity.yml   OTF fidelity (denoise/deblur/dejpeg)
+#   3. this file                   GAN finetune on top of the fidelity model
+# So set path.pretrain_network_g to your fidelity checkpoint.
+#
+# NOTE: the LR keeps `nearest-exact` in the final resize. Do NOT anti-alias the
+# LR — training on aliased (jagged) LR is what teaches the model to de-alias.
+# Also do not enable lq_phase_jitter for the GAN stage (it destabilizes it).
 #########################################################################################
 name: 4x_NERVE_OTF_gan
 scale: 4
@@ -18,14 +21,12 @@ use_amp: true
 amp_bf16: true
 use_channels_last: true
 fast_matmul: false
-use_compile: false  # keep default: inductor errors on discriminator step
+use_compile: false  # keep default: inductor errors on the 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]
@@ -61,9 +62,9 @@ queue_size: 120
 
 datasets:
   train:
-    name: LUCID OTF Train
+    name: Train
     type: realesrgandataset
-    dataroot_gt: [datasets/train/lucid_hc512/hr]
+    dataroot_gt: [datasets/train/hr]
 
     blur_kernel_size: 12
     kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
@@ -86,26 +87,23 @@ datasets:
     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
+    num_worker_per_gpu: 4
     accum_iter: 1
 
   val:
-    name: Urban100 OTF Val
+    name: Val
     type: pairedimagedataset
-    dataroot_gt: [datasets/val/urban100/hr]
-    dataroot_lq: [datasets/val/urban100_otf_x4]
+    dataroot_gt: [datasets/val/hr]
+    dataroot_lq: [datasets/val/lr_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
+  pretrain_network_g: experiments/4x_NERVE_OTF_fidelity/models/net_g_ema_latest.safetensors
   param_key_g: ~
   strict_load_g: true
   resume_state: ~
@@ -147,7 +145,7 @@ train:
 val:
   val_enabled: true
   val_freq: 5000
-  save_img: false  # disk-safe: metrics only; visuals on demand via inference scripts
+  save_img: false
   tile_size: 0
   tile_overlap: 8
   metrics_enabled: true
@@ -163,4 +161,3 @@ logger:
   save_checkpoint_freq: 5000
   save_checkpoint_format: safetensors
   use_tb_logger: true
-

configs/4x_NERVE_release.yml

@@ -1,19 +1,19 @@
 # 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)
+# NERVE 4x — clean release (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.
+# This is the base of the NERVE chain. Train it on clean HR/LR pairs
+# (HR x4-downsampled with any cubic filter), then optionally continue with the
+# OTF fidelity and OTF GAN configs for real-world degraded input.
 #
-# 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.
+# PATH: put nerve_arch.py in traiNNer-redux/traiNNer/archs/ first
+#       (cp nerve_arch.py <traiNNer-redux>/traiNNer/archs/).
+# DATA: dataroot_gt = HR images, dataroot_lq = matching x4-downscaled LRs.
+#       Point them at whatever folders you use, e.g. datasets/train/hr.
+# PRETRAIN: leave as ~ to train from scratch, or warm-start from the released
+#       2x body with:
+#         pretrain_network_g: 2x_NERVE_release.safetensors
+#         strict_load_g: false        # resets the scale-specific head
 #########################################################################################
 name: 4x_NERVE_release
 scale: 4
@@ -22,16 +22,16 @@ 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
+compile_mode: default
 num_gpu: auto
 manual_seed: 1024
 
 datasets:
   train:
-    name: LUCID Train
+    name: Train
     type: pairedimagedataset
-    dataroot_gt: [datasets/train/lucid_hc512/hr]
-    dataroot_lq: [datasets/train/lucid_hc512/lr_x4]
+    dataroot_gt: [datasets/train/hr]
+    dataroot_lq: [datasets/train/lr_x4]
     lq_size: 96
     use_hflip: true
     use_rot: true
@@ -39,21 +39,18 @@ datasets:
     batch_size_per_gpu: 8
     accum_iter: 1
   val:
-    name: Urban100 Val
+    name: Val
     type: pairedimagedataset
-    dataroot_gt: [datasets/val/urban100/hr]
-    dataroot_lq: [datasets/val/urban100/lr_x4]
-    filename_tmpl: "{}x4"
+    dataroot_gt: [datasets/val/hr]
+    dataroot_lq: [datasets/val/lr_x4]
 
 network_g:
   type: nerve
-  dim: 48
-  n_blocks: 16
 
 path:
-  pretrain_network_g: experiments/2x_NERVE_release/models/net_g_ema_latest.safetensors
+  pretrain_network_g: ~
   param_key_g: ~
-  strict_load_g: false  # loads the body; resets the scale-specific head
+  strict_load_g: true
   resume_state: ~
 
 train:
@@ -79,7 +76,7 @@ train:
 val:
   val_enabled: true
   val_freq: 5000
-  save_img: false  # disk-safe: metrics only; visuals on demand via inference scripts
+  save_img: false
   tile_size: 0
   tile_overlap: 8
   metrics_enabled: true
@@ -98,4 +95,3 @@ logger:
   save_checkpoint_freq: 5000
   save_checkpoint_format: safetensors
   use_tb_logger: true
-

models/2x_NERVE_release.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:241dc6a6cb2e3ef8bd66493c36bb6a900bf4ca97491d0d1462d42ecb4d989693
+size 7129504

models/4x_NERVE_release.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07bee958c168b561ec8d31ffe319ba6aa8ad3b67726633994e3d65687708fb98
+size 7212448

ncnn/2x_NERVE_fp16.ncnn.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78aed5af85a2cd140ce304442e87d60172135ee07a941a2aa0f7168624a1768d
+size 3562568

ncnn/2x_NERVE_fp16.ncnn.param

@@ -0,0 +1,106 @@
+7767517
+104 130
+Input                    in0                      0 1 in0
+Split                    splitncnn_0              1 2 in0 1 2
+Convolution              conv_26                  1 1 2 3 0=64 1=5 11=5 12=1 13=1 14=2 2=1 3=1 4=2 5=0 6=4800
+Split                    splitncnn_1              1 3 3 4 5 6
+Interp                   interpolate_0            1 1 1 7 0=3 1=2.0 2=2.0 6=0
+Convolution              convrelu_0               1 1 6 8 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_28                  1 1 8 9 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_0                    2 1 4 9 10 0=0
+Split                    splitncnn_2              1 2 10 11 12
+Convolution              convrelu_1               1 1 12 13 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_30                  1 1 13 14 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_1                    2 1 11 14 15 0=0
+Split                    splitncnn_3              1 2 15 16 17
+Convolution              convrelu_2               1 1 17 18 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_32                  1 1 18 19 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_2                    2 1 16 19 20 0=0
+Split                    splitncnn_4              1 2 20 21 22
+Convolution              convrelu_3               1 1 22 23 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_34                  1 1 23 24 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_3                    2 1 21 24 25 0=0
+Split                    splitncnn_5              1 2 25 26 27
+Convolution              convrelu_4               1 1 27 28 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_36                  1 1 28 29 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_4                    2 1 26 29 30 0=0
+Split                    splitncnn_6              1 2 30 31 32
+Convolution              convrelu_5               1 1 32 33 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_38                  1 1 33 34 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_5                    2 1 31 34 35 0=0
+Split                    splitncnn_7              1 2 35 36 37
+Convolution              convrelu_6               1 1 37 38 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_40                  1 1 38 39 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_6                    2 1 36 39 40 0=0
+Split                    splitncnn_8              1 2 40 41 42
+Convolution              convrelu_7               1 1 42 43 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_42                  1 1 43 44 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_7                    2 1 41 44 45 0=0
+Split                    splitncnn_9              1 2 45 46 47
+Convolution              convrelu_8               1 1 47 48 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_44                  1 1 48 49 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_8                    2 1 46 49 50 0=0
+Split                    splitncnn_10             1 2 50 51 52
+Convolution              convrelu_9               1 1 52 53 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_46                  1 1 53 54 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_9                    2 1 51 54 55 0=0
+Split                    splitncnn_11             1 2 55 56 57
+Convolution              convrelu_10              1 1 57 58 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_48                  1 1 58 59 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_10                   2 1 56 59 60 0=0
+Split                    splitncnn_12             1 2 60 61 62
+Convolution              convrelu_11              1 1 62 63 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_50                  1 1 63 64 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_11                   2 1 61 64 65 0=0
+Split                    splitncnn_13             1 2 65 66 67
+Convolution              convrelu_12              1 1 67 68 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_52                  1 1 68 69 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_12                   2 1 66 69 70 0=0
+Split                    splitncnn_14             1 2 70 71 72
+Convolution              convrelu_13              1 1 72 73 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_54                  1 1 73 74 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_13                   2 1 71 74 75 0=0
+Split                    splitncnn_15             1 2 75 76 77
+Convolution              convrelu_14              1 1 77 78 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_56                  1 1 78 79 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_14                   2 1 76 79 80 0=0
+Split                    splitncnn_16             1 2 80 81 82
+Convolution              convrelu_15              1 1 82 83 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_58                  1 1 83 84 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_15                   2 1 81 84 85 0=0
+Split                    splitncnn_17             1 2 85 86 87
+Convolution              convrelu_16              1 1 87 88 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_60                  1 1 88 89 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_16                   2 1 86 89 90 0=0
+Split                    splitncnn_18             1 2 90 91 92
+Convolution              convrelu_17              1 1 92 93 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_62                  1 1 93 94 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_17                   2 1 91 94 95 0=0
+Split                    splitncnn_19             1 2 95 96 97
+Convolution              convrelu_18              1 1 97 98 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_64                  1 1 98 99 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_18                   2 1 96 99 100 0=0
+Split                    splitncnn_20             1 2 100 101 102
+Convolution              convrelu_19              1 1 102 103 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_66                  1 1 103 104 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_19                   2 1 101 104 105 0=0
+Split                    splitncnn_21             1 2 105 106 107
+Convolution              convrelu_20              1 1 107 108 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_68                  1 1 108 109 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_20                   2 1 106 109 110 0=0
+Split                    splitncnn_22             1 2 110 111 112
+Convolution              convrelu_21              1 1 112 113 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_70                  1 1 113 114 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_21                   2 1 111 114 115 0=0
+Split                    splitncnn_23             1 2 115 116 117
+Convolution              convrelu_22              1 1 117 118 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_72                  1 1 118 119 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_22                   2 1 116 119 120 0=0
+Split                    splitncnn_24             1 2 120 121 122
+Convolution              convrelu_23              1 1 122 123 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_74                  1 1 123 124 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_23                   2 1 121 124 125 0=0
+BinaryOp                 add_24                   2 1 5 125 126 0=0
+Convolution              conv_75                  1 1 126 127 0=12 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=6912
+PixelShuffle             pixel_shuffle_1          1 1 127 128 0=2 1=0
+BinaryOp                 add_25                   2 1 128 7 out0 0=0

ncnn/4x_NERVE_fp16.ncnn.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50dd0a50e4a64496ba11c0627bae2ef76cfead9ddd7e8fb4d185cf01a688b214
+size 3604040

ncnn/4x_NERVE_fp16.ncnn.param

@@ -0,0 +1,106 @@
+7767517
+104 130
+Input                    in0                      0 1 in0
+Split                    splitncnn_0              1 2 in0 1 2
+Convolution              conv_26                  1 1 2 3 0=64 1=5 11=5 12=1 13=1 14=2 2=1 3=1 4=2 5=0 6=4800
+Split                    splitncnn_1              1 3 3 4 5 6
+Interp                   interpolate_0            1 1 1 7 0=3 1=4.0 2=4.0 6=0
+Convolution              convrelu_0               1 1 6 8 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_28                  1 1 8 9 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_0                    2 1 4 9 10 0=0
+Split                    splitncnn_2              1 2 10 11 12
+Convolution              convrelu_1               1 1 12 13 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_30                  1 1 13 14 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_1                    2 1 11 14 15 0=0
+Split                    splitncnn_3              1 2 15 16 17
+Convolution              convrelu_2               1 1 17 18 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_32                  1 1 18 19 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_2                    2 1 16 19 20 0=0
+Split                    splitncnn_4              1 2 20 21 22
+Convolution              convrelu_3               1 1 22 23 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_34                  1 1 23 24 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_3                    2 1 21 24 25 0=0
+Split                    splitncnn_5              1 2 25 26 27
+Convolution              convrelu_4               1 1 27 28 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_36                  1 1 28 29 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_4                    2 1 26 29 30 0=0
+Split                    splitncnn_6              1 2 30 31 32
+Convolution              convrelu_5               1 1 32 33 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_38                  1 1 33 34 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_5                    2 1 31 34 35 0=0
+Split                    splitncnn_7              1 2 35 36 37
+Convolution              convrelu_6               1 1 37 38 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_40                  1 1 38 39 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_6                    2 1 36 39 40 0=0
+Split                    splitncnn_8              1 2 40 41 42
+Convolution              convrelu_7               1 1 42 43 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_42                  1 1 43 44 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_7                    2 1 41 44 45 0=0
+Split                    splitncnn_9              1 2 45 46 47
+Convolution              convrelu_8               1 1 47 48 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_44                  1 1 48 49 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_8                    2 1 46 49 50 0=0
+Split                    splitncnn_10             1 2 50 51 52
+Convolution              convrelu_9               1 1 52 53 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_46                  1 1 53 54 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_9                    2 1 51 54 55 0=0
+Split                    splitncnn_11             1 2 55 56 57
+Convolution              convrelu_10              1 1 57 58 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_48                  1 1 58 59 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_10                   2 1 56 59 60 0=0
+Split                    splitncnn_12             1 2 60 61 62
+Convolution              convrelu_11              1 1 62 63 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_50                  1 1 63 64 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_11                   2 1 61 64 65 0=0
+Split                    splitncnn_13             1 2 65 66 67
+Convolution              convrelu_12              1 1 67 68 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_52                  1 1 68 69 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_12                   2 1 66 69 70 0=0
+Split                    splitncnn_14             1 2 70 71 72
+Convolution              convrelu_13              1 1 72 73 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_54                  1 1 73 74 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_13                   2 1 71 74 75 0=0
+Split                    splitncnn_15             1 2 75 76 77
+Convolution              convrelu_14              1 1 77 78 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_56                  1 1 78 79 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_14                   2 1 76 79 80 0=0
+Split                    splitncnn_16             1 2 80 81 82
+Convolution              convrelu_15              1 1 82 83 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_58                  1 1 83 84 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_15                   2 1 81 84 85 0=0
+Split                    splitncnn_17             1 2 85 86 87
+Convolution              convrelu_16              1 1 87 88 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_60                  1 1 88 89 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_16                   2 1 86 89 90 0=0
+Split                    splitncnn_18             1 2 90 91 92
+Convolution              convrelu_17              1 1 92 93 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_62                  1 1 93 94 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_17                   2 1 91 94 95 0=0
+Split                    splitncnn_19             1 2 95 96 97
+Convolution              convrelu_18              1 1 97 98 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_64                  1 1 98 99 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_18                   2 1 96 99 100 0=0
+Split                    splitncnn_20             1 2 100 101 102
+Convolution              convrelu_19              1 1 102 103 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_66                  1 1 103 104 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_19                   2 1 101 104 105 0=0
+Split                    splitncnn_21             1 2 105 106 107
+Convolution              convrelu_20              1 1 107 108 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_68                  1 1 108 109 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_20                   2 1 106 109 110 0=0
+Split                    splitncnn_22             1 2 110 111 112
+Convolution              convrelu_21              1 1 112 113 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_70                  1 1 113 114 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_21                   2 1 111 114 115 0=0
+Split                    splitncnn_23             1 2 115 116 117
+Convolution              convrelu_22              1 1 117 118 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_72                  1 1 118 119 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_22                   2 1 116 119 120 0=0
+Split                    splitncnn_24             1 2 120 121 122
+Convolution              convrelu_23              1 1 122 123 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864 9=1
+Convolution              conv_74                  1 1 123 124 0=64 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=36864
+BinaryOp                 add_23                   2 1 121 124 125 0=0
+BinaryOp                 add_24                   2 1 5 125 126 0=0
+Convolution              conv_75                  1 1 126 127 0=48 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=0 6=27648
+PixelShuffle             pixel_shuffle_1          1 1 127 128 0=4 1=0
+BinaryOp                 add_25                   2 1 128 7 out0 0=0

nerve_arch.py

@@ -0,0 +1,86 @@
+# NERVE — lightweight super-resolution for mobile/edge (single config).
+#
+# Conv-ReLU-Conv residual blocks, no normalization, no attention, no gating.
+# A 5x5 stem opens the receptive field for real-world degradations; a global
+# residual and a bicubic input residual let the net learn only what bicubic
+# upscaling misses. dim=64, n_blocks=24, ~1.8M params at 4x.
+#
+# Norm-free is decided by ablation (NERVE_ABLATION_LOG.md §4): removing
+# LayerNorm won +1.15 dB at ~2x the training speed. dim=64/n_blocks=24 is the
+# capacity-sweep sweet spot (§15): 0.69M -> 1.80M bought a large GAN quality
+# gain (lpips -0.015), while 1.80M -> 2.39M bought ~4x less gain per parameter.
+# Every op in the graph is standard (Conv/Add/ReLU/PixelShuffle/bicubic
+# Resize), so the checkpoint exports to dynamic-shape ONNX and converts to
+# NCNN/CoreML without fused/unfused pairs.
+
+from spandrel.util import store_hyperparameters
+from torch import Tensor, nn
+from torch.nn import functional as F  # noqa: N812
+
+from traiNNer.utils.icnr import icnr_reinit
+from traiNNer.utils.registry import ARCH_REGISTRY, SPANDREL_REGISTRY
+
+
+class NerveBlock(nn.Module):
+    """Conv-ReLU-Conv residual block. No normalization, no shortcuts."""
+
+    def __init__(self, dim: int) -> None:
+        super().__init__()
+        self.conv1 = nn.Conv2d(dim, dim, 3, padding=1, bias=False)
+        self.conv2 = nn.Conv2d(dim, dim, 3, padding=1, bias=False)
+
+    def forward(self, x: Tensor) -> Tensor:
+        out = self.conv2(F.relu(self.conv1(x)))
+        return x + out
+
+
+@store_hyperparameters()
+class NERVE(nn.Module):
+    """5x5 stem -> n_blocks residual blocks -> PixelShuffle head -> + bicubic(x)."""
+
+    hyperparameters = {}  # noqa: RUF012
+
+    def __init__(
+        self,
+        *,
+        in_ch: int = 3,
+        out_ch: int = 3,
+        dim: int = 64,
+        n_blocks: int = 24,
+        upscale: int = 4,
+    ) -> None:
+        super().__init__()
+        self.upscale = upscale
+
+        self.stem = nn.Conv2d(in_ch, dim, 5, padding=2, bias=False)
+        self.blocks = nn.Sequential(*[NerveBlock(dim) for _ in range(n_blocks)])
+        self.upsampler = nn.Sequential(
+            nn.Conv2d(dim, out_ch * upscale**2, 3, padding=1, bias=False),
+            nn.PixelShuffle(upscale),
+        )
+        # ICNR: every sub-pixel phase starts from the same base kernel, so the
+        # head cannot train up a checkerboard. Init-only, zero graph/runtime cost.
+        icnr_reinit(self.upsampler)
+
+    def forward(self, x: Tensor) -> Tensor:
+        feat = self.stem(x)
+        out = feat + self.blocks(feat)
+        out = self.upsampler(out)
+
+        base = F.interpolate(
+            x, scale_factor=self.upscale, mode="bicubic", align_corners=False
+        )
+        return out + base
+
+
+@SPANDREL_REGISTRY.register()
+@ARCH_REGISTRY.register()
+def nerve(
+    scale: int = 4,
+    *,
+    in_ch: int = 3,
+    out_ch: int = 3,
+    dim: int = 64,
+    n_blocks: int = 24,
+) -> NERVE:
+    return NERVE(in_ch=in_ch, out_ch=out_ch, dim=dim, n_blocks=n_blocks, upscale=scale)

onnx/2x_NERVE_1x3xHxW_fp32_op20.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:df4e8b6c5fecc1fd405c23d7a678fe19d64b2b69f2af189be0765c1c6c57a154
-size 2793602
+oid sha256:a499bc54cb9d442a94f63fab201de61c72b32bf88bab6dffffbd5abbb43cb5f7
+size 7145533

onnx/4x_NERVE_1x3xHxW_fp32_op20.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0a3df17cade4ab62ac6160170a1a1aee0663530d366da09152e4341294fc15e0
-size 2855810
+oid sha256:eeecdc430db756f36096b110ee275d860feab6b5c036a7c94e55c82d2d01b616
+size 7228477

scripts/export_dynamic.py

@@ -0,0 +1,163 @@
+"""Export NERVE (canonical, configurable size) or SPAN to dynamic-shape ONNX and verify.
+
+- fp32 export with dynamic H/W on the lq/sr tensors
+- onnx.checker
+- ONNX Runtime CPU inference vs PyTorch CPU (max abs diff) at several shapes:
+  48x48, 323x711, 720x1280.
+
+Random weights prove the graph is export-clean; pass a trained EMA checkpoint
+(--checkpoint) to validate an actual trained model. Canonical export is
+opset 20 (torch 2.14's dynamo exporter has an opset floor of 18; an opset-17
+request still emits 18).
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+import numpy as np
+import torch
+from torch.export import Dim
+from traiNNer.archs.nerve_arch import nerve
+
+ROOT = Path(__file__).resolve().parents[2]
+OUT_ROOT = ROOT / "experiments" / "nerve" / "export"
+
+SHAPES: list[tuple[int, int]] = [(48, 48), (323, 711), (720, 1280)]
+
+
+def load_checkpoint(model: torch.nn.Module, path: Path) -> None:
+    if path.suffix == ".safetensors":
+        try:
+            from safetensors.torch import load_file
+
+            sd = load_file(str(path), device="cpu")
+        except Exception:  # noqa: BLE001 - legacy pickle with .safetensors name
+            import io
+
+            sd = torch.load(
+                io.BytesIO(path.read_bytes()), map_location="cpu", weights_only=True
+            )
+    else:
+        sd = torch.load(path, map_location="cpu", weights_only=True)
+    missing, unexpected = model.load_state_dict(sd, strict=True)
+    if missing or unexpected:
+        raise RuntimeError(
+            f"checkpoint mismatch: missing={missing} unexpected={unexpected}"
+        )
+
+
+def build(arch: str, scale: int, dim: int, n_blocks: int) -> torch.nn.Module:
+    if arch == "span":
+        from traiNNer.archs.span_arch import span
+
+        return span(scale=scale)
+    return nerve(scale=scale, dim=dim, n_blocks=n_blocks)
+
+
+def export_onnx(
+    model: torch.nn.Module,
+    scale: int,
+    opset: int,
+    out_path: Path,
+) -> None:
+    model.eval()
+    x = torch.randn(1, 3, 64, 64)
+    dynamic_shapes = {"x": {2: Dim("H"), 3: Dim("W")}}
+    torch.onnx.export(
+        model,
+        (x,),
+        str(out_path),
+        opset_version=opset,
+        dynamic_shapes=dynamic_shapes,
+    )
+
+
+def verify_ort(model: torch.nn.Module, onnx_path: Path, scale: int, tol: float) -> dict:
+    import onnx
+    import onnxruntime as ort
+
+    onnx.checker.check_model(onnx_path)
+    sess = ort.InferenceSession(str(onnx_path), providers=["CPUExecutionProvider"])
+    iname = sess.get_inputs()[0].name
+
+    model.eval()
+    results: dict[str, float] = {}
+    with torch.no_grad():
+        for h, w in SHAPES:
+            x = torch.randn(1, 3, h, w)
+            ref = model(x)
+            ort_out = sess.run(None, {iname: x.numpy()})[0]
+            diff = float(np.abs(ref.numpy() - ort_out).max())
+            results[f"{h}x{w}"] = round(diff, 6)
+    return results
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--arch", default="nerve", choices=["nerve", "span"])
+    parser.add_argument("--dim", type=int, default=48)
+    parser.add_argument("--n-blocks", type=int, default=16)
+    parser.add_argument(
+        "--checkpoint", default=None, help="trained EMA safetensors to load"
+    )
+    parser.add_argument(
+        "--scale", type=int, default=4, help="scale when a checkpoint is given"
+    )
+    parser.add_argument("--scales", nargs="+", type=int, default=[4, 2])
+    parser.add_argument("--opsets", nargs="+", type=int, default=[20])
+    parser.add_argument("--tol", type=float, default=1e-3)
+    parser.add_argument("--out-root", default=str(OUT_ROOT))
+    args = parser.parse_args()
+
+    scales = [args.scale] if args.checkpoint else args.scales
+    overall_rc = 0
+    summary: dict = {}
+
+    for scale in scales:
+        model = build(args.arch, scale, args.dim, args.n_blocks)
+        model.to("cpu")
+        model.eval()
+        if args.checkpoint:
+            load_checkpoint(model, Path(args.checkpoint))
+        params = sum(p.numel() for p in model.parameters())
+        print(f"arch {args.arch} scale {scale} params={params:,}", flush=True)
+
+        scale_rc = 0
+        for opset in args.opsets:
+            tag = f"{args.arch}_d{args.dim}_b{args.n_blocks}_s{scale}"
+            out_dir = Path(args.out_root) / tag
+            out_dir.mkdir(parents=True, exist_ok=True)
+            out_path = out_dir / f"nerve_dynamic_op{opset}.onnx"
+            export_onnx(model, scale, opset, out_path)
+            print(f"exported {out_path.name}", flush=True)
+            try:
+                diffs = verify_ort(model, out_path, scale, args.tol)
+            except Exception as e:  # noqa: BLE001
+                print(f"opset {opset} ORT verify FAILED: {e}", flush=True)
+                scale_rc = 1
+                continue
+            worst = max(diffs.values()) if diffs else float("inf")
+            ok = worst <= args.tol
+            print(
+                f"opset {opset} max|diff| per shape: {diffs} -> {'OK' if ok else 'FAIL'}",
+                flush=True,
+            )
+            scale_rc |= 0 if ok else 1
+            summary[f"arch{args.arch}_s{scale}_op{opset}"] = {
+                "max_diff": diffs,
+                "ok": ok,
+            }
+        overall_rc |= scale_rc
+
+    (Path(args.out_root) / "export_summary.json").write_text(
+        json.dumps(summary, indent=2)
+    )
+    return overall_rc
+
+
+if __name__ == "__main__":
+    sys.exit(main())