fomo-iqa 0.1.0__tar.gz

fomo_iqa-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 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.

fomo_iqa-0.1.0/PKG-INFO

@@ -0,0 +1,219 @@
+Metadata-Version: 2.4
+Name: fomo-iqa
+Version: 0.1.0
+Summary: FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance — reference-based perceptual distance models trained on diffusion-generated (FLUX) image pairs
+Author: Jaihyun Lew
+License: Apache-2.0
+Project-URL: Homepage, https://github.com/JHLew/FoMo
+Keywords: image-quality-assessment,perceptual-metric,perceptual-loss,IQA
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: torch>=2.1
+Requires-Dist: torchvision
+Requires-Dist: transformers>=4.56
+Requires-Dist: huggingface_hub
+Requires-Dist: lpips>=0.1.4
+Requires-Dist: DISTS-pytorch>=0.1
+Requires-Dist: dreamsim>=0.2.1
+Requires-Dist: peft
+Requires-Dist: numpy
+Requires-Dist: pillow
+Dynamic: license-file
+
+# FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance
+
+<p align="center">
+  <img src="figures/FoMo_main.png" alt="FoMo: data generation from the forking moment of the diffusion trajectory, and ranked-BCE training" width="100%">
+</p>
+<p align="center">
+  <img src="figures/FoMo_generated_pairs.png" alt="Examples of generated image pairs at different forking moments" width="100%">
+</p>
+
+FoMo trains reference-based perceptual distance metrics **without human labels**. Running a diffusion model (FLUX.1-dev) in img2img mode from a base image, re-entering the trajectory `s` denoising steps before the end, produces a variant whose perceptual deviation from the base grows monotonically with `s` — the *forking moment* of the generative trajectory. The timestep is therefore a free, dense supervision signal: we generate 480k such image sets and train distance models with a pairwise ranking objective on it.
+
+Seven models are released, all trained on the same 480k FoMo dataset with the same ranked-BCE objective; only the frozen backbone / trainable part differs:
+
+| Name | Architecture | Trained part |
+|---|---|---|
+| `lpips_alex` | LPIPS, AlexNet features | linear calibration weights |
+| `lpips_vgg`  | LPIPS, VGG-16 features  | linear calibration weights |
+| `dists`      | DISTS, VGG-16 features  | α / β mixing weights |
+| `dinov3`     | frozen DINOv3 ViT-S/16+ + prediction head | 3-layer 2D-RoPE transformer head |
+| `clip`       | frozen CLIP ViT-B/32 + prediction head    | 3-layer 2D-RoPE transformer head |
+| `mae`        | frozen MAE ViT-B/16 + prediction head     | 3-layer 2D-RoPE transformer head |
+| `dreamsim`   | DreamSim ensemble (DINO + CLIP + OpenCLIP ViT-B/16) | LoRA adapters (r=16) |
+
+## Using the pretrained models
+
+```bash
+pip install fomo-iqa          # or, from this repository:  pip install -e .
+```
+
+(The distribution is `fomo-iqa` because `fomo` is taken on PyPI; the import name is `fomo`.)
+
+```python
+import fomo
+
+# A model is defined by its backbone alone. How the score is presented —
+# `symmetric` and `output` — is chosen per call, so one model serves every use.
+
+# --- a CNN model --------------------------------------------------------------
+lpips_alex = fomo.FoMo('lpips_alex')
+d = lpips_alex(img_ref, img_test)                        # [B]; lower = more similar
+                                                         # already non-negative with d(x, x) = 0,
+                                                         # and symmetric by construction, so
+                                                         # `symmetric` and 'distance' change nothing
+
+# --- a prediction-head model --------------------------------------------------
+dinov3 = fomo.FoMo('dinov3')
+d_raw  = dinov3(img_ref, img_test, symmetric=True)                     # 'raw' (default): the
+                                                                       # published protocol. A ranking
+                                                                       # score with an arbitrary zero
+                                                                       # point, so it may be negative
+d_dist = dinov3(img_ref, img_test, symmetric=True, output='distance')  # non-negative, d(x, x) ≈ 0,
+                                                                       # LPIPS-style. Ranks pairs
+                                                                       # identically to 'raw'
+d_asym = dinov3(img_ref, img_test, symmetric=False)                    # single order s(x, y)
+
+# `symmetric=True` is the default and only affects dinov3 / clip / mae, where the
+# head reads the two images as one sequence: it averages both argument orders
+# from a single backbone pass.
+
+# --- as a perceptual loss ------------------------------------------------------
+# differentiable w.r.t. the inputs. output='loss' is hinged at the identity
+# level, so optimization settles once the images are as close as identical
+# instead of continuing past that point.
+loss = dinov3(prediction, target, output='loss').mean()
+loss.backward()
+```
+
+`fomo.load(name)` returns the underlying `DistanceModel` (method `eval_distance`), and `checkpoint='path/to/file.pth'` loads a local file — including a `ckpt_latest.pt` produced by `train.py` (its EMA weights are used). `python example_usage.py --model dinov3 --mode metric|loss` is a runnable demo.
+
+**Inputs.** Float tensors `[B, 3, H, W]` in `[0, 1]`. Do **not** normalize — each model applies its own preprocessing (ImageNet statistics for DINOv3 / MAE, CLIP statistics for CLIP, `[-1, 1]` for LPIPS, raw for DISTS, bicubic resize to 224 for DreamSim). Both images of a pair must share a spatial size; any size works (the ViT models zero-pad to their patch grid). Benchmarks are scored at native resolution.
+
+**Distance options.**
+
+- `lpips_alex`, `lpips_vgg`, `dists` and `dreamsim` compute a symmetric distance directly (DreamSim as the cosine distance of its embeddings). `dinov3`, `clip` and `mae` read the two images as one concatenated token sequence and output a scalar for that ordered pair.
+- `symmetric` (default `True`): for `dinov3` / `clip` / `mae` the output `s(x, y)` need not equal `s(y, x)`, so the score is `½ [s(x, y) + s(y, x)]`, both orders from a single backbone pass. `symmetric=False` returns the single order `s(x, y)`.
+- `output` (default `'raw'`): which of three presentations the score is returned in. `'raw'` and `'distance'` are the same metric in two scales — softplus is strictly increasing, so they rank every pair identically (SROCC and KROCC agree to the last digit). `'loss'` subtracts a per-pair quantity and is shaped for optimization rather than for ranking.
+
+  | `output` | what it returns | use it for |
+  |---|---|---|
+  | `'raw'` (default) | the model output, unchanged | reproducing the published numbers |
+  | `'distance'` | non-negative, `d(x, x) ≈ 0` | a drop-in LPIPS-style metric |
+  | `'loss'` | `relu( d(x, y) − max(d(x, x), d(y, y)) )` | a perceptual loss / image optimization |
+
+  `'raw'` is the protocol every published number was computed under. For `dinov3` / `clip` / `mae` it is a ranking score whose zero point is arbitrary, so it may be negative; `lpips_alex`, `lpips_vgg`, `dists` and `dreamsim` return a non-negative distance with `d(x, x) = 0` directly.
+
+  `'distance'` applies softplus to `dinov3` / `clip` / `mae` — mapping their identity level to `≈ 0` and giving a positive output everywhere, with no calibration constant to fit or ship — and leaves the other four untouched, since they already have an exact `d(x, x) = 0` (softplus would move it to `log 2 ≈ 0.693`). The rule differs by model family so that the *semantics* are uniform: all seven are non-negative with `d(x, x) ≈ 0`. Rankings are identical to `'raw'`; use `'raw'` to reproduce the published numbers exactly.
+
+  `'loss'` is a hinge whose floor sits at the per-pair identity level: full gradient while the images differ, and zero once they are as close as identical, so optimization stops there rather than pushing past it. Softplus is *not* applied — the relu already makes the output non-negative. The anchor is detached, because it depends on the image being optimized: with gradients flowing through it, the optimizer could lower the loss by raising `d(y, y)` rather than matching the reference.
+
+**Backbone downloads.** The frozen backbones are fetched from their public sources on first use (torchvision for AlexNet / VGG, HuggingFace for CLIP / MAE / DINOv3, the DreamSim release for its three ViTs, ~1.5 GB into `~/.cache/fomo/dreamsim`, override with `FOMO_DREAMSIM_CACHE`). **DINOv3 is gated:** accept the license of `facebook/dinov3-vits16plus-pretrain-lvd1689m` on HuggingFace and run `huggingface-cli login` before using `dinov3`.
+
+## Repository layout
+
+```
+fomo/                        pip package (metric / loss usage)
+├── model.py                 DistanceModel — the seven models
+├── dinov3.py                prediction head (2D-RoPE attention, SwiGLU, registers)
+└── hub.py                   fomo.load() / fomo.FoMo, checkpoint download
+generate_fomo_dataset.py     FLUX img2img data generation
+dataset.py                   FoMoDataset + paired augmentation, benchmark loaders
+train.py                     ranked-BCE training
+evaluation.py / evaluate.py  PIPAL / TID2013 / CSIQ / LIVE metrics, CLI
+example_usage.py             metric + loss demo
+checkpoints/                 model card for the Hub repo (the .pth files live on the Hub)
+scripts/
+├── download_fomo_480k.sh    fetch the released 480k training set
+├── generate_fomo_480k.sh    …or regenerate it with FLUX
+├── setup_benchmarks.sh      fetch the four evaluation benchmarks
+└── train.sh                 train one model with the paper's configuration
+```
+
+## Environment
+
+```bash
+conda env create -f environment.yml     # env "fomo": torch 2.9.1, transformers 4.57.3, diffusers 0.37.1
+conda activate fomo
+```
+
+Training assumes a bf16-capable GPU (an L40 / A100 class card runs every recipe at batch 64); evaluation runs on one GPU.
+
+## Data
+
+**Download the released dataset** (recommended) — 480,000 samples at 512×512, hosted as 240 tar shards (~550 GB) at [JHLew/fomo_480k](https://huggingface.co/datasets/JHLew/fomo_480k):
+
+```bash
+bash scripts/download_fomo_480k.sh      # → ./data/fomo/{train/…, index.json}
+```
+
+Each sample is a base image plus two img2img variants at random timesteps `s ∈ [1, 50]` (FLUX.1-dev, 50 steps, guidance 3.5; the variant files are named by `s`). The `s` used in this repository and in the dataset corresponds to `S − s` in the paper, with `S = 50`. Half of the bases are generated from a blank canvas (`FLUX_*`), half are ImageNet-1k crops (`IN1K_*`). `index.json` lists `{"base": …, "variants": […]}` with paths relative to the dataset root.
+
+**Or regenerate it** — this is the data-generation pipeline, the core of the method:
+
+```bash
+huggingface-cli login                    # FLUX.1-dev is gated
+bash scripts/generate_fomo_480k.sh       # STAGE=scratch|in1k|both, shard with IDX_FROM / IDX_TO
+```
+
+`generate_fomo_dataset.py` writes `<out>/train/<TAG>_<idx>/{base.png, <t1>.png, <t2>.png}` and appends to `index.json`. The IN1K stage needs ImageNet-1k at `./imagenet-1k/train/<class>/*.JPEG`. Generation is expensive (2–3 FLUX calls per sample); shard the index range across GPUs.
+
+**Benchmarks:**
+
+```bash
+bash scripts/setup_benchmarks.sh         # PIPAL, TID2013, CSIQ, LIVE → ./benchmarks/
+```
+
+## Training
+
+One command trains a model with the paper's configuration and evaluates it on the four benchmarks after every epoch:
+
+```bash
+bash scripts/train.sh dino            # lpips_alex | lpips_vgg | dists | dino | clip | mae | dreamsim
+```
+
+which runs
+
+```bash
+python train.py --exp_name dino --backbone dino \
+    --dataroot ./data/fomo --train_index_file ./data/fomo/index.json \
+    --epochs 3 --batch_size 64 --lr 1e-4 --min_crop_size 64 --max_crop_size 256
+```
+
+Each sample yields one training pair: two of its three images in random order (or, with probability 1/51, an image paired with itself), labelled with the FLUX-space value of the larger timestep `s` of the two, recomputed for the sampled resolution. The ranked-BCE loss scores every ordered pair of samples in the batch against this label ordering. Settings:
+
+- `lpips_alex` / `lpips_vgg` / `dists`: `--lr 2e-4`, fixed crops `--min_crop_size 256 --max_crop_size 256`.
+- `dino` / `clip` / `mae`: `--lr 1e-4`, random 64–256 crops batched by zero-padding + attention masking.
+- `dreamsim`: `--lr 1e-4`, fixed `--min_crop_size 224 --max_crop_size 224` (fixed-resolution ensemble).
+- Benchmarks are evaluated after every epoch with the **EMA weights at native resolution**; the final epoch is what is reported. Outputs go to `experiments/<exp_name>/` (`ckpt_latest.pt`, `metrics.json`, `train.log`, TensorBoard); rerunning the same command resumes.
+
+The final-epoch numbers of such a run land close to the table below. Any checkpoint can be re-scored on its own:
+
+```bash
+python evaluate.py --model dinov3                                                  # released weights
+python evaluate.py --model dinov3 --checkpoint experiments/dino/ckpt_latest.pt      # your run
+```
+
+### Results
+
+SROCC / KROCC / PLCC of the released checkpoints (native resolution; PLCC after a 5-parameter logistic fit). `python evaluate.py --model <name>` reproduces every cell:
+
+| Model | PIPAL | TID2013 | CSIQ | LIVE |
+|---|---|---|---|---|
+| LPIPS-Alex| 0.739 / 0.541 / 0.773 | 0.782 / 0.583 / 0.813 | 0.940 / 0.784 / 0.946 | 0.949 / 0.794 / 0.943 |
+| LPIPS-VGG | 0.691 / 0.503 / 0.739 | 0.659 / 0.486 / 0.746 | 0.852 / 0.665 / 0.880 | 0.918 / 0.740 / 0.918 |
+| DISTS     | 0.616 / 0.437 / 0.657 | 0.692 / 0.513 / 0.771 | 0.919 / 0.750 / 0.931 | 0.954 / 0.804 / 0.949 |
+| DINOv3    | 0.705 / 0.504 / 0.709 | 0.711 / 0.524 / 0.769 | 0.810 / 0.626 / 0.874 | 0.897 / 0.721 / 0.913 |
+| CLIP      | 0.668 / 0.479 / 0.679 | 0.744 / 0.556 / 0.800 | 0.911 / 0.735 / 0.928 | 0.935 / 0.770 / 0.936 |
+| MAE       | 0.644 / 0.453 / 0.634 | 0.674 / 0.495 / 0.681 | 0.817 / 0.625 / 0.830 | 0.919 / 0.752 / 0.926 |
+| DreamSim  | 0.790 / 0.585 / 0.779 | 0.806 / 0.610 / 0.836 | 0.894 / 0.710 / 0.905 | 0.933 / 0.773 / 0.936 |
+
+## Citation
+
+The paper is under review; citation information will be added here.
+
+## License
+
+Apache 2.0 — see [LICENSE](LICENSE). The backbones (DINOv3, CLIP, MAE, DreamSim's ViTs) and the `lpips` / `DISTS-pytorch` / `dreamsim` packages keep their own upstream licenses.

fomo_iqa-0.1.0/README.md

@@ -0,0 +1,196 @@
+# FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance
+
+<p align="center">
+  <img src="figures/FoMo_main.png" alt="FoMo: data generation from the forking moment of the diffusion trajectory, and ranked-BCE training" width="100%">
+</p>
+<p align="center">
+  <img src="figures/FoMo_generated_pairs.png" alt="Examples of generated image pairs at different forking moments" width="100%">
+</p>
+
+FoMo trains reference-based perceptual distance metrics **without human labels**. Running a diffusion model (FLUX.1-dev) in img2img mode from a base image, re-entering the trajectory `s` denoising steps before the end, produces a variant whose perceptual deviation from the base grows monotonically with `s` — the *forking moment* of the generative trajectory. The timestep is therefore a free, dense supervision signal: we generate 480k such image sets and train distance models with a pairwise ranking objective on it.
+
+Seven models are released, all trained on the same 480k FoMo dataset with the same ranked-BCE objective; only the frozen backbone / trainable part differs:
+
+| Name | Architecture | Trained part |
+|---|---|---|
+| `lpips_alex` | LPIPS, AlexNet features | linear calibration weights |
+| `lpips_vgg`  | LPIPS, VGG-16 features  | linear calibration weights |
+| `dists`      | DISTS, VGG-16 features  | α / β mixing weights |
+| `dinov3`     | frozen DINOv3 ViT-S/16+ + prediction head | 3-layer 2D-RoPE transformer head |
+| `clip`       | frozen CLIP ViT-B/32 + prediction head    | 3-layer 2D-RoPE transformer head |
+| `mae`        | frozen MAE ViT-B/16 + prediction head     | 3-layer 2D-RoPE transformer head |
+| `dreamsim`   | DreamSim ensemble (DINO + CLIP + OpenCLIP ViT-B/16) | LoRA adapters (r=16) |
+
+## Using the pretrained models
+
+```bash
+pip install fomo-iqa          # or, from this repository:  pip install -e .
+```
+
+(The distribution is `fomo-iqa` because `fomo` is taken on PyPI; the import name is `fomo`.)
+
+```python
+import fomo
+
+# A model is defined by its backbone alone. How the score is presented —
+# `symmetric` and `output` — is chosen per call, so one model serves every use.
+
+# --- a CNN model --------------------------------------------------------------
+lpips_alex = fomo.FoMo('lpips_alex')
+d = lpips_alex(img_ref, img_test)                        # [B]; lower = more similar
+                                                         # already non-negative with d(x, x) = 0,
+                                                         # and symmetric by construction, so
+                                                         # `symmetric` and 'distance' change nothing
+
+# --- a prediction-head model --------------------------------------------------
+dinov3 = fomo.FoMo('dinov3')
+d_raw  = dinov3(img_ref, img_test, symmetric=True)                     # 'raw' (default): the
+                                                                       # published protocol. A ranking
+                                                                       # score with an arbitrary zero
+                                                                       # point, so it may be negative
+d_dist = dinov3(img_ref, img_test, symmetric=True, output='distance')  # non-negative, d(x, x) ≈ 0,
+                                                                       # LPIPS-style. Ranks pairs
+                                                                       # identically to 'raw'
+d_asym = dinov3(img_ref, img_test, symmetric=False)                    # single order s(x, y)
+
+# `symmetric=True` is the default and only affects dinov3 / clip / mae, where the
+# head reads the two images as one sequence: it averages both argument orders
+# from a single backbone pass.
+
+# --- as a perceptual loss ------------------------------------------------------
+# differentiable w.r.t. the inputs. output='loss' is hinged at the identity
+# level, so optimization settles once the images are as close as identical
+# instead of continuing past that point.
+loss = dinov3(prediction, target, output='loss').mean()
+loss.backward()
+```
+
+`fomo.load(name)` returns the underlying `DistanceModel` (method `eval_distance`), and `checkpoint='path/to/file.pth'` loads a local file — including a `ckpt_latest.pt` produced by `train.py` (its EMA weights are used). `python example_usage.py --model dinov3 --mode metric|loss` is a runnable demo.
+
+**Inputs.** Float tensors `[B, 3, H, W]` in `[0, 1]`. Do **not** normalize — each model applies its own preprocessing (ImageNet statistics for DINOv3 / MAE, CLIP statistics for CLIP, `[-1, 1]` for LPIPS, raw for DISTS, bicubic resize to 224 for DreamSim). Both images of a pair must share a spatial size; any size works (the ViT models zero-pad to their patch grid). Benchmarks are scored at native resolution.
+
+**Distance options.**
+
+- `lpips_alex`, `lpips_vgg`, `dists` and `dreamsim` compute a symmetric distance directly (DreamSim as the cosine distance of its embeddings). `dinov3`, `clip` and `mae` read the two images as one concatenated token sequence and output a scalar for that ordered pair.
+- `symmetric` (default `True`): for `dinov3` / `clip` / `mae` the output `s(x, y)` need not equal `s(y, x)`, so the score is `½ [s(x, y) + s(y, x)]`, both orders from a single backbone pass. `symmetric=False` returns the single order `s(x, y)`.
+- `output` (default `'raw'`): which of three presentations the score is returned in. `'raw'` and `'distance'` are the same metric in two scales — softplus is strictly increasing, so they rank every pair identically (SROCC and KROCC agree to the last digit). `'loss'` subtracts a per-pair quantity and is shaped for optimization rather than for ranking.
+
+  | `output` | what it returns | use it for |
+  |---|---|---|
+  | `'raw'` (default) | the model output, unchanged | reproducing the published numbers |
+  | `'distance'` | non-negative, `d(x, x) ≈ 0` | a drop-in LPIPS-style metric |
+  | `'loss'` | `relu( d(x, y) − max(d(x, x), d(y, y)) )` | a perceptual loss / image optimization |
+
+  `'raw'` is the protocol every published number was computed under. For `dinov3` / `clip` / `mae` it is a ranking score whose zero point is arbitrary, so it may be negative; `lpips_alex`, `lpips_vgg`, `dists` and `dreamsim` return a non-negative distance with `d(x, x) = 0` directly.
+
+  `'distance'` applies softplus to `dinov3` / `clip` / `mae` — mapping their identity level to `≈ 0` and giving a positive output everywhere, with no calibration constant to fit or ship — and leaves the other four untouched, since they already have an exact `d(x, x) = 0` (softplus would move it to `log 2 ≈ 0.693`). The rule differs by model family so that the *semantics* are uniform: all seven are non-negative with `d(x, x) ≈ 0`. Rankings are identical to `'raw'`; use `'raw'` to reproduce the published numbers exactly.
+
+  `'loss'` is a hinge whose floor sits at the per-pair identity level: full gradient while the images differ, and zero once they are as close as identical, so optimization stops there rather than pushing past it. Softplus is *not* applied — the relu already makes the output non-negative. The anchor is detached, because it depends on the image being optimized: with gradients flowing through it, the optimizer could lower the loss by raising `d(y, y)` rather than matching the reference.
+
+**Backbone downloads.** The frozen backbones are fetched from their public sources on first use (torchvision for AlexNet / VGG, HuggingFace for CLIP / MAE / DINOv3, the DreamSim release for its three ViTs, ~1.5 GB into `~/.cache/fomo/dreamsim`, override with `FOMO_DREAMSIM_CACHE`). **DINOv3 is gated:** accept the license of `facebook/dinov3-vits16plus-pretrain-lvd1689m` on HuggingFace and run `huggingface-cli login` before using `dinov3`.
+
+## Repository layout
+
+```
+fomo/                        pip package (metric / loss usage)
+├── model.py                 DistanceModel — the seven models
+├── dinov3.py                prediction head (2D-RoPE attention, SwiGLU, registers)
+└── hub.py                   fomo.load() / fomo.FoMo, checkpoint download
+generate_fomo_dataset.py     FLUX img2img data generation
+dataset.py                   FoMoDataset + paired augmentation, benchmark loaders
+train.py                     ranked-BCE training
+evaluation.py / evaluate.py  PIPAL / TID2013 / CSIQ / LIVE metrics, CLI
+example_usage.py             metric + loss demo
+checkpoints/                 model card for the Hub repo (the .pth files live on the Hub)
+scripts/
+├── download_fomo_480k.sh    fetch the released 480k training set
+├── generate_fomo_480k.sh    …or regenerate it with FLUX
+├── setup_benchmarks.sh      fetch the four evaluation benchmarks
+└── train.sh                 train one model with the paper's configuration
+```
+
+## Environment
+
+```bash
+conda env create -f environment.yml     # env "fomo": torch 2.9.1, transformers 4.57.3, diffusers 0.37.1
+conda activate fomo
+```
+
+Training assumes a bf16-capable GPU (an L40 / A100 class card runs every recipe at batch 64); evaluation runs on one GPU.
+
+## Data
+
+**Download the released dataset** (recommended) — 480,000 samples at 512×512, hosted as 240 tar shards (~550 GB) at [JHLew/fomo_480k](https://huggingface.co/datasets/JHLew/fomo_480k):
+
+```bash
+bash scripts/download_fomo_480k.sh      # → ./data/fomo/{train/…, index.json}
+```
+
+Each sample is a base image plus two img2img variants at random timesteps `s ∈ [1, 50]` (FLUX.1-dev, 50 steps, guidance 3.5; the variant files are named by `s`). The `s` used in this repository and in the dataset corresponds to `S − s` in the paper, with `S = 50`. Half of the bases are generated from a blank canvas (`FLUX_*`), half are ImageNet-1k crops (`IN1K_*`). `index.json` lists `{"base": …, "variants": […]}` with paths relative to the dataset root.
+
+**Or regenerate it** — this is the data-generation pipeline, the core of the method:
+
+```bash
+huggingface-cli login                    # FLUX.1-dev is gated
+bash scripts/generate_fomo_480k.sh       # STAGE=scratch|in1k|both, shard with IDX_FROM / IDX_TO
+```
+
+`generate_fomo_dataset.py` writes `<out>/train/<TAG>_<idx>/{base.png, <t1>.png, <t2>.png}` and appends to `index.json`. The IN1K stage needs ImageNet-1k at `./imagenet-1k/train/<class>/*.JPEG`. Generation is expensive (2–3 FLUX calls per sample); shard the index range across GPUs.
+
+**Benchmarks:**
+
+```bash
+bash scripts/setup_benchmarks.sh         # PIPAL, TID2013, CSIQ, LIVE → ./benchmarks/
+```
+
+## Training
+
+One command trains a model with the paper's configuration and evaluates it on the four benchmarks after every epoch:
+
+```bash
+bash scripts/train.sh dino            # lpips_alex | lpips_vgg | dists | dino | clip | mae | dreamsim
+```
+
+which runs
+
+```bash
+python train.py --exp_name dino --backbone dino \
+    --dataroot ./data/fomo --train_index_file ./data/fomo/index.json \
+    --epochs 3 --batch_size 64 --lr 1e-4 --min_crop_size 64 --max_crop_size 256
+```
+
+Each sample yields one training pair: two of its three images in random order (or, with probability 1/51, an image paired with itself), labelled with the FLUX-space value of the larger timestep `s` of the two, recomputed for the sampled resolution. The ranked-BCE loss scores every ordered pair of samples in the batch against this label ordering. Settings:
+
+- `lpips_alex` / `lpips_vgg` / `dists`: `--lr 2e-4`, fixed crops `--min_crop_size 256 --max_crop_size 256`.
+- `dino` / `clip` / `mae`: `--lr 1e-4`, random 64–256 crops batched by zero-padding + attention masking.
+- `dreamsim`: `--lr 1e-4`, fixed `--min_crop_size 224 --max_crop_size 224` (fixed-resolution ensemble).
+- Benchmarks are evaluated after every epoch with the **EMA weights at native resolution**; the final epoch is what is reported. Outputs go to `experiments/<exp_name>/` (`ckpt_latest.pt`, `metrics.json`, `train.log`, TensorBoard); rerunning the same command resumes.
+
+The final-epoch numbers of such a run land close to the table below. Any checkpoint can be re-scored on its own:
+
+```bash
+python evaluate.py --model dinov3                                                  # released weights
+python evaluate.py --model dinov3 --checkpoint experiments/dino/ckpt_latest.pt      # your run
+```
+
+### Results
+
+SROCC / KROCC / PLCC of the released checkpoints (native resolution; PLCC after a 5-parameter logistic fit). `python evaluate.py --model <name>` reproduces every cell:
+
+| Model | PIPAL | TID2013 | CSIQ | LIVE |
+|---|---|---|---|---|
+| LPIPS-Alex| 0.739 / 0.541 / 0.773 | 0.782 / 0.583 / 0.813 | 0.940 / 0.784 / 0.946 | 0.949 / 0.794 / 0.943 |
+| LPIPS-VGG | 0.691 / 0.503 / 0.739 | 0.659 / 0.486 / 0.746 | 0.852 / 0.665 / 0.880 | 0.918 / 0.740 / 0.918 |
+| DISTS     | 0.616 / 0.437 / 0.657 | 0.692 / 0.513 / 0.771 | 0.919 / 0.750 / 0.931 | 0.954 / 0.804 / 0.949 |
+| DINOv3    | 0.705 / 0.504 / 0.709 | 0.711 / 0.524 / 0.769 | 0.810 / 0.626 / 0.874 | 0.897 / 0.721 / 0.913 |
+| CLIP      | 0.668 / 0.479 / 0.679 | 0.744 / 0.556 / 0.800 | 0.911 / 0.735 / 0.928 | 0.935 / 0.770 / 0.936 |
+| MAE       | 0.644 / 0.453 / 0.634 | 0.674 / 0.495 / 0.681 | 0.817 / 0.625 / 0.830 | 0.919 / 0.752 / 0.926 |
+| DreamSim  | 0.790 / 0.585 / 0.779 | 0.806 / 0.610 / 0.836 | 0.894 / 0.710 / 0.905 | 0.933 / 0.773 / 0.936 |
+
+## Citation
+
+The paper is under review; citation information will be added here.
+
+## License
+
+Apache 2.0 — see [LICENSE](LICENSE). The backbones (DINOv3, CLIP, MAE, DreamSim's ViTs) and the `lpips` / `DISTS-pytorch` / `dreamsim` packages keep their own upstream licenses.

fomo_iqa-0.1.0/fomo/__init__.py

@@ -0,0 +1,28 @@
+"""
+FoMo — perceptual image distance metrics trained on diffusion-generated
+(FLUX.1-dev) image pairs.
+
+    import fomo
+    metric = fomo.FoMo('dinov3')            # downloads + caches the weights
+    d = metric(img_ref, img_test)           # [B]; lower = more similar
+
+Models: lpips_alex, lpips_vgg, dists, dinov3, clip, mae, dreamsim.
+Inputs: float tensors [B, 3, H, W] in [0, 1], not pre-normalized.
+A model is defined by its backbone alone; the score presentation is chosen per
+call. symmetric (default True) averages both argument orders for the
+prediction-head models; output selects one of three presentations —
+
+    'raw'       (default) the model output; the protocol the published
+                numbers use. The prediction-head models (dinov3 / clip / mae)
+                score pairs with a ranking value whose zero point is arbitrary,
+                so it may be negative.
+    'distance'  non-negative with d(x, x) ~ 0, for drop-in LPIPS-style use;
+                ranks pairs identically to 'raw'.
+    'loss'      hinged at the identity level, for use as a perceptual loss.
+"""
+
+from fomo.hub import MODELS, FoMo, load
+from fomo.model import DistanceModel
+
+__version__ = '0.1.0'
+__all__ = ['FoMo', 'DistanceModel', 'load', 'MODELS', '__version__']