Upload folder using huggingface_hub

README.md

@@ -4,70 +4,181 @@ library_name: comfyui
 pipeline_tag: image-to-image
 tags:
   - hdr
+  - inverse-tone-mapping
+  - sdr-to-hdr
+  - vae-decoder
   - diffusion
   - comfyui
-  - vae-decoder
   - openexr
   - radiance
   - rudra
 ---
 
-# RUDRA — HDR Decoders for Diffusion Models
+# RUDRA
 
 **Radiometric Dynamic-Range Conditioning for HDR-Aware Diffusion Models**
-FXTD Studios / Radiance Research
+[FXTD Studios](https://fxtdstudios.com) · [github.com/fxtdstudios/RUDRA](https://github.com/fxtdstudios/RUDRA)
 
-Distilled decoders that turn diffusion **latents into scene-linear HDR / OpenEXR** instead
-of tone-mapped SDR. They replace the standard VAE decode inside the *Radiance HDR VAE
-Decode* ComfyUI node, preserving highlights, exposure, and wide-gamut color.
+Two independent families live in this repository.
+
+| family | what it does | files |
+|---|---|---|
+| **`sdr2hdr/`** | takes an ordinary 8-bit frame and reconstructs scene-linear HDR | 6 |
+| **decoders (root)** | decode diffusion latents straight to scene-linear HDR, skipping the tone-mapped VAE | 14 |
+
+They share a name and a radiometric convention and nothing else. Pick the one
+that matches your problem.
+
+---
 
-➡️ **Code, training, and docs:** https://github.com/fxtdstudios/RUDRA
+## `sdr2hdr/` — single-image inverse tone mapping
 
-## Files
+A compact U-Net that predicts a **residual on an analytic inverse-ACES tone
+map**, not the image itself. Six channels in (SDR RGB plus the analytic
+baseline), five out (a log-domain residual and two masks). The curve does the
+mapping; the network supplies only what a curve cannot know.
 
-Each file is a trained `RadianceTurboDecoder` / `RadianceFullDecoder` for one backbone:
+**Start with `sdr2hdr_shadow_v1.safetensors`.**
 
+| file | params | what it is |
+|---|---:|---|
+| **`sdr2hdr_shadow_v1`** | 1,217,318 | **the shipped model.** v5 backbone plus a trained shadow gate, seed 20260901 |
+| `sdr2hdr_shadow_s2` | 1,217,318 | same recipe, seed 2 |
+| `sdr2hdr_shadow_s3` | 1,217,318 | same recipe, seed 3 |
+| `sdr2hdr_image_v5` | 1,196,197 | the backbone alone, no gate |
+| `sdr2hdr_image_v6` | 4,770,117 | 4× capacity ablation, 64 base channels |
+| `sdr2hdr_temporal_v1` | 37,011 | temporal refiner. **Not an `SDR2HDRNet`** and will not load as one |
+
+Input is 8-bit normalised sRGB RGB. Output is scene-linear RGB normalised to
+10,000 nits, with diffuse white at 203 (ITU-R BT.2408).
+
+### Loading
+
+Each file carries its own architecture config in its safetensors metadata, so
+nothing else is needed to build the right network:
+
+```python
+import json, torch
+from safetensors import safe_open
+from safetensors.torch import load_file
+from rudra.sdr2hdr import SDR2HDRNet          # pip install -e . from the GitHub repo
+
+path = "sdr2hdr/sdr2hdr_shadow_v1.safetensors"
+with safe_open(path, framework="pt") as f:
+    config = json.loads(f.metadata()["config"])
+model = SDR2HDRNet.from_config(config)
+model.load_state_dict(load_file(path), strict=True)
+model.eval()
 ```
-rudra_{turbo|full}_decoder_{backbone}_ema.safetensors
+
+`SHA256SUMS` and `index.json` sit beside the weights.
+
+### What it scores
+
+429 held-out frames at native 1280×720, PU21-PSNR and ColorVideoVDP JOD,
+against the analytic inverse-ACES baseline the model sits on top of. `hard` is
+the deployment condition (unknown tone curve, 4:2:0 chroma, banding, JPEG);
+`clean` is well-graded input the baseline already handles.
+
+| model | clean dB | clean JOD | hard dB | hard JOD |
+|---|---:|---:|---:|---:|
+| shadow_v1 | +0.07 | +0.113 | +1.24 | +0.389 |
+| shadow_s2 | +0.71 | +0.068 | +0.96 | +0.308 |
+| shadow_s3 | +0.45 | +0.089 | +1.18 | +0.358 |
+| **mean ± sd** | **+0.41 ± 0.33** | **+0.090 ± 0.023** | **+1.12 ± 0.15** | **+0.352 ± 0.041** |
+| image_v5 | −3.00 | −0.046 | +1.43 | +0.443 |
+| image_v6 | −2.75 | +0.004 | +0.96 | +0.343 |
+
+The three gate seeds are the only configurations that come out positive on all
+four measures.
+
+### Limitations, stated plainly
+
+- **No published method has been scored on our split.** Every number above is
+  against our own analytic baseline. Read them as internal progress, not as a
+  comparison, until that gap is closed.
+- **The backbone alone regresses on clean input.** `image_v5` loses 3.0 dB of
+  PU21-PSNR on well-graded frames, and the failure concentrates in
+  low-dynamic-range scenes: the 60 worst average −10.76 dB at a median
+  ground-truth peak of 238 nits, the 60 best +3.41 dB at 19,590 nits. The gate
+  is what fixes it.
+- **The metrics disagree, and it matters.** That same 3.0 dB loss is −0.046 JOD
+  in CVVDP, far below a just-noticeable difference. Do not quote the PU21 row
+  without the JOD beside it.
+- **The achievable gain is bounded by the input, not the model.** A linear
+  readout of the features explains 3% of the variance in the per-frame scale an
+  oracle would use, and 79% of that variance sits within a condition rather than
+  between. Capacity (4×), corpus (6×) and objective were each varied and none
+  moved it.
+- **Three seeds is a spread, not a distribution.**
+- **The temporal refiner is unevaluated.** Its held-out set is 4 validation and
+  5 test clips of one scene each, too small to report.
+
+Training data is documented in the GitHub README: Poly Haven HDRIs (CC0),
+HdM-HDR-2014 and HdM-HFR-2017 (academic licence), Netflix Chimera (CC BY 4.0),
+plus proprietary FXTD footage that is not redistributed.
+
+---
+
+## Decoders — diffusion latents to scene-linear HDR
+
+Diffusion models are trained on tone-mapped images, so their VAEs decode to a
+world where nothing is brighter than white. These replace that decoder and
+return scene-linear radiance instead.
+
+Two sizes per backbone: `full` reconstructs more faithfully, `turbo` is roughly
+a tenth the size and fast enough to sit in a live graph.
+
+```bash
+huggingface-cli download fxtdstudios/RUDRA \
+    --include "rudra_*_decoder_*.safetensors" \
+    --local-dir "ComfyUI/models/radiance"
 ```
 
-| Backbone | Recommended file | Quality (PSNR_log) |
-|---|---|---|
-| Flux.1 | `rudra_full_decoder_flux_ema.safetensors` | 29.77 |
-| Wan | `rudra_full_decoder_wan_ema.safetensors` | 32.45 |
-| SDXL | `rudra_turbo_decoder_sdxl_ema.safetensors` | 33.86 |
-| Qwen-Image | `rudra_turbo_decoder_qwen_ema.safetensors` | 26.67 |
-| Flux.2 Klein | `rudra_turbo_decoder_flux2-klein_ema.safetensors` | 28.57 |
-| LTX (2.3) | `rudra_full_decoder_ltx-video_ema.safetensors` | 25.47 |
-| Z-Image | use the Flux decoder (shares the FLUX.1 VAE) | — |
-
-`turbo` (~0.5 M params) is fast and strong on simple latents (SDXL); `full` (~5.6 M) wins
-on Flux/Wan/LTX. Both are provided where trained.
-
-## Usage (ComfyUI)
-
-1. Download into `ComfyUI/models/radiance/`:
-   ```bash
-   huggingface-cli download fxtdstudios/RUDRA --include "rudra_*.safetensors" \
-       --local-dir "ComfyUI/models/radiance"
-   ```
-2. In the **Radiance HDR VAE Decode** node: set `rudra_decoder = Enabled`, pick
-   `decoder_size` (`rudra_turbo` or `rudra_full`) per the table above, and set
-   `target_space` to your output color space (Linear / ACEScg / Rec.2020 / LogC4…).
-
-## Notes
-
-- **Backbone-specific:** a decoder is tied to its VAE latent space — use the matching file
-  for the model feeding the node (Flux decoder for a Flux workflow, etc.).
-- **Flux.2 Klein** uses a 128-channel / 16× VAE, so its decoder has an extra upsample stage
-  (requires the updated `fast_vae.py` from the GitHub repo).
-- Quality is reported as held-out **log-space PSNR**; perceptual HDR evaluation uses
-  **ColorVideoVDP** (JOD). SDXL/Qwen/Klein were trained on a smaller pair set and can be
-  improved with more data.
-
-## Citation
-
-> RUDRA: Radiometric Dynamic-Range Conditioning for HDR-Aware Diffusion Models.
-> FXTD Studios / Radiance Research.
-
-License: change the `license:` field above to match your release t
+Then enable `rudra_decoder` in the **Radiance HDR VAE Decode** node and pick a
+`decoder_size`.
+
+| file | size |
+|---|---:|
+| `rudra_full_decoder_flux_ema.safetensors` | 21.5 MB |
+| `rudra_full_decoder_ltx-video_ema.safetensors` | 22.0 MB |
+| `rudra_full_decoder_sdxl_ema.safetensors` | 21.4 MB |
+| `rudra_full_decoder_wan_ema.safetensors` | 21.5 MB |
+| `rudra_full_decoder_zimage_ema.safetensors` | 21.5 MB |
+| `rudra_turbo_decoder_flux_ema.safetensors` | 2.2 MB |
+| `rudra_turbo_decoder_flux2_ema.safetensors` | 2.0 MB |
+| `rudra_turbo_decoder_flux2-klein_ema.safetensors` | 3.0 MB |
+| `rudra_turbo_decoder_ltx_ema.safetensors` | 2.4 MB |
+| `rudra_turbo_decoder_ltx-video_ema.safetensors` | 3.5 MB |
+| `rudra_turbo_decoder_qwen_ema.safetensors` | 2.2 MB |
+| `rudra_turbo_decoder_sdxl_ema.safetensors` | 2.1 MB |
+| `rudra_turbo_decoder_wan_ema.safetensors` | 2.2 MB |
+| `rudra_turbo_decoder_zimage_ema.safetensors` | 2.2 MB |
+
+Measured reconstruction quality, for the size we recommend per backbone. The
+table covers the recommended decoder per backbone, not every file above:
+
+| backbone | VAE latent | recommended | PSNR_log |
+|---|---|---|---:|
+| Flux.1 | 16ch / 8× | full | 29.77 |
+| Wan | 16ch / 8× | full | 32.45 |
+| LTX | 128ch / 8× | full | 25.47 |
+| SDXL | 4ch / 8× | turbo | 33.86 |
+| Qwen-Image | 16ch / 8× | turbo | 26.67 |
+| Flux.2 Klein | 128ch / **16×** | turbo | 28.57 |
+
+---
+
+## Licence and citation
+
+Apache 2.0. If the SDR→HDR work is useful in yours:
+
+```bibtex
+@misc{rudra2026,
+  title  = {What an 8-Bit Frame Can and Cannot Say About the Scene Behind It:
+            bounds for inverse tone mapping, and a gate that reaches one of them},
+  author = {{FXTD Studios}},
+  year   = {2026},
+  note   = {https://github.com/fxtdstudios/RUDRA}
+}
+```

sdr2hdr/SHA256SUMS

@@ -0,0 +1,6 @@
+a25ac506dd5cb2f20f1f428628c9bb4c6d09bfd5aed8f8b93d783314f98bd47c  sdr2hdr_shadow_v1.safetensors
+f46b1a367ae9aca6df9d256243189462b7c504fdda9bb089bfe9588e3c213a87  sdr2hdr_shadow_s2.safetensors
+46ab0e49bd246ef4e9e1394f3f736b5274afc04e24efd6afff994cf1b1d7cc7e  sdr2hdr_shadow_s3.safetensors
+3e0581ce5f4c20c53220955ce23af394a7ad5a091755c2ee5b72e9a02ecca715  sdr2hdr_image_v5.safetensors
+44c08cc202e4e0a3f7c59b13b0b25cec029116ec1b7d79102d07e269a5d44b90  sdr2hdr_image_v6.safetensors
+9ee451688b46ce2f11948d451dfe32f1816448c876231d7299eb29031aeb060c  sdr2hdr_temporal_v1.safetensors

sdr2hdr/index.json

@@ -0,0 +1,44 @@
+[
+  {
+    "file": "sdr2hdr/sdr2hdr_shadow_v1.safetensors",
+    "note": "the shipped model: v5 backbone + trained shadow gate",
+    "params": 1217318,
+    "step": 2400,
+    "sha256": "a25ac506dd5cb2f20f1f428628c9bb4c6d09bfd5aed8f8b93d783314f98bd47c"
+  },
+  {
+    "file": "sdr2hdr/sdr2hdr_shadow_s2.safetensors",
+    "note": "same recipe, seed 2",
+    "params": 1217318,
+    "step": 2600,
+    "sha256": "f46b1a367ae9aca6df9d256243189462b7c504fdda9bb089bfe9588e3c213a87"
+  },
+  {
+    "file": "sdr2hdr/sdr2hdr_shadow_s3.safetensors",
+    "note": "same recipe, seed 3",
+    "params": 1217318,
+    "step": 2800,
+    "sha256": "46ab0e49bd246ef4e9e1394f3f736b5274afc04e24efd6afff994cf1b1d7cc7e"
+  },
+  {
+    "file": "sdr2hdr/sdr2hdr_image_v5.safetensors",
+    "note": "the backbone alone, no gate",
+    "params": 1196197,
+    "step": 81000,
+    "sha256": "3e0581ce5f4c20c53220955ce23af394a7ad5a091755c2ee5b72e9a02ecca715"
+  },
+  {
+    "file": "sdr2hdr/sdr2hdr_image_v6.safetensors",
+    "note": "4x capacity ablation, 64 base channels",
+    "params": 4770117,
+    "step": 83000,
+    "sha256": "44c08cc202e4e0a3f7c59b13b0b25cec029116ec1b7d79102d07e269a5d44b90"
+  },
+  {
+    "file": "sdr2hdr/sdr2hdr_temporal_v1.safetensors",
+    "note": "temporal refiner -- NOT an SDR2HDRNet",
+    "params": 37011,
+    "step": 0,
+    "sha256": "9ee451688b46ce2f11948d451dfe32f1816448c876231d7299eb29031aeb060c"
+  }
+]

sdr2hdr/sdr2hdr_image_v5.config.json

@@ -0,0 +1,41 @@
+{
+  "mode": "image",
+  "manifest": "E:\\RUDRA_v3_20260822\\sdr_hdr_manifest.jsonl",
+  "output_dir": "E:\\RUDRA_v3_20260822\\checkpoints\\sdr2hdr_image_v5",
+  "image_checkpoint": null,
+  "best_eval": "hard",
+  "resume": "E:\\RUDRA_v3_20260822\\checkpoints\\sdr2hdr_image_v5\\step_0020000.pt",
+  "init_checkpoint": null,
+  "reset_best": false,
+  "device": "cuda",
+  "steps": 100000,
+  "batch_size": 2,
+  "crop_size": 384,
+  "workers": 2,
+  "base_channels": 32,
+  "temporal_channels": 24,
+  "lr": 0.0002,
+  "weight_decay": 0.0001,
+  "grad_accum": 2,
+  "max_grad_norm": 1.0,
+  "augmentation_strength": 1.0,
+  "degradation_probability": 0.65,
+  "video_sample_fraction": 0.4,
+  "scene_balanced_sampling": true,
+  "temporal_weight": 0.5,
+  "shadow_chroma_weight": 0.15,
+  "shadow_smoothness_weight": 0.02,
+  "best_metric": "composite_gain",
+  "val_fraction": 0.1,
+  "eval_every": 1000,
+  "eval_batches": 128,
+  "save_every": 2000,
+  "log_every": 20,
+  "max_items": null,
+  "max_val_items": null,
+  "seed": 20260715,
+  "manifest_sha256": "2fc33f59704351099e6941ea40fa50f83674a630ee779729ad870455e52ef5f7",
+  "architecture": "SDR2HDRNet",
+  "input_contract": "8-bit normalized sRGB RGB",
+  "output_contract": "scene-linear RGB normalized to 10000 nits"
+}

sdr2hdr/sdr2hdr_image_v5.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e0581ce5f4c20c53220955ce23af394a7ad5a091755c2ee5b72e9a02ecca715
+size 4794076

sdr2hdr/sdr2hdr_image_v6.config.json

@@ -0,0 +1,41 @@
+{
+  "mode": "image",
+  "manifest": "E:\\RUDRA_v3_20260822\\sdr_hdr_manifest.jsonl",
+  "output_dir": "E:\\RUDRA_v3_20260822\\checkpoints\\sdr2hdr_image_v6_c64",
+  "image_checkpoint": null,
+  "best_eval": "hard",
+  "resume": null,
+  "init_checkpoint": null,
+  "reset_best": false,
+  "device": "cuda",
+  "steps": 100000,
+  "batch_size": 2,
+  "crop_size": 384,
+  "workers": 2,
+  "base_channels": 64,
+  "temporal_channels": 24,
+  "lr": 0.0002,
+  "weight_decay": 0.0001,
+  "grad_accum": 2,
+  "max_grad_norm": 1.0,
+  "augmentation_strength": 1.0,
+  "degradation_probability": 0.65,
+  "video_sample_fraction": 0.4,
+  "scene_balanced_sampling": true,
+  "temporal_weight": 0.5,
+  "shadow_chroma_weight": 0.15,
+  "shadow_smoothness_weight": 0.02,
+  "best_metric": "composite_gain",
+  "val_fraction": 0.1,
+  "eval_every": 1000,
+  "eval_batches": 128,
+  "save_every": 2000,
+  "log_every": 20,
+  "max_items": null,
+  "max_val_items": null,
+  "seed": 20260715,
+  "manifest_sha256": "2fc33f59704351099e6941ea40fa50f83674a630ee779729ad870455e52ef5f7",
+  "architecture": "SDR2HDRNet",
+  "input_contract": "8-bit normalized sRGB RGB",
+  "output_contract": "scene-linear RGB normalized to 10000 nits"
+}

sdr2hdr/sdr2hdr_image_v6.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44c08cc202e4e0a3f7c59b13b0b25cec029116ec1b7d79102d07e269a5d44b90
+size 19089804

sdr2hdr/sdr2hdr_shadow_s2.config.json

@@ -0,0 +1,19 @@
+{
+  "manifest": "E:\\RUDRA_v3_20260822\\sdr_hdr_manifest.jsonl",
+  "init_checkpoint": "E:\\RUDRA_v3_20260822\\checkpoints\\sdr2hdr_image_v5\\shipped_v5_step81000.pt",
+  "output_dir": "E:\\RUDRA_v3_20260822\\checkpoints\\sdr2hdr_shadow_s2",
+  "steps": 3000,
+  "lr": 0.003,
+  "grad_accum": 8,
+  "max_side": 512,
+  "degradation_probability": 0.5,
+  "eval_every": 200,
+  "eval_batches": 128,
+  "best_smoothing": 5,
+  "workers": 2,
+  "seed": 2,
+  "device": "cuda",
+  "base_channels": 32,
+  "shadow_conditioning": true,
+  "objective": "degradation_bce"
+}

sdr2hdr/sdr2hdr_shadow_s2.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f46b1a367ae9aca6df9d256243189462b7c504fdda9bb089bfe9588e3c213a87
+size 4878384

sdr2hdr/sdr2hdr_shadow_s3.config.json

@@ -0,0 +1,19 @@
+{
+  "manifest": "E:\\RUDRA_v3_20260822\\sdr_hdr_manifest.jsonl",
+  "init_checkpoint": "E:\\RUDRA_v3_20260822\\checkpoints\\sdr2hdr_image_v5\\shipped_v5_step81000.pt",
+  "output_dir": "E:\\RUDRA_v3_20260822\\checkpoints\\sdr2hdr_shadow_s3",
+  "steps": 3000,
+  "lr": 0.003,
+  "grad_accum": 8,
+  "max_side": 512,
+  "degradation_probability": 0.5,
+  "eval_every": 200,
+  "eval_batches": 128,
+  "best_smoothing": 5,
+  "workers": 2,
+  "seed": 3,
+  "device": "cuda",
+  "base_channels": 32,
+  "shadow_conditioning": true,
+  "objective": "degradation_bce"
+}

sdr2hdr/sdr2hdr_shadow_s3.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46ab0e49bd246ef4e9e1394f3f736b5274afc04e24efd6afff994cf1b1d7cc7e
+size 4878384

sdr2hdr/sdr2hdr_shadow_v1.config.json

@@ -0,0 +1,19 @@
+{
+  "manifest": "E:\\RUDRA_v3_20260822\\sdr_hdr_manifest.jsonl",
+  "init_checkpoint": "E:\\RUDRA_v3_20260822\\checkpoints\\sdr2hdr_image_v5\\shipped_v5_step81000.pt",
+  "output_dir": "E:\\RUDRA_v3_20260822\\checkpoints\\sdr2hdr_shadow_v1",
+  "steps": 3000,
+  "lr": 0.003,
+  "grad_accum": 8,
+  "max_side": 512,
+  "degradation_probability": 0.5,
+  "eval_every": 200,
+  "eval_batches": 128,
+  "best_smoothing": 5,
+  "workers": 2,
+  "seed": 20260901,
+  "device": "cuda",
+  "base_channels": 32,
+  "shadow_conditioning": true,
+  "objective": "degradation_bce"
+}

sdr2hdr/sdr2hdr_shadow_v1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a25ac506dd5cb2f20f1f428628c9bb4c6d09bfd5aed8f8b93d783314f98bd47c
+size 4878392

sdr2hdr/sdr2hdr_temporal_v1.config.json

@@ -0,0 +1,40 @@
+{
+  "mode": "temporal",
+  "manifest": "E:\\RUDRA_postfix_20260818\\hdrdata\\video_manifest_9f_unc.jsonl",
+  "output_dir": "E:\\RUDRA_postfix_20260818\\hdrdata\\checkpoints\\sdr2hdr_temporal_postfix_v1",
+  "image_checkpoint": "E:\\RUDRA_postfix_20260818\\hdrdata\\checkpoints\\sdr2hdr_image_postfix_v1\\best.pt",
+  "resume": "E:\\RUDRA_postfix_20260818\\hdrdata\\checkpoints\\sdr2hdr_temporal_postfix_v1\\step_0016000.pt",
+  "init_checkpoint": null,
+  "reset_best": false,
+  "device": "cuda",
+  "steps": 30000,
+  "batch_size": 2,
+  "crop_size": 256,
+  "workers": 2,
+  "base_channels": 32,
+  "temporal_channels": 24,
+  "lr": 0.0002,
+  "weight_decay": 0.0001,
+  "grad_accum": 1,
+  "max_grad_norm": 1.0,
+  "augmentation_strength": 1.0,
+  "degradation_probability": 0.65,
+  "video_sample_fraction": 0.25,
+  "scene_balanced_sampling": true,
+  "temporal_weight": 0.5,
+  "shadow_chroma_weight": 0.15,
+  "shadow_smoothness_weight": 0.02,
+  "best_metric": "loss",
+  "val_fraction": 0.1,
+  "eval_every": 500,
+  "eval_batches": 16,
+  "save_every": 2000,
+  "log_every": 20,
+  "max_items": null,
+  "max_val_items": null,
+  "seed": 20260715,
+  "manifest_sha256": "b4a1791294282b912662b6d7293f04ec2cbe561eb79ffa72b4737f06812808bd",
+  "architecture": "TemporalHDRRefiner",
+  "input_contract": "8-bit normalized sRGB RGB",
+  "output_contract": "scene-linear RGB normalized to 10000 nits"
+}

sdr2hdr/sdr2hdr_temporal_v1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ee451688b46ce2f11948d451dfe32f1816448c876231d7299eb29031aeb060c
+size 150196