Update README.md

README.md

@@ -9,46 +9,9 @@ tags:
 library_name: pytorch
 ---
 
-# GUARD — GenImage-SD1.4 Controlled Detector Panel
+# GUARD-AI Checkpoints
 
-Checkpoints for the **GUARD** benchmark's controlled comparison: AI-generated-image
-detectors whose training fakes all come from a single, fixed generator —
-**Stable Diffusion v1.4**. This isolates the detection *method* from its training
-*data*: every detector below was trained (by us, or by its original authors) on
-SD1.4 output, so differences in benchmark performance reflect the architecture and
-training recipe, not the generator each detector happened to see.
-
-The panel covers **13 detector families** across all four GUARD taxonomy groups
-(spatial, frequency, hybrid/reconstruction, transformer/VLM) — all now final.
-
-## Checkpoints
-
-| File | Detector | Family | Provenance | Status |
-|---|---|---|---|---|
-| `wang2020/genimage_sd14_best.pth` | CNNDetection | Spatial | retrained by us, official GenImage SD1.4 split | ✅ final (val AUC 0.99997) |
-| `corvi2023/genimage_sd14_best.pth` | Corvi et al. | Spatial | retrained by us, official GenImage SD1.4 split | ✅ final (val AUC 1.0000) |
-| `npr2024/genimage_sd14_best.pth` | NPR | Spatial | retrained by us, official GenImage SD1.4 split | ✅ final (val AUC 1.0000) |
-| `cavia2024/genimage_sd14_best.pth` | LaDeDa (Cavia et al.) | Spatial | retrained by us, official GenImage SD1.4 split | ✅ final (val AUC 0.9999) |
-| `freqnet/genimage_sd14_best.pth` | FreqNet | Frequency | retrained by us, official GenImage SD1.4 split | ✅ final (val AUC 0.9996) |
-| `spai/genimage_sd14_best.pth` | SPAI | Frequency | retrained by us (frozen MFM ViT-B/16 backbone; released ckpt trains on a non-SD1.4 corpus — SD1.4/1.3 appear only as held-out test generators in the original paper) | ✅ final (val AUC 0.9998) |
-| `ojha2023/genimage_sd14_best.pth` | UniversalFakeDetect | Transformer/VLM | retrained by us, official GenImage SD1.4 split (retried after an initial GPU-memory conflict during the panel run) | ✅ final (val AUC 0.9973) |
-| `d3/genimage_sd14_best.pth` | D³ | Hybrid | retrained by us (released ckpt trains on a broader merged UFD+GenImage pool; we retrain SD1.4-only) | ✅ final (val AUC 1.0000) |
-| `chen2024_convnext/genimage_sd14_released.pth` | DRCT-ConvNeXt | Hybrid | authors' released checkpoint — SD1.4/1.5 diffusion-reconstruction training | ✅ released |
-| `chen2024_clip/genimage_sd14_released.pth` | DRCT-CLIP | Hybrid | authors' released checkpoint — SD1.4/1.5 diffusion-reconstruction training | ✅ released |
-| `aide/genimage_sd14_released.pth` | AIDE | Hybrid | authors' released checkpoint, trained on official SD1.4 split | ✅ released |
-| `effort/genimage_sd14_released.pth` | Effort | Transformer/VLM | authors' released checkpoint (`effort_clip_L14_trainOn_sdv14.pth`), trained on official SD1.4 split | ✅ released |
-| `c2pclip/genimage_sd14_released.pth` | C2P-CLIP | Transformer/VLM | authors' released checkpoint, GenImage/SD1.4 variant (not the ProGAN one) | ✅ released |
-| `rajan2025/genimage_sd14_released.pth` | AlignedForensics | Spatial | authors' released "Stay-Positive" checkpoint, SD1.4/LDM-reconstruction training | ✅ released |
-
-**All 13 detector families in the panel are now final.** 8 were retrained by us
-on the official GenImage SD1.4 split (wang2020, corvi2023, npr2024, cavia2024,
-freqnet, spai, ojha2023, d3); 5 are the original authors' own SD1.4-trained
-checkpoints, kept as released (aide, effort, c2pclip, rajan2025, DRCT). Only
-**SPAI** and **D³** required a fresh retrain to join the panel despite already
-having released checkpoints — SPAI's original paper trains its classifier head
-on a different (non-SD1.4) latent-diffusion corpus, with SD1.4/SD1.3 used solely
-as held-out zero-shot test generators in their own Table 1; D³'s released
-checkpoint trains on a broader merged UFD+GenImage pool rather than SD1.4 alone.
+Checkpoints used in the **GUARD-AI** benchmark.
 
 ## Checkpoint format
 
@@ -62,11 +25,3 @@ threshold  = ckpt["threshold"]  # Youden-J calibrated decision threshold (retrai
 All detectors output a single fake-logit; `sigmoid(logit) >= threshold` ⇒ "AI-generated".
 Architecture definitions live in each detector's original repository; the GUARD
 codebase vendors and wires them for evaluation.
-
-## Why SD1.4?
-
-GenImage (Zhu et al., NeurIPS 2023) identifies SD v1.4 as the training generator
-with the best cross-generator transfer, and the majority of GenImage-family released
-checkpoints (AIDE, Effort, C2P-CLIP, D³, DRCT, AlignedForensics) are already trained
-on SD1.4 data. Anchoring the retrained panel on the same split makes the detection
-*method* — not its training *data* — the object of comparison.