Update model card, app.json (patch size + VRAM plan) and inference configs

CBCT/app.json

@@ -4,37 +4,72 @@
     "description": "<b>Description:</b><br>Supervised whole-body CBCT-to-sCT model trained using SynthRAD 2023 and SynthRAD 2025 Task 2 datasets, distributed as five validation-fold checkpoints (<code>CV_0.pt</code> to <code>CV_4.pt</code>). Training pairs are aligned with IMPACTReg. The available Task 2 data include 291 patients from SynthRAD 2023 and 638 patients from SynthRAD 2025, for 929 patients in total.<br><br><b>Architecture and loss:</b><br>2.5D UNet++ with a ResNet34 encoder, trained with KonfAI and optimized using IMPACT-Synth, a perceptual loss leveraging semantic priors from <b>SAM 2.1-s</b>.<br><br><b>Training data:</b><br>Whole-body paired CT-CBCT volumes come from SynthRAD 2023 Task 2 (<code>brain</code>, <code>pelvis</code>) and SynthRAD 2025 Task 2 (<code>AB</code>, <code>HN</code>, <code>TH</code>). Additional OOD evaluation cohorts are available for <code>AB_ood</code>, <code>HN_ood</code>, and <code>TH_ood</code>. Pairs are aligned with IMPACTReg, and the associated transform files are available at <a href=\"https://huggingface.co/datasets/VBoussot/synthrad2023-impact-registration\">synthrad2023-impact-registration</a> and <a href=\"https://huggingface.co/datasets/VBoussot/synthrad2025-impact-registration\">synthrad2025-impact-registration</a>. The files <code>CrossValidation_0.txt</code> to <code>CrossValidation_4.txt</code> correspond to validation folds for the distributed checkpoints, and <code>Validation.txt</code> contains 150 release-validation patients.<br><br><b>Release performance (CV overall, n=150):</b><br>Dice 0.758, MAE 53.55 HU, SSIM 0.939, PSNR 32.09 dB, SAM 13.70, Reg 1.04, Uncertainty 784.24.<br><br><b>Inference note:</b><br>The prediction pipeline supports test-time augmentation and writes an averaged <code>InferenceStack</code> output that can also be used for uncertainty estimation.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
     "tta": 2,
     "mc_dropout": false,
-    "models": ["CV_0.pt", "CV_1.pt", "CV_2.pt", "CV_3.pt", "CV_4.pt"],
+    "models": [
+        "CV_0.pt",
+        "CV_1.pt",
+        "CV_2.pt",
+        "CV_3.pt",
+        "CV_4.pt"
+    ],
     "inputs": {
-      "CBCT": {
-        "display_name": "CBCT",
-        "volume_type": "VOLUME",
-        "required": true
-      }
+        "CBCT": {
+            "display_name": "CBCT",
+            "volume_type": "VOLUME",
+            "required": true
+        }
     },
-
     "outputs": {
-      "sCT": {
-        "display_name": "sCT",
-        "volume_type": "VOLUME",
-        "required": true
-      }
-    },
-
-    "inputs_evaluations": {
-      "Image": {
-        "Evaluation.yml": {
-          "sCT": {
+        "sCT": {
             "display_name": "sCT",
             "volume_type": "VOLUME",
             "required": true
-          },
-          "CT": {
-            "display_name": "CT",
-            "volume_type": "VOLUME",
-            "required": true
-          }
         }
-      }
+    },
+    "inputs_evaluations": {
+        "Image": {
+            "Evaluation.yml": {
+                "sCT": {
+                    "display_name": "sCT",
+                    "volume_type": "VOLUME",
+                    "required": true
+                },
+                "CT": {
+                    "display_name": "CT",
+                    "volume_type": "VOLUME",
+                    "required": true
+                }
+            }
+        }
+    },
+    "patch_size": [
+        1,
+        512,
+        512
+    ],
+    "vram_plan": {
+        "8": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 16
+        },
+        "16": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 28
+        },
+        "24": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 48
+        }
     }
 }

Finetune/app.json

@@ -4,37 +4,72 @@
     "description": "<b>Description:</b><br>Supervised whole-body CBCT-to-sCT model trained using SynthRAD 2023 and SynthRAD 2025 Task 2 datasets, distributed as five validation-fold checkpoints (<code>CV_0.pt</code> to <code>CV_4.pt</code>). Training pairs are aligned with IMPACTReg. The available Task 2 data include 291 patients from SynthRAD 2023 and 638 patients from SynthRAD 2025, for 929 patients in total.<br><br><b>Architecture and loss:</b><br>2.5D UNet++ with a ResNet34 encoder, trained with KonfAI and optimized using IMPACT-Synth, a perceptual loss leveraging semantic priors from <b>SAM 2.1-s</b>.<br><br><b>Training data:</b><br>Whole-body paired CT-CBCT volumes come from SynthRAD 2023 Task 2 (<code>brain</code>, <code>pelvis</code>) and SynthRAD 2025 Task 2 (<code>AB</code>, <code>HN</code>, <code>TH</code>). Additional OOD evaluation cohorts are available for <code>AB_ood</code>, <code>HN_ood</code>, and <code>TH_ood</code>. Pairs are aligned with IMPACTReg, and the associated transform files are available at <a href=\"https://huggingface.co/datasets/VBoussot/synthrad2023-impact-registration\">synthrad2023-impact-registration</a> and <a href=\"https://huggingface.co/datasets/VBoussot/synthrad2025-impact-registration\">synthrad2025-impact-registration</a>. The files <code>CrossValidation_0.txt</code> to <code>CrossValidation_4.txt</code> correspond to validation folds for the distributed checkpoints, and <code>Validation.txt</code> contains 150 release-validation patients.<br><br><b>Release performance (CV overall, n=150):</b><br>Dice 0.758, MAE 53.55 HU, SSIM 0.939, PSNR 32.09 dB, SAM 13.70, Reg 1.04, Uncertainty 784.24.<br><br><b>Inference note:</b><br>The prediction pipeline supports test-time augmentation and writes an averaged <code>InferenceStack</code> output that can also be used for uncertainty estimation.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
     "tta": 2,
     "mc_dropout": false,
-    "models": ["CV_0.pt", "CV_1.pt", "CV_2.pt", "CV_3.pt", "CV_4.pt"],
+    "models": [
+        "CV_0.pt",
+        "CV_1.pt",
+        "CV_2.pt",
+        "CV_3.pt",
+        "CV_4.pt"
+    ],
     "inputs": {
-      "CBCT": {
-        "display_name": "CBCT",
-        "volume_type": "VOLUME",
-        "required": true
-      }
+        "CBCT": {
+            "display_name": "CBCT",
+            "volume_type": "VOLUME",
+            "required": true
+        }
     },
-
     "outputs": {
-      "sCT": {
-        "display_name": "sCT",
-        "volume_type": "VOLUME",
-        "required": true
-      }
-    },
-
-    "inputs_evaluations": {
-      "Image": {
-        "Evaluation.yml": {
-          "sCT": {
+        "sCT": {
             "display_name": "sCT",
             "volume_type": "VOLUME",
             "required": true
-          },
-          "CT": {
-            "display_name": "CT",
-            "volume_type": "VOLUME",
-            "required": true
-          }
         }
-      }
+    },
+    "inputs_evaluations": {
+        "Image": {
+            "Evaluation.yml": {
+                "sCT": {
+                    "display_name": "sCT",
+                    "volume_type": "VOLUME",
+                    "required": true
+                },
+                "CT": {
+                    "display_name": "CT",
+                    "volume_type": "VOLUME",
+                    "required": true
+                }
+            }
+        }
+    },
+    "patch_size": [
+        1,
+        512,
+        512
+    ],
+    "vram_plan": {
+        "8": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 16
+        },
+        "16": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 28
+        },
+        "24": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 48
+        }
     }
 }

MR/app.json

@@ -1,38 +1,75 @@
 {
-  "display_name": "Synthesis: MR",
-  "short_description": "Supervised whole-body MR-to-sCT model for SynthRAD Task 1.<br><br><b>Training data:</b><br>805 paired CT-MR cases including OOD cohorts.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
-  "description": "<b>Description:</b><br>Supervised whole-body MR-to-sCT model trained using SynthRAD 2023 and SynthRAD 2025 Task 1 datasets, distributed as five validation-fold checkpoints (<code>CV_0.pt</code> to <code>CV_4.pt</code>). Training pairs are aligned with IMPACTReg. The available Task 1 data include 317 patients from SynthRAD 2023, 410 patients from SynthRAD 2025, and 78 additional OOD paired CT-MR cases, for 805 training patients in total.<br><br><b>Architecture and loss:</b><br>2.5D UNet++ with a ResNet34 encoder, trained with KonfAI and optimized using IMPACT-Synth, a perceptual loss leveraging semantic priors from <b>SAM 2.1-s</b>.<br><br><b>Training data:</b><br>Whole-body paired CT-MR volumes come from SynthRAD 2023 Task 1 (<code>brain</code>, <code>pelvis</code>), SynthRAD 2025 Task 1 (<code>AB</code>, <code>HN</code>, <code>TH</code>), and additional OOD cohorts including <code>CSIRO</code>, <code>GA</code>, <code>MR_Linac</code>, and <code>MR_Rennes</code>. Pairs are aligned with IMPACTReg, and the associated transform files are available at <a href=\"https://huggingface.co/datasets/VBoussot/synthrad2023-impact-registration\">synthrad2023-impact-registration</a> and <a href=\"https://huggingface.co/datasets/VBoussot/synthrad2025-impact-registration\">synthrad2025-impact-registration</a>. The files <code>CrossValidation_0.txt</code> to <code>CrossValidation_4.txt</code> correspond to validation folds for the distributed checkpoints, and <code>Validation.txt</code> contains 140 release-validation patients.<br><br><b>Release performance (CV overall, n=140):</b><br>Dice 0.698, MAE 67.49 HU, SSIM 0.959, PSNR 29.28 dB, SAM 21.98, Reg 0.622.<br><br><b>Inference note:</b><br>A body mask is estimated automatically at inference time and used to crop valid anatomy support before writing the output volume.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
-  "tta": 2,
-  "mc_dropout": false,
-  "models": ["CV_0.pt", "CV_1.pt", "CV_2.pt", "CV_3.pt", "CV_4.pt"],
-  "inputs": {
-    "MR": {
-      "display_name": "MR",
-      "volume_type": "VOLUME",
-      "required": true
-    }
-  },
-  "outputs": {
-    "sCT": {
-      "display_name": "sCT",
-      "volume_type": "VOLUME",
-      "required": true
-    }
-  },
-  "inputs_evaluations": {
-    "Image": {
-      "Evaluation.yml": {
+    "display_name": "Synthesis: MR",
+    "short_description": "Supervised whole-body MR-to-sCT model for SynthRAD Task 1.<br><br><b>Training data:</b><br>805 paired CT-MR cases including OOD cohorts.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
+    "description": "<b>Description:</b><br>Supervised whole-body MR-to-sCT model trained using SynthRAD 2023 and SynthRAD 2025 Task 1 datasets, distributed as five validation-fold checkpoints (<code>CV_0.pt</code> to <code>CV_4.pt</code>). Training pairs are aligned with IMPACTReg. The available Task 1 data include 317 patients from SynthRAD 2023, 410 patients from SynthRAD 2025, and 78 additional OOD paired CT-MR cases, for 805 training patients in total.<br><br><b>Architecture and loss:</b><br>2.5D UNet++ with a ResNet34 encoder, trained with KonfAI and optimized using IMPACT-Synth, a perceptual loss leveraging semantic priors from <b>SAM 2.1-s</b>.<br><br><b>Training data:</b><br>Whole-body paired CT-MR volumes come from SynthRAD 2023 Task 1 (<code>brain</code>, <code>pelvis</code>), SynthRAD 2025 Task 1 (<code>AB</code>, <code>HN</code>, <code>TH</code>), and additional OOD cohorts including <code>CSIRO</code>, <code>GA</code>, <code>MR_Linac</code>, and <code>MR_Rennes</code>. Pairs are aligned with IMPACTReg, and the associated transform files are available at <a href=\"https://huggingface.co/datasets/VBoussot/synthrad2023-impact-registration\">synthrad2023-impact-registration</a> and <a href=\"https://huggingface.co/datasets/VBoussot/synthrad2025-impact-registration\">synthrad2025-impact-registration</a>. The files <code>CrossValidation_0.txt</code> to <code>CrossValidation_4.txt</code> correspond to validation folds for the distributed checkpoints, and <code>Validation.txt</code> contains 140 release-validation patients.<br><br><b>Release performance (CV overall, n=140):</b><br>Dice 0.698, MAE 67.49 HU, SSIM 0.959, PSNR 29.28 dB, SAM 21.98, Reg 0.622.<br><br><b>Inference note:</b><br>A body mask is estimated automatically at inference time and used to crop valid anatomy support before writing the output volume.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
+    "tta": 2,
+    "mc_dropout": false,
+    "models": [
+        "CV_0.pt",
+        "CV_1.pt",
+        "CV_2.pt",
+        "CV_3.pt",
+        "CV_4.pt"
+    ],
+    "inputs": {
+        "MR": {
+            "display_name": "MR",
+            "volume_type": "VOLUME",
+            "required": true
+        }
+    },
+    "outputs": {
         "sCT": {
-          "display_name": "sCT",
-          "volume_type": "VOLUME",
-          "required": true
+            "display_name": "sCT",
+            "volume_type": "VOLUME",
+            "required": true
+        }
+    },
+    "inputs_evaluations": {
+        "Image": {
+            "Evaluation.yml": {
+                "sCT": {
+                    "display_name": "sCT",
+                    "volume_type": "VOLUME",
+                    "required": true
+                },
+                "CT": {
+                    "display_name": "CT",
+                    "volume_type": "VOLUME",
+                    "required": true
+                }
+            }
+        }
+    },
+    "patch_size": [
+        1,
+        512,
+        512
+    ],
+    "vram_plan": {
+        "8": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 16
+        },
+        "16": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 28
         },
-        "CT": {
-          "display_name": "CT",
-          "volume_type": "VOLUME",
-          "required": true
+        "24": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 48
         }
-      }
     }
-  }
 }

MR_CBCT/app.json

@@ -1,38 +1,75 @@
 {
-  "display_name": "Synthesis: MR/CBCT",
-  "short_description": "Supervised whole-body MR/CBCT-to-sCT model.<br><br><b>Training data:</b><br>1734 paired cases from Task 1 + Task 2.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
-  "description": "<b>Description:</b><br>Supervised whole-body synthesis model distributed as a single checkpoint (<code>CV_0.pt</code>) for MR/CBCT-style inputs. The model follows the same 2.5D synthesis setup used in the ImpactSynth family and produces an sCT volume from a single input volume after automatic body-mask estimation and intensity normalization.<br><br><b>Architecture and loss:</b><br>2.5D UNet++ with a ResNet34 encoder, trained with KonfAI and optimized using IMPACT-Synth, a perceptual loss leveraging semantic priors from <b>SAM 2.1-s</b>.<br><br><b>Training data:</b><br>The model is trained on 1734 IMPACTReg-aligned paired synthesis cases spanning both Task 1 and Task 2 cohorts, combining MR and CBCT supervision in a unified synthesis setup.<br><br><b>Input handling:</b><br>The prediction pipeline accepts a single input volume exposed here as <code>MR or CBCT</code>, estimates a body mask automatically, resamples to the working resolution, applies normalization, and writes an averaged sCT output using test-time augmentation.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
-  "tta": 2,
-  "mc_dropout": false,
-  "models": ["CV_0.pt", "CV_1.pt", "CV_2.pt", "CV_3.pt", "CV_4.pt"],
-  "inputs": {
-    "MR": {
-      "display_name": "MR or CBCT",
-      "volume_type": "VOLUME",
-      "required": true
-    }
-  },
-  "outputs": {
-    "sCT": {
-      "display_name": "sCT",
-      "volume_type": "VOLUME",
-      "required": true
-    }
-  },
-  "inputs_evaluations": {
-    "Image": {
-      "Evaluation.yml": {
+    "display_name": "Synthesis: MR/CBCT",
+    "short_description": "Supervised whole-body MR/CBCT-to-sCT model.<br><br><b>Training data:</b><br>1734 paired cases from Task 1 + Task 2.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
+    "description": "<b>Description:</b><br>Supervised whole-body synthesis model distributed as a single checkpoint (<code>CV_0.pt</code>) for MR/CBCT-style inputs. The model follows the same 2.5D synthesis setup used in the ImpactSynth family and produces an sCT volume from a single input volume after automatic body-mask estimation and intensity normalization.<br><br><b>Architecture and loss:</b><br>2.5D UNet++ with a ResNet34 encoder, trained with KonfAI and optimized using IMPACT-Synth, a perceptual loss leveraging semantic priors from <b>SAM 2.1-s</b>.<br><br><b>Training data:</b><br>The model is trained on 1734 IMPACTReg-aligned paired synthesis cases spanning both Task 1 and Task 2 cohorts, combining MR and CBCT supervision in a unified synthesis setup.<br><br><b>Input handling:</b><br>The prediction pipeline accepts a single input volume exposed here as <code>MR or CBCT</code>, estimates a body mask automatically, resamples to the working resolution, applies normalization, and writes an averaged sCT output using test-time augmentation.<br><br><b>How to cite:</b><br><cite>V. Boussot et al., <i>Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration</i>, arXiv preprint arXiv:2510.21358, <a href=\"https://doi.org/10.48550/arXiv.2510.21358\">https://doi.org/10.48550/arXiv.2510.21358</a>.</cite>",
+    "tta": 2,
+    "mc_dropout": false,
+    "models": [
+        "CV_0.pt",
+        "CV_1.pt",
+        "CV_2.pt",
+        "CV_3.pt",
+        "CV_4.pt"
+    ],
+    "inputs": {
+        "MR": {
+            "display_name": "MR or CBCT",
+            "volume_type": "VOLUME",
+            "required": true
+        }
+    },
+    "outputs": {
         "sCT": {
-          "display_name": "sCT",
-          "volume_type": "VOLUME",
-          "required": true
+            "display_name": "sCT",
+            "volume_type": "VOLUME",
+            "required": true
+        }
+    },
+    "inputs_evaluations": {
+        "Image": {
+            "Evaluation.yml": {
+                "sCT": {
+                    "display_name": "sCT",
+                    "volume_type": "VOLUME",
+                    "required": true
+                },
+                "CT": {
+                    "display_name": "CT",
+                    "volume_type": "VOLUME",
+                    "required": true
+                }
+            }
+        }
+    },
+    "patch_size": [
+        1,
+        512,
+        512
+    ],
+    "vram_plan": {
+        "8": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 16
+        },
+        "16": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 28
         },
-        "CT": {
-          "display_name": "CT",
-          "volume_type": "VOLUME",
-          "required": true
+        "24": {
+            "patch_size": [
+                1,
+                512,
+                512
+            ],
+            "batch_size": 48
         }
-      }
     }
-  }
 }

README.md

@@ -1,3 +1,58 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+library_name: konfai
+pipeline_tag: image-to-image
+tags:
+- medical-imaging
+- synthetic-ct
+- image-synthesis
+- radiotherapy
+- mri
+- cbct
+- konfai
+---
+
+# ImpactSynth — Synthetic CT (sCT) from MR / CBCT
+
+Whole-body **synthetic CT** generation from **MR** or **CBCT**, built with
+[**KonfAI**](https://github.com/vboussot/KonfAI). Top-ranking models from the
+**[SynthRAD](https://synthrad2023.grand-challenge.org/)** challenge (Tasks 1 & 2).
+
+## 🧩 Models
+
+| Model | Input | Output | Description | Ensemble |
+|:--|:--|:--|:--|:--:|
+| `MR` | MR | sCT | MR → sCT (SynthRAD Task 1) | 5 |
+| `CBCT` | CBCT | sCT | CBCT → sCT (SynthRAD Task 2) | 5 |
+| `MR_CBCT` | MR / CBCT | sCT | Joint MR/CBCT → sCT | 5 |
+| `Finetune` | CBCT | sCT | BICMAC fine-tuned CBCT → sCT | 5 |
+
+2.5D UNet++ · patch `[1, 512, 512]` · 5-model ensemble.
+
+## 🚀 Usage
+
+```bash
+pip install impact_synth_konfai
+impact-synth-konfai synthesize MR -i input_mr.nii.gz -o output/
+```
+
+- **Generic runner:** `konfai-apps infer VBoussot/ImpactSynth:MR -i input_mr.nii.gz -o output/`
+- **Interactive:** [**SlicerKonfAI**](https://github.com/vboussot/SlicerKonfAI) — the ⚙ *Advanced* dialog overrides patch size and batch size.
+
+## ⚡ Performance & VRAM
+
+Benchmarked on an **NVIDIA RTX PRO 5000 (24 GB)**. The batch size is **auto-selected from your free GPU VRAM**.
+
+| Free VRAM | Batch (auto) | Peak VRAM |
+|:--|:--|:--|
+| 8 GB  | 16 | ~7.6 GB |
+| 16 GB | 28 | ~15 GB |
+| 24 GB | 48 | ~21.7 GB |
+
+≈ **25 s / case** on the benchmark volume (scales with case size). Override with `--patch-size` / `--batch-size`.
+
+## 🔗 Links & Citation
+
+- 🧠 **KonfAI:** [github.com/vboussot/KonfAI](https://github.com/vboussot/KonfAI)
+- 📦 **PyPI:** [impact_synth_konfai](https://pypi.org/project/impact_synth_konfai/)
+- 📄 **Paper:** [arXiv:2510.21358](https://arxiv.org/abs/2510.21358)