feat: no VRAM plan, every preset leaves its patch free and blends its tiles

No app declares a `vram_plan` any more. Every preset leaves its patch free (`patch_size: [0, 0, 0]`): the volume runs as one patch when it fits, and the engine is handed exactly what it was without a patch. Otherwise konfai tiles it on the measured VRAM, the fewest equal patches that fit, overlapping by 20 % and blended with a `Cosinus` window: without a declared overlap and window the last patch written won, which left a seam in the displacement field where a patch starts. A registration forwards one pair: its batch stays 1.

CBCT_CT_HeadNeck/Prediction.yml

@@ -138,8 +138,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -163,7 +163,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-CBCT-CT-HeadNeck

CBCT_CT_MRSeg/Prediction.yml

@@ -171,8 +171,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -196,7 +196,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-CBCT-CT-MRSeg

ConvexAdam_Coarse/Prediction.yml

@@ -78,8 +78,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -103,7 +103,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-ConvexAdam-Coarse

ConvexAdam_Composite/Prediction.yml

@@ -79,8 +79,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -104,7 +104,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-ConvexAdam-Composite

ConvexAdam_Fine/Prediction.yml

@@ -78,8 +78,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -103,7 +103,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-ConvexAdam-Fine

ConvexAdam_Fine/app.json

@@ -1,44 +1,10 @@
 {
     "display_name": "ConvexAdam Fine (MIND)",
     "short_description": "Fine Adam refinement (tileable; expects a coarse/pre-aligned start).",
-    "description": "Second stage of the ConvexAdam pipeline: the Adam instance-optimisation refinement driven by the IMPACT metric on MIND features, with no linear pre-align (it assumes the moving is already on the fixed grid, e.g. resampled by the coarse stage). Runs from a zero warm-start and tiles naturally for large images (set a patch size).",
+    "description": "Second stage of the ConvexAdam pipeline: the Adam instance-optimisation refinement driven by the IMPACT metric on MIND features, with no linear pre-align (it assumes the moving is already on the fixed grid, e.g. resampled by the coarse stage). Runs from a zero warm-start. A volume that fits the GPU runs as one patch; one that does not is tiled on its own, the tiles overlapping by 20 % and blended (a patch size can still be forced).",
     "task": "registration",
     "tta": 0,
     "mc_dropout": 0,
-    "vram_plan": {
-        "8": {
-            "patch_size": [
-                128,
-                128,
-                128
-            ],
-            "batch_size": 1
-        },
-        "16": {
-            "patch_size": [
-                192,
-                192,
-                192
-            ],
-            "batch_size": 1
-        },
-        "24": {
-            "patch_size": [
-                256,
-                256,
-                256
-            ],
-            "batch_size": 1
-        },
-        "40": {
-            "patch_size": [
-                320,
-                320,
-                320
-            ],
-            "batch_size": 1
-        }
-    },
     "models": [],
     "inputs": {
         "Fixed": {

FireANTs_Affine/Prediction.yml

@@ -75,8 +75,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -100,7 +100,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-FireANTs-Affine

FireANTs_IMPACT/Prediction.yml

@@ -82,8 +82,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -107,7 +107,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-FireANTs-IMPACT

FireANTs_SyN/Prediction.yml

@@ -75,8 +75,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -100,7 +100,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-FireANTs-SyN

Generic_Rigid/Prediction.yml

@@ -59,8 +59,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -84,7 +84,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-Generic-Rigid

Generic_Rigid_BSpline/Prediction.yml

@@ -61,8 +61,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -86,7 +86,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-Generic-Rigid-BSpline

MR_CT_HeadNeck/Prediction.yml

@@ -123,8 +123,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -148,7 +148,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-MR-CT-HeadNeck

MR_CT_MRSeg/Prediction.yml

@@ -171,8 +171,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -196,7 +196,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-MR-CT-MRSeg

MR_CT_TS/Prediction.yml

@@ -171,8 +171,8 @@ Predictor:
             prob: 1
         nb: 2
     Patch:
-      patch_size: None
-      overlap: None
+      patch_size: [0, 0, 0]
+      overlap: 20%
       pad_value: None
       extend_slice: 0
     subset: None
@@ -196,7 +196,7 @@ Predictor:
         dataset_filename: Transform:itktransform
         group: Transform
         same_as_group: Volume_0:Fixed
-        patch_combine: None
+        patch_combine: Cosinus
         reduction: Mean
         Mean: {}
   train_name: ImpactReg-MR-CT-TS

README.md

@@ -75,6 +75,21 @@ Per-stage breakdown: linear pre-align **≈ 4.2 s** (ITK affine, MI) · coarse *
 One-time TorchScript feature-model load **≈ 7 s** (amortised across a batch). Times scale with case size;
 `--tta k` multiplies runtime. The `elastix + IMPACT` presets run through elastix and scale differently.
 
+## 🧩 Large volumes
+
+Every preset leaves its patch free (`patch_size: [0, 0, 0]`). A pair that fits the GPU runs as one patch, and
+the engine receives exactly what it did before. When the GPU runs out of memory, KonfAI cuts the volume into
+the fewest equal patches that fit, registers each, and blends the fields over a 20 % overlap with a cosine
+window. A patch size can still be forced with `--patch-size`.
+
+The `elastix + IMPACT` presets rarely get there. IMPACT computes its features on the whole image and, when that
+does not fit the GPU, retries on smaller feature patches by itself: the run is slower, and the registration
+stays one global registration.
+
+Tiling suits a local, deformable stage (`ConvexAdam_Fine`, `FireANTs_SyN`). A global stage
+(rigid, affine, the coarse ConvexAdam initialisation) is then estimated per patch, which is no longer one
+global transform: for those, prefer a GPU the volume fits on, or a coarser input.
+
 ## 🔗 Links & Citation
 
 - 🧠 **KonfAI:** [github.com/vboussot/KonfAI](https://github.com/vboussot/KonfAI)