Raise image cap from 512x512 to configurable max_image_pixels (default 2048x2048)

DecisionModel gains a max_image_pixels kwarg (default 2048*2048); processor_config longest_edge 262144 -> 4194304; README documents aspect-ratio-preserving resize and token budget; manifest checksums updated.

README.md

@@ -83,8 +83,38 @@ probabilities = model.predict([row])[0]
 print(answer(question, probabilities))
 ```
 
-`source/src/autojev/model.py` is copied byte-for-byte from the evaluated training
-run. `release-manifest.json` records file checksums and checkpoint provenance.
+### Images
+
+Pass images as file paths, PIL images, or `data:image/...;base64` URLs in
+`row["images"]`. The processor resizes them **while preserving aspect ratio**.
+Each side is rounded to a multiple of 32 px, and the total area is kept between
+256×256 (65,536 px) and `max_image_pixels`. Each 32×32 patch costs one visual token,
+so an image uses about `width * height / 1024` tokens.
+
+`max_image_pixels` defaults to 2048×2048 (4,194,304 px, ≤4,096 tokens per image).
+Larger images are downscaled automatically, so you don't need to resize them
+yourself. To lower or raise the cap:
+
+```python
+model = DecisionModel(checkpoint, device="cuda", max_image_pixels=1600 * 1000)
+```
+
+Keep in mind:
+
+- The text and all images of a decision must fit in `prepare`'s 8,192-token
+  limit. Inputs are rejected rather than truncated. With several images per
+  decision, lower `max_image_pixels` or downscale the images first.
+- Training and the reported evaluation capped images at 512×512 (262,144 px,
+  ≤256 tokens). Higher resolutions keep more detail but fall outside the training
+  distribution, and they cost more memory and latency. To reproduce the evaluated
+  setup exactly, use `max_image_pixels=512 * 512`.
+- If you resize images yourself, scale both sides by the same factor
+  `min(1, sqrt(max_image_pixels / (width * height)))` to keep the aspect ratio.
+
+`source/src/autojev/model.py` is the evaluated training-run code, with one change:
+the image-size cap is now the configurable `max_image_pixels`. It was previously
+hardcoded at 512×512 pixels. `release-manifest.json` records file checksums and
+checkpoint provenance.
 
 
 ## Acknowledgment

processor_config.json

@@ -20,7 +20,7 @@
     "resample": 3,
     "rescale_factor": 0.00392156862745098,
     "size": {
-      "longest_edge": 262144,
+      "longest_edge": 4194304,
       "shortest_edge": 65536
     },
     "temporal_patch_size": 2

release-manifest.json

@@ -20,8 +20,8 @@
       "sha256": "3792cbe9e964a7ed65fff010d3f94bb27c6e726e2940ab6b0e1db5d30df92510"
     },
     "README.md": {
-      "bytes": 4339,
-      "sha256": "7fd072ef02e89803dede5d57ea578d8eeab075137d39ab89da856d925d83aa8d"
+      "bytes": 5696,
+      "sha256": "265e16155cbc9c65bcf2a797521b8214eb659cf327f7bbed19ddcc298491eae4"
     },
     "chat_template.jinja": {
       "bytes": 8952,
@@ -92,8 +92,8 @@
       "sha256": "dc39d771cb240f775399e1845b11d081d8e007f14eecae3b780bc6272b128a46"
     },
     "processor_config.json": {
-      "bytes": 1218,
-      "sha256": "abb488b9a03a53407dc9c8e43c0a2caec9142e5fc7a1203807fdf54b2149cc8f"
+      "bytes": 1219,
+      "sha256": "f76f21cad63dc9da928676e84e4f4889f154b9122e8c3f6b38093d39729fc266"
     },
     "readout.safetensors": {
       "bytes": 2611288,
@@ -132,8 +132,8 @@
       "sha256": "2e4f4ae4b8aafac0247f931220b592a140bd2f5cd032dce8a517a158fbeedda3"
     },
     "source/src/autojev/model.py": {
-      "bytes": 17038,
-      "sha256": "824721da4d8f8ba1e534f3ee35081869243c4457f6b3bd0a9a91b611b2ec1076"
+      "bytes": 17304,
+      "sha256": "02940963e9fba5736b8f625f860812f31de2e6658a49606fbed49fd7215c20f9"
     },
     "source/src/autojev/optim.py": {
       "bytes": 4702,

source/src/autojev/model.py

@@ -23,6 +23,8 @@ from autojev.types import Answer, Content, DecisionInput, ImageInput, JSONValue,
 BASE_MODEL = "Qwen/Qwen3.8-27B"
 BASE_REVISION = "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0"
 MAX_OPTIONS = 255
+MIN_IMAGE_PIXELS = 256 * 256
+MAX_IMAGE_PIXELS = 2048 * 2048
 
 
 def enable_noncausal_full_attention(text_model: torch.nn.Module) -> None:
@@ -170,6 +172,7 @@ class DecisionModel(torch.nn.Module):
         cache_dir: str | Path | None = None,
         attention_mode: str | None = None,
         pooling: str | None = None,
+        max_image_pixels: int = MAX_IMAGE_PIXELS,
     ) -> None:
         super().__init__()
         torch.set_num_threads(cpu_threads)
@@ -205,7 +208,9 @@ class DecisionModel(torch.nn.Module):
             ),
         )
         self.processor.tokenizer.padding_side = "left"
-        self.processor.image_processor.size = {"shortest_edge": 65536, "longest_edge": 262144}
+        if max_image_pixels < MIN_IMAGE_PIXELS:
+            raise ValueError(f"max_image_pixels must be at least {MIN_IMAGE_PIXELS}.")
+        self.processor.image_processor.size = {"shortest_edge": MIN_IMAGE_PIXELS, "longest_edge": max_image_pixels}
         tokenizer = self.processor.tokenizer
         candidates = list(string.ascii_uppercase) + [
             "".join(pair) for pair in itertools.product(string.ascii_uppercase, repeat=2)