Update README.md

README.md

@@ -9,29 +9,32 @@ tags:
   - reasoning-segmentation
   - earth-observation
   - custom_code
-license: other  # TODO: choose license (InternVL3.5: Apache-2.0; SAM3: SAM License)
+license: other  # TODO: SAM License on sam3_delta.safetensors, Apache-2.0 on the rest, some training datasets might be research use only
 ---
 
 # MLRS
 
-<!-- TODO: paper title, authors, links -->
+**More with Less: a Large Scale Remote Sensing VLM with a Simple Recipe**
 
-MLRS is a remote-sensing vision-language model that answers questions and performs
-reasoning segmentation. A LoRA-finetuned **InternVL3.5-8B** reasons about the image and either
+Accepted to **ACCV 2026**. 📄 Paper: https://arxiv.org/abs/2607.15942
+
+MLRS is a remote-sensing vision-language model that answers questions and performs reasoning segmentation. 
+MLRS supports wide range of inputs: ultra-high-res, SAR, thermal, false-color, multi-temporal, multi-view, visual prompting, multi-image. 
+A LoRA-finetuned **InternVL3.5-8B** reasons about the image (or multiple images in rgb format) and either
 answers in text or calls a segmentation tool:
 
 ```
 <think>...</think><segmentation>{"noun phrase": "...", "objects": [{"bbox": [x1, y1, x2, y2]}]}</segmentation>
 ```
 
-Each noun phrase and its boxes (on a 0–999 grid) are passed to **SAM3**, whose mask decoder was finetuned
-jointly. SAM3 returns one mask per phrase, clipped to the predicted boxes. Text-only tasks
+Each noun phrase, its boxes (on a 0–999 grid), and the last rgb image are passed to **SAM3**, whose mask decoder was finetuned
+jointly via GRTO. SAM3 returns one mask per phrase, clipped to the predicted boxes. Text-only tasks
 answer with `<answer>...</answer>` and return no masks.
 
 This repository only holds the finetuned deltas (a LoRA adapter and SAM3 mask-decoder weights).
 The base models are downloaded from the Hub at the revisions pinned in `config.json`.
 
-> `facebook/sam3` is gated. Accept its license on the Hub and log in (`hf auth login`) before loading.
+> `facebook/sam3` is gated. Accept its license on the Hub and log in (`hf auth login` or set `HF_TOKEN` env.variable) before loading.
 
 ## Usage
 
@@ -40,7 +43,7 @@ import torch
 from PIL import Image
 from transformers import AutoModel, AutoProcessor
 
-repo = "TODO-org/MLRS"
+repo = "INSAIT-Institute/MLRS"
 model = AutoModel.from_pretrained(repo, trust_remote_code=True, dtype=torch.bfloat16, device_map="cuda")
 processor = AutoProcessor.from_pretrained(repo, trust_remote_code=True)
 
@@ -76,10 +79,18 @@ back to the size of the input image.
 
 - The defaults reproduce the evaluation setup: greedy decoding, `max_new_tokens=1024`, and the
   LoRA adapter kept unmerged. `AutoModel.from_pretrained(..., merge_lora=True)` generates faster, with small numerical differences.
-- Tested with `transformers==TODO`, `peft==0.18.0`, `torch==2.9.0`.
+- Tested with `transformers==5.18.0`, `peft==0.18.0`, `torch==2.9.0`.
 
 ## Citation
 
 ```bibtex
-TODO
+@misc{ailuro2026mlrs,
+      title={{More with Less}: a Large Scale Remote Sensing {VLM} with a Simple Recipe}, 
+      author={Stefan Maria Ailuro and Mario Markov and Mohammad Mahdi and Luc Van Gool and Danda Pani Paudel},
+      year={2026},
+      eprint={2607.15942},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2607.15942}, 
+}
 ```