diff --git a/README.md b/README.md
index 0d5bb4e..8fd98da 100644
--- a/README.md
+++ b/README.md
@@ -50,6 +50,8 @@ curl -H "Authorization: Bearer dev" "localhost:8080/generate?prompt=orange+cat"
 
 ## Model and license
 
-`model/` holds an ONNX conversion of [AuroraAI-Research/AuroraIMG-6M](https://huggingface.co/AuroraAI-Research/AuroraIMG-6M) at revision `3d66743`, licensed under the Apache License 2.0 (copy in [`model/LICENSE`](model/LICENSE)). The weights are unchanged; they were converted from the original safetensors to ONNX, and the latent un-normalization and output scaling were folded into `decoder.onnx`. Output matches the original pipeline to within 1/255 per pixel.
+`model/` holds an ONNX conversion of [AuroraAI-Research/AuroraIMG-6M](https://huggingface.co/AuroraAI-Research/AuroraIMG-6M) at revision `3d66743`, licensed under the Apache License 2.0 (copy in [`model/LICENSE`](model/LICENSE), attribution and changes in [`model/NOTICE`](model/NOTICE); both ship in the Docker image). The weights are unchanged; they were converted from the original safetensors to ONNX, and the latent un-normalization and output scaling were folded into `decoder.onnx`. Output matches the original pipeline to within 1/255 per pixel.
+
+The model was trained on [CC12M](https://huggingface.co/datasets/pixparse/cc12m-wds), which is distributed as image URLs and captions; the underlying images carry their own licenses.
 
 `scripts/export_onnx.py` regenerates `model/` from the original repo (needs torch, onnx, onnxscript, onnxruntime, safetensors).
diff --git a/model/NOTICE b/model/NOTICE
new file mode 100644
index 0000000..de3941d
--- /dev/null
+++ b/model/NOTICE
@@ -0,0 +1,15 @@
+AuroraIMG-6M (ONNX conversion)
+
+The files in this directory are derived from AuroraIMG-6M by AuroraAI-Research (2026),
+https://huggingface.co/AuroraAI-Research/AuroraIMG-6M, revision 3d66743c3139d034b832a670d067221f5d544bbb,
+licensed under the Apache License, Version 2.0 (see LICENSE in this directory).
+
+Changes from the original:
+- text_encoder.onnx, denoiser.onnx and decoder.onnx were converted from the original
+  diffusion_pytorch_model.safetensors weights to ONNX with scripts/export_onnx.py. The weights are unchanged.
+- decoder.onnx also includes the latent un-normalization (lat_mu, lat_sd from vae/config.json) and the
+  output scaling to 0-1 RGB, which the original pipeline.py does outside the VAE.
+- tokenizer.json is copied unchanged from tokenizer/tokenizer.json.
+
+The model was trained on CC12M (https://huggingface.co/datasets/pixparse/cc12m-wds), which is distributed as
+image URLs and captions; the underlying images carry their own licenses.