Initial release

README.md

@@ -1,8 +1,5 @@
 ---
 license: mit
-base_model:
-- Qwen/Qwen3.5-4B
-- Qwen/Qwen3.5-0.8B
 pipeline_tag: text-to-audio
 tags:
 - spatial-audio
@@ -20,10 +17,7 @@ Unified AMBIT supports native First-Order Ambisonics (FOA) generation and editin
 **text → AR → ScenePlan → DiT → FOA**. Editing additionally takes reference FOA audio.
 The AR-generated ScenePlan is used directly for rendering; no user-supplied plan is required.
 
-This release is the Unified 275k checkpoint. It continues the Unified 250k model
-for 25k updates after upgrading the AR request encoder to Qwen3.5-4B and expanding
-the ScenePlan output head. The renderer retains Qwen3.5-0.8B text conditioning.
-AR planning and DiT rendering share a Transformer backbone.
+One unified model supports both generation and editing. AR and DiT share the same Transformer backbone. All necessary model assets are included in the download.
 
 ## Use
 
@@ -56,15 +50,7 @@ python -m ambit.infer edit --bundle models/AMBIT \
   --output outputs/edited.wav --seed 42
 ```
 
-Both commands save a float WAV and a JSON containing the generated ScenePlan.
-Audio is 44,100 Hz, four-channel ACN/SN3D FOA in W, Y, Z, X order. Editing accepts
-FOA clips up to 663,552 samples (about 15 seconds); mono and stereo are not accepted
-by this entry point. Use an ambisonic decoder for headphone playback.
-
-Default sampling: generation 100 Euler steps / CFG 3; editing 20 steps / CFG 1.
-The text request must fit 512 tokens and the generated plan supports up to four sources.
-Planning, verbatim speech and source preservation may be imperfect. This is a model
-release, not a claim of perfect instruction following or unchanged voice identity.
+Both commands save FOA audio and the generated ScenePlan. Editing uses a 44.1 kHz, four-channel ACN/SN3D FOA reference in W, Y, Z, X order.
 
 ## Files
 
@@ -72,17 +58,10 @@ release, not a claim of perfect instruction following or unchanged voice identit
 - `vae.safetensors`: FOA audio encoder/decoder.
 - `bundle.json`, `renderer.json`, `vae.json`: inference configurations.
 - `codec/`: ScenePlan tokenization assets.
-- `planner_encoder/`, `text_encoder/`: frozen Qwen text encoders and tokenizers.
-- `SHA256SUMS`: artifact checksums.
-
-This is an inference export of the 275k checkpoint, not a resumable training checkpoint.
-It contains no optimizer state or training data. Use the AMBIT loader, not AutoModel
-on the root of this repository. The two Qwen subdirectories are encoder assets.
+- `planner_encoder/`, `text_encoder/`: text encoders and tokenizers.
 
 ## License and attribution
 
-AMBIT code/weights are distributed under the repository MIT license. The included
-Qwen3.5-4B and Qwen3.5-0.8B text encoders and tokenizer assets retain their Apache-2.0
-licenses, included under `LICENSES/`. These are frozen text-encoder exports rather
-than complete vision/chat model packages. AMBIT builds on
-[stable-audio-tools](https://github.com/Stability-AI/stable-audio-tools).
+AMBIT is distributed under the MIT license. Included third-party components
+retain their respective licenses; see `LICENSES/`.
+AMBIT builds on [stable-audio-tools](https://github.com/Stability-AI/stable-audio-tools).