Update README.md

README.md

@@ -14,20 +14,19 @@ tags:
   <img src="https://github.com/wileewang/LongLive2.0/blob/release-clean-merge/assets/longlive2/logo.png?raw=true" alt="LongLive2.0 logo" width="100%">
 </p>
 
-# LongLive2.0 5B Checkpoints
+# LongLive2.0 5B Checkpoint
 
-This repository hosts temporary LongLive2.0 5B checkpoints for inference with
+This repository hosts the LongLive2.0 5B checkpoint for inference with
 the LongLive2.0 release code:
 
 https://github.com/wileewang/LongLive2.0
 
-The checkpoint package contains two parts:
+The checkpoint package contains a ready-to-use merged generator checkpoint. The
+AR-trained base generator and DMD-distilled LoRA adapter have already been
+merged, so standard inference only needs to load one `generator_ckpt`.
 
-- **Base generator checkpoint**: the AR-trained Wan2.2-TI2V-5B generator.
-- **LoRA checkpoint**: the DMD-distilled few-step LoRA adapter.
-
-LongLive2.0 inference loads the base generator first, applies the LoRA modules,
-and then loads the LoRA weights.
+Do not configure a separate `lora_ckpt` or `adapter` section when using the
+merged checkpoint, otherwise the LoRA adapter would be applied a second time.
 
 ## Installation
 
@@ -65,15 +64,7 @@ Edit `configs/inference.yaml`:
 
 ```yaml
 checkpoints:
-  generator_ckpt: checkpoints/longlive2_5b/path/to/base_generator.pt
-  lora_ckpt: checkpoints/longlive2_5b/path/to/dmd_lora.pt
-
-adapter:
-  type: lora
-  rank: 128
-  alpha: 128
-  dropout: 0.0
-  verbose: true
+  generator_ckpt: checkpoints/longlive2_5b/merged_generator.pt
 
 data:
   data_path: /path/to/inference_prompts
@@ -89,9 +80,10 @@ inference:
   multi_shot_rope_offset: 8
 ```
 
-Replace the checkpoint filenames above with the actual files in this repository.
-If the LoRA checkpoint is not used, remove the `adapter` section and leave
-`lora_ckpt` unset.
+Replace `merged_generator.pt` with the actual merged checkpoint filename in this
+repository. If your local config was copied from a base+LoRA setup, remove
+`checkpoints.lora_ckpt` and the top-level `adapter` section before running
+inference.
 
 ## Prompt Folder
 
@@ -159,8 +151,10 @@ Outputs are written to `output_folder`.
 
 ## Notes
 
-- The base checkpoint and LoRA checkpoint should be loaded together for the
-  few-step DMD model.
+- The released BF16 checkpoint is already merged; standard inference only needs
+  `checkpoints.generator_ckpt`.
+- Do not set `checkpoints.lora_ckpt` or `adapter` when using the merged
+  checkpoint.
 - `inference.sampling_steps` controls the number of denoising steps.
 - `inference.multi_shot_sink` enables the multi-shot attention sink.
 - `inference.multi_shot_rope_offset` controls the multi-shot RoPE offset.