Add model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
pipeline_tag: text-to-video
|
| 3 |
+
tags:
|
| 4 |
+
- video-generation
|
| 5 |
+
- autoregressive-video
|
| 6 |
+
- diffusion-distillation
|
| 7 |
+
- wan2.1
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
<h1 align="center">DistillAlign: Coordinating Mode Covering and Mode Seeking in Autoregressive Video Distillation</h1>
|
| 11 |
+
|
| 12 |
+
<p align="center"><b>Anonymous release for ICLR 2027 submission</b></p>
|
| 13 |
+
|
| 14 |
+
**DistillAlign** aligns and balances the **mode-covering** and **mode-seeking**
|
| 15 |
+
objectives of the multi-stage video distillation pipeline — using only a
|
| 16 |
+
**1.3B DMD teacher**, it already surpasses baselines refined with a **14B DMD
|
| 17 |
+
teacher**.
|
| 18 |
+
|
| 19 |
+
## Checkpoints
|
| 20 |
+
|
| 21 |
+
All released generators are **Wan2.1-1.3B students**; the size in each name
|
| 22 |
+
refers to the **teacher** used during training.
|
| 23 |
+
|
| 24 |
+
| Model | Checkpoint | Description |
|
| 25 |
+
|---|---|---|
|
| 26 |
+
| Initializer (1.3B teacher) | [distillalign_init_1p3b_teacher.pt](https://huggingface.co/DistillAlign-Anonymous/DistillAlign/blob/main/checkpoints/distillalign_init_1p3b_teacher.pt) | Pre-DMD initializer, trained toward a Wan2.1-T2V-1.3B teacher |
|
| 27 |
+
| Initializer (14B teacher) | [distillalign_init_14b_teacher.pt](https://huggingface.co/DistillAlign-Anonymous/DistillAlign/blob/main/checkpoints/distillalign_init_14b_teacher.pt) | Pre-DMD initializer, trained toward a Wan2.1-T2V-14B teacher |
|
| 28 |
+
| Distilled (1.3B teacher) | [distillalign_distill_1p3b_teacher.pt](https://huggingface.co/DistillAlign-Anonymous/DistillAlign/blob/main/checkpoints/distillalign_distill_1p3b_teacher.pt) | Final joint-distilled generator, Wan2.1-T2V-1.3B DMD teacher |
|
| 29 |
+
| Distilled (14B teacher) | [distillalign_distill_14b_teacher.pt](https://huggingface.co/DistillAlign-Anonymous/DistillAlign/blob/main/checkpoints/distillalign_distill_14b_teacher.pt) | Final joint-distilled generator, Wan2.1-T2V-14B DMD teacher |
|
| 30 |
+
|
| 31 |
+
Each checkpoint stores a plain `{"generator": state_dict}` and loads directly
|
| 32 |
+
with the inference and evaluation code in the supplementary material.
|
| 33 |
+
|
| 34 |
+
## Teacher Reference Caches
|
| 35 |
+
|
| 36 |
+
Ready-made teacher reference features for the teacher-normalized
|
| 37 |
+
distribution evaluation. Pass the `.npz` file to `--teacher-features` and
|
| 38 |
+
only the student side needs to be generated.
|
| 39 |
+
|
| 40 |
+
| Cache | Download | Description |
|
| 41 |
+
|---|---|---|
|
| 42 |
+
| Wan2.1-1.3B teacher reference features | [wan2.1_t2v_1.3b_reference_vjepa2.npz](https://huggingface.co/DistillAlign-Anonymous/DistillAlign/blob/main/teacher_caches/wan2.1_t2v_1.3b_reference_vjepa2.npz) | 256 x 2560 V-JEPA2 features, ready for `--teacher-features` |
|
| 43 |
+
| Wan2.1-14B teacher reference features | [wan2.1_t2v_14b_reference_vjepa2.npz](https://huggingface.co/DistillAlign-Anonymous/DistillAlign/blob/main/teacher_caches/wan2.1_t2v_14b_reference_vjepa2.npz) | 256 x 2560 V-JEPA2 features, ready for `--teacher-features` |
|
| 44 |
+
|
| 45 |
+
See `teacher_caches/README.md` for the exact sampling and extraction protocol.
|