Release TA-only step5750 checkpoint and config

README.md

@@ -4,7 +4,7 @@ language:
 - en
 - ru
 base_model:
-- kandinskylab/Kandinsky-5.0-I2V-Lite-sft-5s
+  - kandinskylab/Kandinsky-5.0-I2V-Lite-5s
 pipeline_tag: image-to-video
 tags:
 - image-to-video
@@ -15,49 +15,52 @@ tags:
 - latent-time
 widget:
 - text: >-
-    A close-up of a green-eyed girl in a dress stands in the city. The girl
-    smiles and waves. Then she turns her back and walks away.
+    Target FPS: 15. A close-up of a green-eyed girl in a dress stands in the
+    city. The girl smiles and waves. Then she turns her back and walks away.
   output:
-    url: examples/ftta_pose_15.mp4
+    url: examples/ta_pose_15.mp4
 - text: >-
-    A close-up of a green-eyed girl in a dress stands in the city. The girl
-    smiles and waves. Then she turns her back and walks away.
+    Target FPS: 30. A close-up of a green-eyed girl in a dress stands in the
+    city. The girl smiles and waves. Then she turns her back and walks away.
   output:
-    url: examples/ftta_pose_30.mp4
+    url: examples/ta_pose_30.mp4
 - text: >-
-    A close-up of a green-eyed girl in a dress stands in the city. The girl
-    smiles and waves. Then she turns her back and walks away.
+    Target FPS: 60. A close-up of a green-eyed girl in a dress stands in the
+    city. The girl smiles and waves. Then she turns her back and walks away.
   output:
-    url: examples/ftta_pose_60.mp4
+    url: examples/ta_pose_60.mp4
 - text: >-
-    Beautiful views of the forest with fog. The wind dispels the fog and the
-    sun peeks out over the forest. The sun's rays illuminate a beautiful green
-    forest.
+    Target FPS: 15. Beautiful views of the forest with fog. The wind dispels
+    the fog and the sun peeks out over the forest. The sun's rays illuminate a
+    beautiful green forest.
   output:
-    url: examples/ftta_forest_15.mp4
+    url: examples/ta_forest_15.mp4
 - text: >-
-    Beautiful views of the forest with fog. The wind dispels the fog and the
-    sun peeks out over the forest. The sun's rays illuminate a beautiful green
-    forest.
+    Target FPS: 30. Beautiful views of the forest with fog. The wind dispels
+    the fog and the sun peeks out over the forest. The sun's rays illuminate a
+    beautiful green forest.
   output:
-    url: examples/ftta_forest_30.mp4
+    url: examples/ta_forest_30.mp4
 - text: >-
-    Beautiful views of the forest with fog. The wind dispels the fog and the
-    sun peeks out over the forest. The sun's rays illuminate a beautiful green
-    forest.
+    Target FPS: 60. Beautiful views of the forest with fog. The wind dispels
+    the fog and the sun peeks out over the forest. The sun's rays illuminate a
+    beautiful green forest.
   output:
-    url: examples/ftta_forest_60.mp4
+    url: examples/ta_forest_60.mp4
 ---
 <div style="background-color: #f8f9fa; padding: 20px; border-radius: 10px; margin-bottom: 20px;">
   <h1 style="color: #24292e; margin-top: 0;">Time Adapter for Kandinsky-5 I2V</h1>
   <div style="background-color: white; padding: 15px; border-radius: 8px; margin: 15px 0; box-shadow: 0 2px 4px rgba(0,0,0,0.1);">
     <h2 style="color: #24292e; margin-top: 0;">Overview</h2>
-    <p>This repository contains a lightweight temporal adapter for Kandinsky-5 image-to-video generation.</p>
+    <p>This repository contains a lightweight TA-only temporal adapter checkpoint for Kandinsky-5 image-to-video generation.</p>
     <p>Lower target FPS typically leads to slower perceived motion, while higher target FPS leads to faster perceived motion.</p>
+    <p>The released checkpoint keeps the base DiT weights unchanged and adds only the time-adapter parameters for global FPS conditioning and latent-time control.</p>
+    <p>Use this checkpoint with the `time_adapter` branch of <code>kandinsky-5</code> and the config <code>configs/k5_i2v_time_adapter_step5750_ta_only.yaml</code>.</p>
   </div>
   <div style="background-color: white; padding: 15px; border-radius: 8px; margin: 15px 0; box-shadow: 0 2px 4px rgba(0,0,0,0.1);">
     <h2 style="color: #24292e; margin-top: 0;">Examples</h2>
-    <p>The gallery below shows FTTA examples for pose motion control and forest motion control at different target FPS values.</p>
+    <p>The gallery below shows TA-only examples for pose motion control and forest motion control at different target FPS values.</p>
+    <p>The config follows the standard `kandinsky-5` layout and expects model files under <code>weights/model</code>, <code>weights/vae</code>, <code>weights/text_encoder</code>, and <code>weights/text_encoder2</code>.</p>
   </div>
 </div>
 <Gallery />

configs/k5_i2v_time_adapter_step5750_ta_only.yaml

@@ -0,0 +1,54 @@
+metrics:
+  scheduler_scale: 10
+  legacy_resize: true
+  scale_factor:
+  - 1.0
+  - 2.0
+  - 2.0
+  resolution: 512
+model:
+  checkpoint_path: "./weights/model/time_adapter_step5750_ta_only.safetensors"
+  num_steps: 50
+  guidance_weight: 5.0
+  dit_params:
+    in_visual_dim: 16
+    out_visual_dim: 16
+    time_dim: 512
+    patch_size:
+    - 1
+    - 2
+    - 2
+    model_dim: 1792
+    ff_dim: 7168
+    num_text_blocks: 2
+    num_visual_blocks: 32
+    axes_dims:
+    - 16
+    - 24
+    - 24
+    visual_cond: true
+    in_text_dim: 3584
+    in_text_dim2: 768
+    use_condition_fps: true
+    use_latent_time: true
+  attention:
+    type: flash
+    causal: false
+    local: false
+    glob: false
+    window: 3
+  vae:
+    checkpoint_path: "./weights/vae/"
+    name: "hunyuan"
+  text_embedder:
+    qwen:
+      checkpoint_path: "./weights/text_encoder/"
+      emb_size: 3584
+      max_length: 256
+      mode: "i2v"
+    clip:
+      checkpoint_path: "./weights/text_encoder2/"
+      emb_size: 768
+      max_length: 77
+magcache:
+  mag_ratios: [0.90524, 0.89864, 0.95225, 0.95382, 1.04828, 1.04658, 1.07989, 1.07921, 1.04345, 1.04345, 1.03547, 1.03589, 1.03715, 1.03745, 1.02909, 1.02907, 1.03391, 1.03416, 1.02366, 1.02392, 1.02359, 1.02415, 1.02256, 1.02289, 1.02447, 1.02502, 1.01824, 1.01857, 1.02236, 1.02261, 1.01704, 1.01736, 1.01838, 1.01857, 1.01886, 1.01961, 1.01619, 1.0169, 1.01655, 1.01702, 1.01477, 1.01545, 1.01926, 1.01981, 1.01145, 1.01227, 1.01276, 1.01308, 1.01583, 1.01619, 1.01587, 1.01641, 1.01386, 1.01418, 1.01441, 1.01505, 1.01473, 1.01491, 1.01346, 1.01407, 1.01319, 1.0137, 1.01283, 1.01313, 1.01412, 1.01407, 1.01222, 1.0127, 1.0131, 1.01315, 1.01398, 1.0143, 1.01483, 1.01498, 1.01158, 1.01176, 1.01342, 1.01365, 1.00893, 1.0095, 1.01223, 1.01237, 1.01005, 1.01045, 1.00694, 1.00679, 1.00691, 1.00742, 0.99846, 0.99916, 0.98563, 0.98549, 0.97116, 0.97145, 0.93116, 0.93069, 0.83437, 0.83375]

examples/{ftta_pose_15.mp4 → ta_pose_15.mp4}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1b7cb4a42f8b3e4ed1bc1ae6474758fbfc9e7603906221715715917c47012598
-size 3228724
+oid sha256:d4062a876d137300161c3125908fabe29fc69fce79086af802814f3ade95623f
+size 6104410

examples/{ftta_pose_30.mp4 → ta_pose_30.mp4}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5617cca1f9afac88e788ed958a53e08d2643924a8d97a076d73e9036e9f6dce8
-size 2503432
+oid sha256:53a42b5da401cade02b15b321f787ff6f87284e871facea1cfb1857440d9f00b
+size 4241931

examples/{ftta_pose_60.mp4 → ta_pose_60.mp4}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:954359847cd8a241fae9e50b448a19138ed3a316532e62f94d957109049b6878
-size 1985107
+oid sha256:2b2f19aa9d0cb1bb3fb364f6f1830b9a5d9434ae2bf75330fffbcd7d61b4716e
+size 2610097

model/config.json

@@ -0,0 +1,6 @@
+{
+  "architectures": [
+    "DiffusionTransformer3D"
+  ],
+  "model_type": "kandinsky5"
+}

weights/model/time_adapter_step5750_ta_only.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:995c7bea67011471a1ecafd8725c3f0431cd1da6d0e5f3a8befc908faf82d1e4
+size 8061347376