feat: add SPEED and LDF-VFI interpolation

Integrate checksum-pinned runtimes, harden interpolation and cleanup paths, and add a SPEED/LDF model-lab workflow.
2026-08-15 21:16:14 +02:00
parent 2d96d5aa5d
commit 5b10a1a594
11 changed files with 2286 additions and 150 deletions
@@ -0,0 +1,11 @@
+ComfyUI-Tween
+
+The LDF-VFI sequence sampling adapter in ldf_backend.py is based on the
+official LDF-VFI implementation:
+https://github.com/xypeng9903/LDF-VFI
+
+LDF-VFI is distributed under the Apache License, Version 2.0. Its official
+runtime is downloaded on demand and retains its upstream LICENSE file.
+
+SPEED runtime source is not redistributed by this project. It is downloaded
+on demand from the official repository at a checksum-pinned commit.
@@ -3,9 +3,9 @@
 [![ComfyUI](https://img.shields.io/badge/ComfyUI-Custom_Node-0a7ef0)](https://registry.comfy.org/)
 [![Python 3.10+](https://img.shields.io/badge/Python-3.10+-3776AB?logo=python&logoColor=white)](https://www.python.org/)
 [![License](https://img.shields.io/badge/License-Apache_2.0-green.svg)](https://www.apache.org/licenses/LICENSE-2.0)
-[![Models](https://img.shields.io/badge/VFI_Models-4-8B5CF6)](#which-model-should-i-use)
+[![Models](https://img.shields.io/badge/VFI_Models-6-8B5CF6)](#which-model-should-i-use)

-Four video frame interpolation models in one package — **BIM-VFI**, **EMA-VFI**, **SGM-VFI**, and **GIMM-VFI**. Designed for long videos with thousands of frames without running out of VRAM.
+Six video frame interpolation models in one package — **BIM-VFI**, **EMA-VFI**, **SGM-VFI**, **GIMM-VFI**, **SPEED**, and **LDF-VFI**. Pairwise models include chunked/segmented processing; LDF-VFI adds holistic long-sequence diffusion interpolation.

 <p align="center">
  <img src="assets/model-comparison.svg" alt="Model Comparison" width="720"/>
@@ -21,11 +21,19 @@ git clone https://github.com/Ethanfel/ComfyUI-Tween.git
 pip install -r requirements.txt
 ```

-All dependencies (`gdown`, `timm`, `omegaconf`, `easydict`, `yacs`, `einops`, `huggingface_hub`) are declared in `pyproject.toml` and `requirements.txt`, installed automatically by ComfyUI Manager or pip.
+Dependencies are declared in `pyproject.toml` and `requirements.txt` and are installed automatically by ComfyUI Manager or pip. LDF-VFI requires PyTorch 2.5+ plus a current `diffusers`/`accelerate` stack.

-### cupy (required for BIM-VFI, SGM-VFI, GIMM-VFI)
+### Demo workflow

-[cupy](https://cupy.dev/) provides GPU-accelerated optical flow warping. **EMA-VFI works without it.**
+Import [`example_workflows/tween_speed_ldf_model_lab.json`](example_workflows/tween_speed_ldf_model_lab.json) for the recommended starter graph. It requires [ComfyUI-VideoHelperSuite](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite) for video loading and encoding.
+
+- The enabled SPEED branch loads a 25-frame, 24 FPS sample, automatically tunes memory settings, interpolates to 48 FPS, preserves audio, and writes `Tween/demo_speed_24_to_48`.
+- The LDF-VFI branch is visibly grouped and muted by default so the workflow does not unexpectedly download ~6.4 GB or reserve ~20 GB VRAM. Enable its three coral nodes when you want to compare the sequence-native model.
+- Keep the loader's `force_rate`, Tween's `source_fps`/`target_fps`, and Video Combine's `frame_rate` synchronized when changing cadence.
+
+### cupy (accelerates BIM-VFI, SGM-VFI, and GIMM-VFI)
+
+[cupy](https://cupy.dev/) provides GPU-accelerated optical flow warping. **EMA-VFI, SPEED, and LDF-VFI do not use it.**

 1. Find your CUDA version:
   ```bash
@@ -55,19 +63,16 @@ All dependencies (`gdown`, `timm`, `omegaconf`, `easydict`, `yacs`, `einops`, `h

 ## Which model should I use?

-| | BIM-VFI | EMA-VFI | SGM-VFI | GIMM-VFI |
-|---|---------|---------|---------|----------|
-| **Best for** | General-purpose | Fast, low VRAM | Large motion | High multipliers (4x/8x) |
-| **Quality** | Highest | Good | Best on large motion | Good |
-| **Speed** | Moderate | Fastest | Slowest | Fast for 4x/8x |
-| **VRAM** | ~2 GB/pair | ~1.5 GB/pair | ~3 GB/pair | ~2.5 GB/pair |
-| **Params** | ~17 M | ~14–65 M | ~15 M + GMFlow | ~80 M (RAFT) / ~123 M (FlowFormer) |
-| **Arbitrary timestep** | Yes | Yes (`_t` checkpoint) | No (fixed 0.5) | Yes (native) |
-| **4x/8x** | Recursive passes | Recursive passes | Recursive passes | Single forward pass |
-| **Requires cupy** | Yes | No | Yes | Yes |
-| **Paper** | CVPR 2025 | CVPR 2023 | CVPR 2024 | NeurIPS 2024 |
+| Model | Best for | Multiplier path | Typical VRAM | Trade-off |
+|-------|----------|-----------------|--------------|-----------|
+| **BIM-VFI** | Strong general pairwise quality | Recursive 2x/4x/8x | ~2 GB/pair | Research/education license |
+| **EMA-VFI** | Speed and lower VRAM | Recursive 2x/4x/8x | ~1.5 GB/pair | Less robust on extreme motion |
+| **SGM-VFI** | Large motion | Recursive 2x/4x/8x | ~3 GB/pair | Slowest pairwise option |
+| **GIMM-VFI** | Arbitrary timesteps, efficient 4x/8x | Native multi-frame per pair | ~2.5 GB/pair | Still frame-pair-centric |
+| **SPEED** | New high-quality midpoint generation | One diffusion step at 2x; recursive 4x/8x | ~2.3–2.6 GB at benchmark resolutions | Stochastic, ~447 MB checkpoint |
+| **LDF-VFI** | Long-range temporal coherence and 2x–16x | Native sequence diffusion | ~20 GB | ~6.4 GB weights; much slower |

-**TL;DR:** Start with **BIM-VFI** for best quality. Use **EMA-VFI** for speed or if you can't install cupy. Use **SGM-VFI** for large camera motion. Use **GIMM-VFI** for 4x/8x without recursive passes.
+**TL;DR:** Try **SPEED** as the modern pairwise default. Use **EMA-VFI** when latency matters, **SGM-VFI** for difficult large motion, **GIMM-VFI** for lightweight arbitrary timesteps, and **LDF-VFI** when sequence consistency matters more than speed or memory.

 ## VRAM Guide

@@ -78,9 +83,11 @@ All dependencies (`gdown`, `timm`, `omegaconf`, `easydict`, `yacs`, `einops`, `h
 | 48 GB+ | `batch_size=4–16, all_on_gpu=true` |
 | 96 GB+ | `batch_size=8–16, all_on_gpu=true, chunk_size=0` |

+SPEED generally fits the 24 GB tier at HD resolutions. LDF-VFI is a separate workload: its official 8x quick start requires about 20 GB, and higher resolutions may require smaller VAE tiles or more VRAM.
+
 ## Nodes

-All Interpolate nodes share a common set of controls:
+The pairwise Interpolate nodes (BIM/EMA/SGM/GIMM/SPEED) share these controls:

 | Input | Description |
 |-------|-------------|
@@ -93,7 +100,7 @@ All Interpolate nodes share a common set of controls:
 | **all_on_gpu** | Keep all intermediate frames on GPU (fast, needs large VRAM) |
 | **clear_cache_after_n_frames** | Clear CUDA cache every N pairs to prevent VRAM buildup |
 | **source_fps** | Input frame rate. Required when target_fps > 0 |
-| **target_fps** | Target output FPS. When > 0, overrides multiplier — auto-computes the optimal power-of-2 oversample then selects frames at exact target timestamps. 0 = use multiplier |
+| **target_fps** | Target output FPS. When > 0, overrides multiplier — auto-computes a power-of-2 oversample up to 8x, then selects the nearest generated frame for each target timestamp. 0 = use multiplier |

 | Output | Description |
 |--------|-------------|
@@ -202,14 +209,65 @@ Same pattern as other Segment nodes.

 </details>

+<details>
+<summary><strong>SPEED</strong></summary>
+
+#### Load SPEED Model
+
+Downloads the official `speed.pt` checkpoint from [zhZ524/SPEED](https://huggingface.co/zhZ524/SPEED) to `ComfyUI/models/speed-vfi/`. The loader also fetches a checksum-pinned snapshot of the official runtime on first use; Tween does not bundle that source.
+
+| Input | Description |
+|-------|-------------|
+| **model_path** | Checkpoint from `models/speed-vfi/` (official default is ~447 MB) |
+| **precision** | `auto` prefers BF16, then FP16; FP32 is available for comparison |
+
+#### SPEED Interpolate / Segment Interpolate
+
+Uses the same batching, chunking, segment, and exact-target-FPS controls as BIM-VFI, plus a `seed` input for repeatable starting pixel noise. Keeping the seed on the interpolation node lets it change without reloading the model. SPEED is repeatable for the same seed and execution settings; changing batch, chunk, or segment boundaries can change how its stochastic noise is assigned. The released model predicts only the midpoint, so 4x and 8x are recursive passes. Inputs are padded to the model's 64-pixel divisor and cropped back automatically.
+
+</details>
+
+<details>
+<summary><strong>LDF-VFI</strong></summary>
+
+#### Load LDF-VFI Model
+
+Downloads the official transformer and conditional VAE from [onecat-ai/LDF-VFI](https://huggingface.co/onecat-ai/LDF-VFI) to `ComfyUI/models/ldf-vfi/` (~6.4 GB total). A checksum-pinned Apache-2.0 runtime snapshot is fetched on first use. Loading stays on CPU until the interpolation node executes.
+
+| Input | Description |
+|-------|-------------|
+| **tile_size / tile_overlap** | Spatial VAE tiling and seam blending; default 256/64 |
+| **vae_batch_size** | Lower first if VAE encode/decode runs out of VRAM |
+| **attention_type** | Official `slide_chunk_all_block_2x1x1` sparse attention is recommended |
+
+#### LDF-VFI Sequence Interpolate
+
+LDF-VFI is not a pairwise node. It processes the ordered source batch with the paper's skip-concat autoregressive sampler and internally chunks long sequences without breaking temporal context.
+
+| Input | Description |
+|-------|-------------|
+| **temporal_factor** | Any integer from 2x through 16x |
+| **sampling_steps** | Diffusion steps per temporal block; official quick start uses 16 |
+| **t_shift / t_cond** | Official defaults are 8.0 / 0.1 |
+| **seed** | Repeatable VAE and diffusion sampling |
+| **offload_after** | Return transformer and VAE to CPU after generation |
+| **source_fps / target_fps** | Optional exact-FPS selection using the smallest sufficient native factor |
+
+The second output, `generated_sequence`, is the full native-factor sequence before exact-FPS selection. LDF has no Segment node because externally splitting the sequence would discard the long-range context it is designed to preserve.
+
+</details>
+
 ### Tween Concat Videos

-Concatenates segment video files into a single video using ffmpeg. Connect from any Segment Interpolate's model output to ensure it runs after all segments are saved. Works with all four models.
+Concatenates segment video files into a single video using ffmpeg. Connect from any pairwise Segment Interpolate's model output to ensure it runs after all segments are saved.

 ### Output frame count

- **Multiplier mode:** 2x = 2N-1, 4x = 4N-3, 8x = 8N-7
- **Target FPS mode:** `floor((N-1) / source_fps * target_fps) + 1` frames. Automatically oversamples to the nearest power-of-2 above the ratio, then selects frames at exact target timestamps. Downsampling (target < source) also works — frames are selected from the input with no model calls.
+- **Pairwise multiplier mode:** 2x = 2N-1, 4x = 4N-3, 8x = 8N-7
+- **LDF-VFI native factor:** factor `F` = `F(N-1)+1`, for any integer `F` from 2 through 16
+- **Target FPS mode:** `floor((N-1) / source_fps * target_fps) + 1` frames. Pairwise nodes oversample to the nearest power-of-2 above the ratio (up to 8x), then select the nearest generated frame for each target timestamp. Downsampling (target < source) also works — frames are selected from the input with no model calls. LDF-VFI supports native factors up to 16x.
+
+In target-FPS Segment mode, a very small `segment_size` can cover less than one output-frame interval while downsampling. Increase `segment_size` if the node reports that the segment contains no target timestamps; returning a placeholder frame would make concatenated timing incorrect.

 ## Acknowledgments

@@ -219,8 +277,10 @@ Concatenates segment video files into a single video using ffmpeg. Connect from
 | **EMA-VFI** | Zhang et al. (MCG-NJU) | CVPR 2023 | [Paper](https://arxiv.org/abs/2303.00440) · [Code](https://github.com/MCG-NJU/EMA-VFI) |
 | **SGM-VFI** | Zhang et al. (MCG-NJU) | CVPR 2024 | [Paper](https://arxiv.org/abs/2404.06913) · [Code](https://github.com/MCG-NJU/SGM-VFI) |
 | **GIMM-VFI** | Guo, Li, Loy (S-Lab NTU) | NeurIPS 2024 | [Paper](https://arxiv.org/abs/2407.08680) · [Code](https://github.com/GSeanCDAT/GIMM-VFI) |
+| **SPEED** | Zhang et al. | ACM MM 2026 | [Paper](https://arxiv.org/abs/2607.15585) · [Code](https://github.com/bbldCVer/SPEED) · [Model](https://huggingface.co/zhZ524/SPEED) |
+| **LDF-VFI** | Peng et al. | CVPR 2026 | [Paper](https://arxiv.org/abs/2601.14959) · [Code](https://github.com/xypeng9903/LDF-VFI) · [Model](https://huggingface.co/onecat-ai/LDF-VFI) |

-GIMM-VFI adaptation from [kijai/ComfyUI-GIMM-VFI](https://github.com/kijai/ComfyUI-GIMM-VFI) with checkpoints from [Kijai/GIMM-VFI_safetensors](https://huggingface.co/Kijai/GIMM-VFI_safetensors). Architecture files in `bim_vfi_arch/`, `ema_vfi_arch/`, `sgm_vfi_arch/`, and `gimm_vfi_arch/` are vendored from their respective repositories with minimal modifications.
+GIMM-VFI adaptation from [kijai/ComfyUI-GIMM-VFI](https://github.com/kijai/ComfyUI-GIMM-VFI) with checkpoints from [Kijai/GIMM-VFI_safetensors](https://huggingface.co/Kijai/GIMM-VFI_safetensors). Architecture files in `bim_vfi_arch/`, `ema_vfi_arch/`, `sgm_vfi_arch/`, and `gimm_vfi_arch/` are vendored from their respective repositories with minimal modifications. SPEED and LDF-VFI use checksum-pinned official source snapshots downloaded into their model directories on demand.

 <details>
 <summary>BibTeX citations</summary>
@@ -253,6 +313,22 @@ GIMM-VFI adaptation from [kijai/ComfyUI-GIMM-VFI](https://github.com/kijai/Comfy
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2024}
 }
+
+@misc{zhang2026speed,
+  title={SPEED: One-Step Pixel Diffusion for High-quality Video Frame Interpolation},
+  author={Zhang, Zihao and Zhao, Haoyu and Yang, Siqian and Wu, Yidi and Jiang, Yudong and Wu, Zuxuan},
+  year={2026},
+  eprint={2607.15585},
+  archivePrefix={arXiv}
+}
+
+@misc{peng2026holistic,
+  title={Towards Holistic Modeling for Video Frame Interpolation with Auto-regressive Diffusion Transformers},
+  author={Peng, Xinyu and Li, Han and Huang, Yuyang and Zheng, Ziyang and Wang, Yaoming and Chen, Xin and Dai, Wenrui and Li, Chenglin and Zou, Junni and Xiong, Hongkai},
+  year={2026},
+  eprint={2601.14959},
+  archivePrefix={arXiv}
+}
 ```

 </details>
@@ -261,6 +337,8 @@ GIMM-VFI adaptation from [kijai/ComfyUI-GIMM-VFI](https://github.com/kijai/Comfy

 **BIM-VFI:** Research and education only. Commercial use requires permission from Prof. Munchurl Kim (mkimee@kaist.ac.kr). See the [original repository](https://github.com/KAIST-VICLab/BiM-VFI).

-**EMA-VFI, SGM-VFI, GIMM-VFI:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0). GIMM-VFI ComfyUI adaptation based on [kijai/ComfyUI-GIMM-VFI](https://github.com/kijai/ComfyUI-GIMM-VFI).
+**EMA-VFI, SGM-VFI, GIMM-VFI, LDF-VFI:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0). GIMM-VFI ComfyUI adaptation based on [kijai/ComfyUI-GIMM-VFI](https://github.com/kijai/ComfyUI-GIMM-VFI).
+
+**SPEED:** The official source repository did not include a license file when this integration was pinned. Tween does not redistribute that source; the loader downloads it directly from the official repository. Review the upstream terms before redistribution or commercial use. The checkpoint is likewise downloaded from its official Hugging Face repository.

 **This wrapper code:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)
@@ -3,6 +3,8 @@ from .nodes import (
    LoadEMAVFIModel, EMAVFIInterpolate, EMAVFISegmentInterpolate,
    LoadSGMVFIModel, SGMVFIInterpolate, SGMVFISegmentInterpolate,
    LoadGIMMVFIModel, GIMMVFIInterpolate, GIMMVFISegmentInterpolate,
+    LoadSPEEDVFIModel, SPEEDVFIInterpolate, SPEEDVFISegmentInterpolate,
+    LoadLDFVFIModel, LDFVFIInterpolate,
    VFIOptimizer,
 )

@@ -20,6 +22,11 @@ NODE_CLASS_MAPPINGS = {
    "LoadGIMMVFIModel": LoadGIMMVFIModel,
    "GIMMVFIInterpolate": GIMMVFIInterpolate,
    "GIMMVFISegmentInterpolate": GIMMVFISegmentInterpolate,
+    "LoadSPEEDVFIModel": LoadSPEEDVFIModel,
+    "SPEEDVFIInterpolate": SPEEDVFIInterpolate,
+    "SPEEDVFISegmentInterpolate": SPEEDVFISegmentInterpolate,
+    "LoadLDFVFIModel": LoadLDFVFIModel,
+    "LDFVFIInterpolate": LDFVFIInterpolate,
    "VFIOptimizer": VFIOptimizer,
 }

@@ -37,5 +44,10 @@ NODE_DISPLAY_NAME_MAPPINGS = {
    "LoadGIMMVFIModel": "Load GIMM-VFI Model",
    "GIMMVFIInterpolate": "GIMM-VFI Interpolate",
    "GIMMVFISegmentInterpolate": "GIMM-VFI Segment Interpolate",
+    "LoadSPEEDVFIModel": "Load SPEED Model",
+    "SPEEDVFIInterpolate": "SPEED Interpolate",
+    "SPEEDVFISegmentInterpolate": "SPEED Segment Interpolate",
+    "LoadLDFVFIModel": "Load LDF-VFI Model",
+    "LDFVFIInterpolate": "LDF-VFI Sequence Interpolate",
    "VFIOptimizer": "VFI Optimizer",
 }
@@ -1,4 +1,4 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 720 320" width="720" height="320">
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 720 475" width="720" height="475">
  <defs>
    <linearGradient id="gQ" x1="0" y1="0" x2="1" y2="0">
      <stop offset="0%" stop-color="#7aa2f7"/><stop offset="100%" stop-color="#7dcfff"/>
@@ -12,13 +12,13 @@
  </defs>

  <!-- Background -->
-  <rect width="720" height="320" rx="16" fill="#0d1117"/>
+  <rect width="720" height="475" rx="16" fill="#0d1117"/>

  <!-- ═══ BIM-VFI (top-left) ═══ -->
  <rect x="10" y="10" width="340" height="145" rx="10" fill="#161b22" stroke="#30363d" stroke-width="1"/>
  <rect x="11" y="22" width="3" height="121" fill="#3fb950"/>
  <text x="30" y="38" fill="#e6edf3" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="15" font-weight="600">BIM-VFI</text>
-  <text x="30" y="56" fill="#3fb950" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="11">&#9733; Recommended &#183; Best quality &#183; CVPR 2025</text>
+  <text x="30" y="56" fill="#3fb950" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="11">Strong pairwise quality &#183; CVPR 2025</text>
  <line x1="30" y1="64" x2="330" y2="64" stroke="#30363d" stroke-width="0.5"/>
  <text x="30" y="82" fill="#7aa2f7" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="11">Quality</text>
  <rect x="88" y="72" width="244" height="11" rx="3" fill="#21262d"/>
@@ -81,4 +81,38 @@
  <rect x="448" y="263" width="244" height="11" rx="3" fill="#21262d"/>
  <rect x="448" y="263" width="146" height="11" rx="3" fill="url(#gV)" opacity="0.85"/>
  <text x="632" y="298" fill="#8b949e" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="10">Apache 2.0</text>
+
+  <!-- ═══ SPEED (third-row left) ═══ -->
+  <rect x="10" y="320" width="340" height="145" rx="10" fill="#161b22" stroke="#30363d" stroke-width="1"/>
+  <rect x="11" y="332" width="3" height="121" fill="#f2cc60"/>
+  <text x="30" y="348" fill="#e6edf3" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="15" font-weight="600">SPEED</text>
+  <text x="30" y="366" fill="#f2cc60" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="11">Modern pairwise default &#183; ACM MM 2026</text>
+  <line x1="30" y1="374" x2="330" y2="374" stroke="#30363d" stroke-width="0.5"/>
+  <text x="30" y="392" fill="#7aa2f7" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="11">Quality</text>
+  <rect x="88" y="382" width="244" height="11" rx="3" fill="#21262d"/>
+  <rect x="88" y="382" width="220" height="11" rx="3" fill="url(#gQ)" opacity="0.85"/>
+  <text x="30" y="410" fill="#9ece6a" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="11">Speed</text>
+  <rect x="88" y="400" width="244" height="11" rx="3" fill="#21262d"/>
+  <rect x="88" y="400" width="190" height="11" rx="3" fill="url(#gS)" opacity="0.85"/>
+  <text x="30" y="428" fill="#bb9af7" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="11">VRAM</text>
+  <rect x="88" y="418" width="244" height="11" rx="3" fill="#21262d"/>
+  <rect x="88" y="418" width="175" height="11" rx="3" fill="url(#gV)" opacity="0.85"/>
+  <text x="253" y="453" fill="#f0883e" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="10">Check upstream terms</text>
+
+  <!-- ═══ LDF-VFI (third-row right) ═══ -->
+  <rect x="370" y="320" width="340" height="145" rx="10" fill="#161b22" stroke="#30363d" stroke-width="1"/>
+  <rect x="371" y="332" width="3" height="121" fill="#ff7b72"/>
+  <text x="390" y="348" fill="#e6edf3" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="15" font-weight="600">LDF-VFI</text>
+  <text x="390" y="366" fill="#ff7b72" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="11">Long-sequence coherence &#183; CVPR 2026</text>
+  <line x1="390" y1="374" x2="690" y2="374" stroke="#30363d" stroke-width="0.5"/>
+  <text x="390" y="392" fill="#7aa2f7" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="11">Quality</text>
+  <rect x="448" y="382" width="244" height="11" rx="3" fill="#21262d"/>
+  <rect x="448" y="382" width="230" height="11" rx="3" fill="url(#gQ)" opacity="0.85"/>
+  <text x="390" y="410" fill="#9ece6a" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="11">Speed</text>
+  <rect x="448" y="400" width="244" height="11" rx="3" fill="#21262d"/>
+  <rect x="448" y="400" width="48" height="11" rx="3" fill="url(#gS)" opacity="0.85"/>
+  <text x="390" y="428" fill="#bb9af7" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="11">VRAM</text>
+  <rect x="448" y="418" width="244" height="11" rx="3" fill="#21262d"/>
+  <rect x="448" y="418" width="40" height="11" rx="3" fill="url(#gV)" opacity="0.85"/>
+  <text x="632" y="453" fill="#8b949e" font-family="-apple-system,BlinkMacSystemFont,'Segoe UI','Noto Sans',Helvetica,Arial,sans-serif" font-size="10">Apache 2.0</text>
 </svg>
@@ -0,0 +1,757 @@
+{
+  "id": "d1b1b840-08a9-4f93-9a1a-5bb9bdb05c39",
+  "revision": 0,
+  "last_node_id": 13,
+  "last_link_id": 11,
+  "nodes": [
+    {
+      "id": 1,
+      "type": "Note",
+      "pos": [
+        0,
+        0
+      ],
+      "size": [
+        1420,
+        112
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "Note"
+      },
+      "widgets_values": [
+        "TWEEN MODEL LAB — SPEED + LDF-VFI\n1) Choose a short input video below.  2) Queue Prompt to run the enabled SPEED path.  3) The LDF-VFI path is muted by default; enable its three coral nodes only when you want the heavier sequence model."
+      ],
+      "color": "#1f3a4d",
+      "bgcolor": "#294d66"
+    },
+    {
+      "id": 2,
+      "type": "VHS_LoadVideoPath",
+      "pos": [
+        20,
+        190
+      ],
+      "size": [
+        300,
+        310
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "meta_batch",
+          "shape": 7,
+          "type": "VHS_BatchManager",
+          "link": null
+        },
+        {
+          "name": "vae",
+          "shape": 7,
+          "type": "VAE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            1,
+            8
+          ]
+        },
+        {
+          "name": "frame_count",
+          "type": "INT",
+          "links": null
+        },
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "links": [
+            7,
+            11
+          ]
+        },
+        {
+          "name": "video_info",
+          "type": "VHS_VIDEOINFO",
+          "links": null
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-videohelpersuite",
+        "Node name for S&R": "VHS_LoadVideoPath"
+      },
+      "widgets_values": {
+        "video": "",
+        "force_rate": 24,
+        "custom_width": 0,
+        "custom_height": 0,
+        "frame_load_cap": 25,
+        "skip_first_frames": 0,
+        "select_every_nth": 1,
+        "format": "AnimateDiff",
+        "videopreview": {
+          "hidden": false,
+          "paused": false,
+          "params": {
+            "filename": "",
+            "type": "path",
+            "format": "video/",
+            "force_rate": 24,
+            "custom_width": 0,
+            "custom_height": 0,
+            "frame_load_cap": 25,
+            "skip_first_frames": 0,
+            "select_every_nth": 1
+          }
+        }
+      }
+    },
+    {
+      "id": 3,
+      "type": "Note",
+      "pos": [
+        20,
+        520
+      ],
+      "size": [
+        300,
+        166
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "Note"
+      },
+      "widgets_values": [
+        "INPUT SAFETY DEFAULTS\n• force_rate = 24 FPS\n• frame_load_cap = 25 (~1 second)\n• dimensions = source size\n\nFor another source rate, change force_rate here and source_fps in the chosen Tween node. The Video Combine frame rate must match target_fps."
+      ],
+      "color": "#3f3520",
+      "bgcolor": "#594b2d"
+    },
+    {
+      "id": 4,
+      "type": "LoadSPEEDVFIModel",
+      "pos": [
+        390,
+        170
+      ],
+      "size": [
+        300,
+        105
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "model",
+          "type": "SPEED_VFI_MODEL",
+          "links": [
+            2,
+            4
+          ]
+        }
+      ],
+      "properties": {
+        "aux_id": "ComfyUI-Tween.git",
+        "Node name for S&R": "LoadSPEEDVFIModel"
+      },
+      "widgets_values": [
+        "speed.pt",
+        "auto"
+      ],
+      "color": "#243b32",
+      "bgcolor": "#315244"
+    },
+    {
+      "id": 5,
+      "type": "VFIOptimizer",
+      "pos": [
+        390,
+        330
+      ],
+      "size": [
+        310,
+        150
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 1
+        },
+        {
+          "name": "model",
+          "type": "*",
+          "link": 2
+        }
+      ],
+      "outputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": [
+            3
+          ]
+        },
+        {
+          "name": "settings",
+          "type": "VFI_SETTINGS",
+          "links": [
+            5
+          ]
+        }
+      ],
+      "properties": {
+        "aux_id": "ComfyUI-Tween.git",
+        "Node name for S&R": "VFIOptimizer"
+      },
+      "widgets_values": [
+        2,
+        0
+      ],
+      "color": "#243b32",
+      "bgcolor": "#315244"
+    },
+    {
+      "id": 6,
+      "type": "SPEEDVFIInterpolate",
+      "pos": [
+        760,
+        240
+      ],
+      "size": [
+        355,
+        340
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 3
+        },
+        {
+          "name": "model",
+          "type": "SPEED_VFI_MODEL",
+          "link": 4
+        },
+        {
+          "name": "settings",
+          "shape": 7,
+          "type": "VFI_SETTINGS",
+          "link": 5
+        }
+      ],
+      "outputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": [
+            6
+          ]
+        },
+        {
+          "name": "oversampled",
+          "type": "IMAGE",
+          "links": null
+        }
+      ],
+      "properties": {
+        "aux_id": "ComfyUI-Tween.git",
+        "Node name for S&R": "SPEEDVFIInterpolate"
+      },
+      "widgets_values": [
+        2,
+        10,
+        true,
+        false,
+        1,
+        0,
+        24,
+        48,
+        0
+      ],
+      "color": "#243b32",
+      "bgcolor": "#315244"
+    },
+    {
+      "id": 7,
+      "type": "VHS_VideoCombine",
+      "pos": [
+        1170,
+        240
+      ],
+      "size": [
+        360,
+        340
+      ],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 6
+        },
+        {
+          "name": "audio",
+          "shape": 7,
+          "type": "AUDIO",
+          "link": 7
+        },
+        {
+          "name": "meta_batch",
+          "shape": 7,
+          "type": "VHS_BatchManager",
+          "link": null
+        },
+        {
+          "name": "vae",
+          "shape": 7,
+          "type": "VAE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "Filenames",
+          "type": "VHS_FILENAMES",
+          "links": null
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-videohelpersuite",
+        "Node name for S&R": "VHS_VideoCombine"
+      },
+      "widgets_values": {
+        "frame_rate": 48,
+        "loop_count": 0,
+        "filename_prefix": "Tween/demo_speed_24_to_48",
+        "format": "video/h264-mp4",
+        "pix_fmt": "yuv420p",
+        "crf": 19,
+        "save_metadata": true,
+        "trim_to_audio": false,
+        "pingpong": false,
+        "save_output": true,
+        "videopreview": {
+          "hidden": false,
+          "paused": false,
+          "params": {}
+        }
+      },
+      "color": "#243b32",
+      "bgcolor": "#315244"
+    },
+    {
+      "id": 8,
+      "type": "Note",
+      "pos": [
+        390,
+        20
+      ],
+      "size": [
+        1140,
+        110
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "Note"
+      },
+      "widgets_values": [
+        "SPEED — RECOMMENDED, ENABLED\nOne-step diffusion midpoint interpolation. This branch uses target-FPS mode (24→48), a fixed seed, and VFI Optimizer settings. First use downloads the official ~447 MB checkpoint and checksum-pinned runtime."
+      ],
+      "color": "#234936",
+      "bgcolor": "#2d6046"
+    },
+    {
+      "id": 9,
+      "type": "Note",
+      "pos": [
+        390,
+        760
+      ],
+      "size": [
+        1140,
+        125
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "Note"
+      },
+      "widgets_values": [
+        "LDF-VFI — ADVANCED, MUTED BY DEFAULT\nRequires an NVIDIA BF16 GPU and roughly 20 GB VRAM; first use downloads ~6.4 GB. To run it, select the three coral nodes below and set Mode → Always. Disable/mute the SPEED output node if you only want LDF. Use 16 sampling steps for the official quality-oriented quick-start setting."
+      ],
+      "color": "#5a2d2d",
+      "bgcolor": "#743b3b"
+    },
+    {
+      "id": 10,
+      "type": "LoadLDFVFIModel",
+      "pos": [
+        390,
+        930
+      ],
+      "size": [
+        340,
+        190
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 2,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "model",
+          "type": "LDF_VFI_MODEL",
+          "links": [
+            9
+          ]
+        }
+      ],
+      "properties": {
+        "aux_id": "ComfyUI-Tween.git",
+        "Node name for S&R": "LoadLDFVFIModel"
+      },
+      "widgets_values": [
+        "onecat-ai/LDF-VFI",
+        256,
+        64,
+        4,
+        "slide_chunk_all_block_2x1x1"
+      ],
+      "color": "#5a2d2d",
+      "bgcolor": "#743b3b"
+    },
+    {
+      "id": 11,
+      "type": "LDFVFIInterpolate",
+      "pos": [
+        790,
+        930
+      ],
+      "size": [
+        370,
+        320
+      ],
+      "flags": {},
+      "order": 10,
+      "mode": 2,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 8
+        },
+        {
+          "name": "model",
+          "type": "LDF_VFI_MODEL",
+          "link": 9
+        }
+      ],
+      "outputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": [
+            10
+          ]
+        },
+        {
+          "name": "generated_sequence",
+          "type": "IMAGE",
+          "links": null
+        }
+      ],
+      "properties": {
+        "aux_id": "ComfyUI-Tween.git",
+        "Node name for S&R": "LDFVFIInterpolate"
+      },
+      "widgets_values": [
+        2,
+        8,
+        8,
+        0.1,
+        42,
+        true,
+        24,
+        48
+      ],
+      "color": "#5a2d2d",
+      "bgcolor": "#743b3b"
+    },
+    {
+      "id": 12,
+      "type": "VHS_VideoCombine",
+      "pos": [
+        1210,
+        930
+      ],
+      "size": [
+        360,
+        340
+      ],
+      "flags": {},
+      "order": 11,
+      "mode": 2,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 10
+        },
+        {
+          "name": "audio",
+          "shape": 7,
+          "type": "AUDIO",
+          "link": 11
+        },
+        {
+          "name": "meta_batch",
+          "shape": 7,
+          "type": "VHS_BatchManager",
+          "link": null
+        },
+        {
+          "name": "vae",
+          "shape": 7,
+          "type": "VAE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "Filenames",
+          "type": "VHS_FILENAMES",
+          "links": null
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-videohelpersuite",
+        "Node name for S&R": "VHS_VideoCombine"
+      },
+      "widgets_values": {
+        "frame_rate": 48,
+        "loop_count": 0,
+        "filename_prefix": "Tween/demo_ldf_24_to_48",
+        "format": "video/h264-mp4",
+        "pix_fmt": "yuv420p",
+        "crf": 19,
+        "save_metadata": true,
+        "trim_to_audio": false,
+        "pingpong": false,
+        "save_output": true,
+        "videopreview": {
+          "hidden": false,
+          "paused": false,
+          "params": {}
+        }
+      },
+      "color": "#5a2d2d",
+      "bgcolor": "#743b3b"
+    },
+    {
+      "id": 13,
+      "type": "Note",
+      "pos": [
+        20,
+        930
+      ],
+      "size": [
+        300,
+        236
+      ],
+      "flags": {},
+      "order": 12,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "Note"
+      },
+      "widgets_values": [
+        "WHY TWO PATHS?\n\nSPEED\n• Best default for normal pairwise VFI\n• Much smaller and faster\n• 2x midpoint; 4x/8x recursive\n\nLDF-VFI\n• Sequence-native temporal modeling\n• Better suited to long-range coherence\n• 2x–16x, but far heavier\n\nThe LDF demo uses 8 steps for a quicker test. Raise to 16 after the graph is working."
+      ],
+      "color": "#40303b",
+      "bgcolor": "#594353"
+    }
+  ],
+  "links": [
+    [
+      1,
+      2,
+      0,
+      5,
+      0,
+      "IMAGE"
+    ],
+    [
+      2,
+      4,
+      0,
+      5,
+      1,
+      "*"
+    ],
+    [
+      3,
+      5,
+      0,
+      6,
+      0,
+      "IMAGE"
+    ],
+    [
+      4,
+      4,
+      0,
+      6,
+      1,
+      "SPEED_VFI_MODEL"
+    ],
+    [
+      5,
+      5,
+      1,
+      6,
+      2,
+      "VFI_SETTINGS"
+    ],
+    [
+      6,
+      6,
+      0,
+      7,
+      0,
+      "IMAGE"
+    ],
+    [
+      7,
+      2,
+      2,
+      7,
+      1,
+      "AUDIO"
+    ],
+    [
+      8,
+      2,
+      0,
+      11,
+      0,
+      "IMAGE"
+    ],
+    [
+      9,
+      10,
+      0,
+      11,
+      1,
+      "LDF_VFI_MODEL"
+    ],
+    [
+      10,
+      11,
+      0,
+      12,
+      0,
+      "IMAGE"
+    ],
+    [
+      11,
+      2,
+      2,
+      12,
+      1,
+      "AUDIO"
+    ]
+  ],
+  "groups": [
+    {
+      "id": 1,
+      "title": "1 · INPUT CLIP",
+      "bounding": [
+        -20,
+        135,
+        360,
+        575
+      ],
+      "color": "#8a6d2f",
+      "font_size": 24,
+      "flags": {}
+    },
+    {
+      "id": 2,
+      "title": "2 · SPEED — RECOMMENDED / ENABLED",
+      "bounding": [
+        350,
+        -20,
+        1240,
+        650
+      ],
+      "color": "#347a56",
+      "font_size": 24,
+      "flags": {}
+    },
+    {
+      "id": 3,
+      "title": "3 · LDF-VFI — ADVANCED / MUTED",
+      "bounding": [
+        350,
+        730,
+        1240,
+        590
+      ],
+      "color": "#8a4646",
+      "font_size": 24,
+      "flags": {}
+    }
+  ],
+  "config": {},
+  "extra": {
+    "workflowRendererVersion": "LG",
+    "ue_links": [],
+    "links_added_by_ue": [],
+    "ds": {
+      "scale": 0.76,
+      "offset": [
+        80,
+        80
+      ]
+    },
+    "frontendVersion": "1.45.19",
+    "VHS_latentpreview": true,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true
+  },
+  "version": 0.4
+}
@@ -0,0 +1,173 @@
+"""Pinned, lazy installers for optional upstream model runtimes.
+
+Tween does not redistribute these projects.  Their official source archives are
+downloaded only when a corresponding loader node is executed, verified against
+a pinned SHA-256 digest, and kept next to that model's checkpoints.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import logging
+import os
+from pathlib import Path
+import shutil
+import tempfile
+import threading
+import urllib.error
+import urllib.request
+import zipfile
+
+
+logger = logging.getLogger("Tween")
+
+
+UPSTREAM_SOURCES = {
+    "speed": {
+        "project": "SPEED",
+        "commit": "40fadbe85c88cc6e4015062389da464fd7e85ab9",
+        "url": (
+            "https://codeload.github.com/bbldCVer/SPEED/zip/"
+            "40fadbe85c88cc6e4015062389da464fd7e85ab9"
+        ),
+        "sha256": "9e9cc71bfeaf7a62008950b8f234f5f035df27b65a5fc0464caee2542f47f68c",
+        "required": "src/models/model.py",
+    },
+    "ldf": {
+        "project": "LDF-VFI",
+        "commit": "61b34d2379df8a313e8e4cb467cc2f74c52b45d7",
+        "url": (
+            "https://codeload.github.com/xypeng9903/LDF-VFI/zip/"
+            "61b34d2379df8a313e8e4cb467cc2f74c52b45d7"
+        ),
+        "sha256": "3a903aeb5353c7e5eb932f129d975d1750246502937d8f7b283b61a269c23668",
+        "required": "training/models/precond.py",
+    },
+}
+_SOURCE_LOCKS = {name: threading.Lock() for name in UPSTREAM_SOURCES}
+
+
+def _download(url: str, destination: Path) -> None:
+    request = urllib.request.Request(url, headers={"User-Agent": "ComfyUI-Tween"})
+    try:
+        with urllib.request.urlopen(request, timeout=60) as response, destination.open("wb") as output:
+            shutil.copyfileobj(response, output, length=1024 * 1024)
+    except (OSError, urllib.error.URLError) as exc:
+        raise RuntimeError(
+            f"Could not download optional upstream runtime from {url}. "
+            "Check network access and retry the loader node."
+        ) from exc
+
+
+def _sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as stream:
+        for chunk in iter(lambda: stream.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def _safe_extract(archive: Path, destination: Path) -> Path:
+    with zipfile.ZipFile(archive) as source_zip:
+        members = source_zip.infolist()
+        if not members:
+            raise RuntimeError(f"Downloaded source archive is empty: {archive}")
+
+        destination_resolved = destination.resolve()
+        for member in members:
+            member_path = (destination / member.filename).resolve()
+            if os.path.commonpath((destination_resolved, member_path)) != str(destination_resolved):
+                raise RuntimeError(f"Unsafe path in source archive: {member.filename}")
+        source_zip.extractall(destination)
+
+    top_level = {Path(member.filename).parts[0] for member in members if member.filename}
+    if len(top_level) != 1:
+        raise RuntimeError("Expected one top-level directory in the upstream source archive")
+    return destination / top_level.pop()
+
+
+def _ensure_upstream_source_unlocked(name: str, model_dir: str | os.PathLike[str]) -> str:
+    try:
+        spec = UPSTREAM_SOURCES[name]
+    except KeyError as exc:
+        raise ValueError(f"Unknown Tween upstream source: {name}") from exc
+
+    model_root = Path(model_dir)
+    source_dir = model_root / "_upstream"
+    required_file = source_dir / spec["required"]
+    if required_file.is_file():
+        marker_path = source_dir / ".tween-source.json"
+        if marker_path.is_file():
+            try:
+                marker = json.loads(marker_path.read_text(encoding="utf-8"))
+            except (OSError, json.JSONDecodeError) as exc:
+                raise RuntimeError(f"Invalid source marker: {marker_path}") from exc
+            if (
+                marker.get("commit") != spec["commit"]
+                or marker.get("archive_sha256") != spec["sha256"]
+            ):
+                raise RuntimeError(
+                    f"{spec['project']} runtime at {source_dir} is pinned to a different commit. "
+                    "Remove _upstream and run the loader again."
+                )
+        else:
+            logger.warning(
+                "Using manually installed %s runtime at %s (no Tween verification marker)",
+                spec["project"], source_dir,
+            )
+        return str(source_dir)
+
+    if source_dir.exists():
+        raise RuntimeError(
+            f"Incomplete {spec['project']} runtime at {source_dir}. "
+            "Remove that _upstream directory and run the loader again."
+        )
+
+    model_root.mkdir(parents=True, exist_ok=True)
+    logger.info(
+        "Downloading pinned %s runtime (%s) to %s",
+        spec["project"], spec["commit"][:12], source_dir,
+    )
+
+    with tempfile.TemporaryDirectory(prefix="tween-source-", dir=model_root) as temp_name:
+        temp_dir = Path(temp_name)
+        archive = temp_dir / "source.zip"
+        _download(spec["url"], archive)
+
+        actual_digest = _sha256(archive)
+        if actual_digest != spec["sha256"]:
+            raise RuntimeError(
+                f"Checksum mismatch for {spec['project']} source archive: "
+                f"expected {spec['sha256']}, got {actual_digest}"
+            )
+
+        extracted = _safe_extract(archive, temp_dir / "extract")
+        if not (extracted / spec["required"]).is_file():
+            raise RuntimeError(
+                f"The {spec['project']} archive does not contain {spec['required']}"
+            )
+
+        marker = {
+            "project": spec["project"],
+            "commit": spec["commit"],
+            "archive_sha256": spec["sha256"],
+            "source_url": spec["url"],
+        }
+        (extracted / ".tween-source.json").write_text(
+            json.dumps(marker, indent=2) + "\n", encoding="utf-8"
+        )
+        shutil.move(str(extracted), str(source_dir))
+
+    logger.info("Installed %s runtime at %s", spec["project"], source_dir)
+    return str(source_dir)
+
+
+def ensure_upstream_source(name: str, model_dir: str | os.PathLike[str]) -> str:
+    """Return a verified upstream checkout, downloading it once per process."""
+    try:
+        source_lock = _SOURCE_LOCKS[name]
+    except KeyError as exc:
+        raise ValueError(f"Unknown Tween upstream source: {name}") from exc
+    with source_lock:
+        return _ensure_upstream_source_unlocked(name, model_dir)
@@ -0,0 +1,437 @@
+"""Sequence-native ComfyUI adapter for the official Apache-2.0 LDF-VFI runtime."""
+
+from __future__ import annotations
+
+import importlib
+import logging
+from pathlib import Path
+import sys
+import threading
+import types
+
+from einops import rearrange, repeat
+import torch
+import torch.nn.functional as F
+
+
+logger = logging.getLogger("Tween")
+_LDF_NAMESPACE = "_tween_ldf_upstream"
+_LDF_IMPORT_LOCK = threading.RLock()
+
+
+def _cuda_bf16_supported(device: torch.device) -> bool:
+    if device.type != "cuda":
+        return False
+    with torch.cuda.device(device):
+        return torch.cuda.is_bf16_supported()
+
+
+def _namespace_package(name: str, path: Path):
+    package = sys.modules.get(name)
+    if package is not None:
+        return package
+    package = types.ModuleType(name)
+    package.__path__ = [str(path)]
+    package.__package__ = name
+    sys.modules[name] = package
+    return package
+
+
+def load_ldf_runtime(source_root: str):
+    """Load LDF under an isolated namespace without polluting ``training``."""
+    root = Path(source_root).resolve()
+    required = root / "training" / "models" / "precond.py"
+    if not required.is_file():
+        raise RuntimeError(f"Invalid LDF-VFI source directory: missing {required}")
+
+    with _LDF_IMPORT_LOCK:
+        cached = sys.modules.get(f"{_LDF_NAMESPACE}.models.precond")
+        if cached is not None:
+            transformer = importlib.import_module(f"{_LDF_NAMESPACE}.models.transformer_wan")
+            return {
+                "Precond": cached.Precond,
+                "ConditionalVAE": cached.Wan2_1SpatialTiledConditionEncoder3Dv2,
+                "MaskEncoder": cached.MaskSpatialTiledEncoder3D,
+                "Transformer": transformer.WanTransformer3DModel,
+            }
+
+        training_root = _namespace_package(_LDF_NAMESPACE, root / "training")
+        saved_training_modules = {
+            name: module for name, module in tuple(sys.modules.items())
+            if name == "training" or name.startswith("training.")
+        }
+        for name in saved_training_modules:
+            sys.modules.pop(name, None)
+        # One upstream transformer import is absolute (training.distributed.util).
+        # Temporarily alias only while importing, then restore the host process.
+        sys.modules["training"] = training_root
+        try:
+            precond = importlib.import_module(f"{_LDF_NAMESPACE}.models.precond")
+            transformer = importlib.import_module(f"{_LDF_NAMESPACE}.models.transformer_wan")
+        except (ImportError, AttributeError) as exc:
+            for name in tuple(sys.modules):
+                if name == _LDF_NAMESPACE or name.startswith(f"{_LDF_NAMESPACE}."):
+                    sys.modules.pop(name, None)
+            raise RuntimeError(
+                "LDF-VFI requires PyTorch 2.5+ and diffusers 0.33+. "
+                "Install Tween's current requirements and restart ComfyUI."
+            ) from exc
+        finally:
+            for name in tuple(sys.modules):
+                if name == "training" or name.startswith("training."):
+                    sys.modules.pop(name, None)
+            sys.modules.update(saved_training_modules)
+
+    return {
+        "Precond": precond.Precond,
+        "ConditionalVAE": precond.Wan2_1SpatialTiledConditionEncoder3Dv2,
+        "MaskEncoder": precond.MaskSpatialTiledEncoder3D,
+        "Transformer": transformer.WanTransformer3DModel,
+    }
+
+
+def _upsample_nearest(frames: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
+    """Expand sparse source frames to every temporal position in a window."""
+    kept_indices = torch.where(mask)[0]
+    if kept_indices.numel() == 0:
+        raise RuntimeError("LDF-VFI received a temporal window with no source frames")
+    positions = torch.arange(mask.shape[0], device=mask.device)
+    nearest = (positions[:, None] - kept_indices[None, :]).abs().argmin(dim=1)
+    return frames[nearest]
+
+
+class LDFVFIModel:
+    """Long-sequence diffusion interpolation using LDF's skip-concat sampler."""
+
+    TRAIN_FRAMES = 60
+    TILE_TIME = 20
+    CONDITION_TILES = 1
+
+    def __init__(
+        self,
+        model_root: str,
+        vae_path: str,
+        source_root: str,
+        tile_size: int = 256,
+        tile_overlap: int = 64,
+        vae_batch_size: int = 8,
+        attention_type: str = "slide_chunk_all_block_2x1x1",
+    ):
+        if tile_size % 8 or tile_overlap % 8:
+            raise ValueError("LDF-VFI tile size and overlap must be divisible by 8")
+        if tile_overlap >= tile_size:
+            raise ValueError("LDF-VFI tile overlap must be smaller than tile size")
+
+        runtime = load_ldf_runtime(source_root)
+        self.device = "cpu"
+        self.dtype = torch.bfloat16
+        self.vae_path = vae_path
+        self._ConditionalVAE = runtime["ConditionalVAE"]
+        self._MaskEncoder = runtime["MaskEncoder"]
+        self._Precond = runtime["Precond"]
+
+        logger.info("Loading LDF-VFI transformer from %s", model_root)
+        try:
+            transformer = runtime["Transformer"].from_pretrained(
+                model_root,
+                subfolder="transformer",
+                torch_dtype=self.dtype,
+                low_cpu_mem_usage=True,
+            )
+        except TypeError:
+            transformer = runtime["Transformer"].from_pretrained(
+                model_root, subfolder="transformer", torch_dtype=self.dtype
+            )
+        transformer.set_attention_type(attention_type)
+        transformer.requires_grad_(False).eval()
+
+        stride = tile_size - tile_overlap
+        tiled_kwargs = {
+            "tile_sample_min_height": tile_size,
+            "tile_sample_min_width": tile_size,
+            "tile_sample_min_time": self.TILE_TIME,
+            "tile_sample_stride_height": stride,
+            "tile_sample_stride_width": stride,
+            "spatial_compression_ratio": 8,
+            "temporal_compression_ratio": 4,
+        }
+        # The same conditional VAE can encode conditions and decode predictions;
+        # sharing it avoids loading a second ~800 MB copy as the reference CLI does.
+        self.vae = self._ConditionalVAE(vae_path, vae_batch_size, **tiled_kwargs)
+        self.mask_encoder = self._MaskEncoder(**tiled_kwargs)
+        self.model = self._Precond(
+            transformer=transformer,
+            vae=self.vae,
+            lq_encoder=self.vae,
+            msk_encoder=self.mask_encoder,
+        )
+        self.model.requires_grad_(False).eval()
+
+    @property
+    def transformer(self):
+        return self.model.transformer
+
+    def _move_auxiliary_models(self, device: torch.device) -> None:
+        self.mask_encoder.mask_encoder.to(device=device, dtype=self.dtype)
+        if self.vae.vae is None:
+            if device.type != "cpu":
+                self.vae.init(device)
+            return
+        self.vae.vae.to(device=device, dtype=self.dtype)
+        if hasattr(self.vae, "mean") and hasattr(self.vae, "std"):
+            self.vae.mean = self.vae.mean.to(device=device, dtype=self.dtype)
+            self.vae.std = self.vae.std.to(device=device, dtype=self.dtype)
+            self.vae.scale = [self.vae.mean, 1.0 / self.vae.std]
+
+    def to(self, device):
+        target = torch.device(device)
+        if target.type == "cuda" and not _cuda_bf16_supported(target):
+            raise RuntimeError("LDF-VFI requires a CUDA GPU with BF16 support (Ampere or newer)")
+        self.transformer.to(device=target, dtype=self.dtype)
+        self._move_auxiliary_models(target)
+        self.device = str(target)
+        return self
+
+    def clear_cache(self) -> None:
+        for name in (
+            "_swin_attention_mask",
+            "_sliding_chunk_attention_mask",
+            "_sliding_window_attention_mask",
+        ):
+            method = getattr(self.transformer, name, None)
+            cache_clear = getattr(method, "cache_clear", None)
+            if cache_clear is not None:
+                cache_clear()
+
+    def _prepare_condition(self, frames, mask, device):
+        if mask.shape[0] > self.TRAIN_FRAMES:
+            raise ValueError("Internal LDF temporal window exceeds the training window")
+        if mask.shape[0] < self.TRAIN_FRAMES:
+            mask = F.pad(mask, (0, self.TRAIN_FRAMES - mask.shape[0]))
+        dense = _upsample_nearest(frames, mask)
+        dense = rearrange(dense, "t c h w -> 1 c t h w")
+        dense = dense.to(device=device, dtype=self.dtype, non_blocking=True).mul(2).sub(1)
+        dense_mask = repeat(
+            mask, "t -> 1 1 t h w", h=dense.shape[-2], w=dense.shape[-1]
+        ).to(device=device, dtype=self.dtype)
+        condition = self.vae.encode(dense, for_train=True)
+        encoded_mask = self.mask_encoder.encode(dense_mask, for_train=True)
+        return dense, dense_mask, condition, encoded_mask
+
+    def _time_schedule(self, num_steps: int, t_shift: float) -> torch.Tensor:
+        schedule = torch.linspace(1.0, 0.0, steps=num_steps + 1)
+        return t_shift * schedule / (1 + (t_shift - 1) * schedule)
+
+    def _predict_step(self, latent, timestep, condition, encoded_mask):
+        return self.model.predict_v(latent, timestep, condition, encoded_mask)
+
+    def _sample_free(self, condition, encoded_mask, schedule, device, progress):
+        latent = torch.randn_like(condition)
+        for index in range(schedule.shape[0] - 1):
+            progress()
+            timestep = torch.full(
+                condition.shape[:-4], float(schedule[index]), device=device, dtype=self.dtype
+            )
+            velocity = self._predict_step(latent, timestep, condition, encoded_mask)
+            step_size = float(schedule[index + 1] - schedule[index])
+            latent = latent + velocity * step_size
+        return latent
+
+    def _sample_between(self, previous, following, condition, encoded_mask,
+                        schedule, t_cond, device, progress):
+        previous_noisy = previous * (1 - t_cond) + torch.randn_like(previous) * t_cond
+        following_noisy = following * (1 - t_cond) + torch.randn_like(following) * t_cond
+        middle = torch.randn_like(condition[:, self.CONDITION_TILES:-self.CONDITION_TILES])
+        previous_t = torch.full(
+            previous.shape[:-4], t_cond, device=device, dtype=self.dtype
+        )
+        following_t = torch.full(
+            following.shape[:-4], t_cond, device=device, dtype=self.dtype
+        )
+        for index in range(schedule.shape[0] - 1):
+            progress()
+            latent = torch.cat((previous_noisy, middle, following_noisy), dim=1)
+            middle_t = torch.full(
+                middle.shape[:-4], float(schedule[index]), device=device, dtype=self.dtype
+            )
+            timestep = torch.cat((previous_t, middle_t, following_t), dim=1)
+            velocity = self._predict_step(
+                latent, timestep, condition, encoded_mask
+            )[:, self.CONDITION_TILES:-self.CONDITION_TILES]
+            step_size = float(schedule[index + 1] - schedule[index])
+            middle = middle + velocity * step_size
+        return middle
+
+    def _sample_tail(self, previous, condition, encoded_mask, schedule,
+                     t_cond, device, progress):
+        previous_noisy = previous * (1 - t_cond) + torch.randn_like(previous) * t_cond
+        tail = torch.randn_like(condition[:, self.CONDITION_TILES:])
+        previous_t = torch.full(
+            previous.shape[:-4], t_cond, device=device, dtype=self.dtype
+        )
+        for index in range(schedule.shape[0] - 1):
+            progress()
+            latent = torch.cat((previous_noisy, tail), dim=1)
+            tail_t = torch.full(
+                tail.shape[:-4], float(schedule[index]), device=device, dtype=self.dtype
+            )
+            timestep = torch.cat((previous_t, tail_t), dim=1)
+            velocity = self._predict_step(
+                latent, timestep, condition, encoded_mask
+            )[:, self.CONDITION_TILES:]
+            step_size = float(schedule[index + 1] - schedule[index])
+            tail = tail + velocity * step_size
+        return tail
+
+    def _decode(self, latent, dense, dense_mask, height, width):
+        latent = rearrange(
+            latent, "1 nt nh nw c t h w -> 1 nt c t (nh h) (nw w)"
+        )
+        prediction = self.vae.decode(latent, dense, dense_mask)[..., :height, :width]
+        return rearrange(prediction, "1 c t h w -> t c h w").add(1).mul(0.5).clamp_(0, 1).float().cpu()
+
+    @staticmethod
+    def sampling_block_count(num_input_frames: int, temporal_factor: int) -> int:
+        total_length = num_input_frames * temporal_factor
+        t0 = 40
+        stride = 20
+        blocks = 1
+        while t0 + stride <= total_length - 1:
+            blocks += 2
+            t0 += stride * 2
+        if t0 < total_length:
+            blocks += 1
+        return blocks
+
+    @torch.no_grad()
+    def interpolate_sequence(
+        self,
+        frames: torch.Tensor,
+        temporal_factor: int,
+        num_steps: int = 16,
+        t_shift: float = 8.0,
+        t_cond: float = 0.1,
+        seed: int = 42,
+        progress_callback=None,
+    ) -> torch.Tensor:
+        if not 2 <= temporal_factor <= 16:
+            raise ValueError("LDF-VFI temporal factor must be between 2 and 16")
+        if num_steps < 1:
+            raise ValueError("LDF-VFI num_steps must be at least 1")
+        if t_shift <= 0:
+            raise ValueError("LDF-VFI t_shift must be greater than 0")
+        if not 0 <= t_cond <= 1:
+            raise ValueError("LDF-VFI t_cond must be between 0 and 1")
+        if frames.shape[0] < 2:
+            return frames
+
+        device = next(self.transformer.parameters()).device
+        if device.type != "cuda":
+            raise RuntimeError("Move LDF-VFI to a CUDA device before interpolation")
+        source = frames.detach().float().cpu()
+        height, width = source.shape[-2:]
+        schedule = self._time_schedule(num_steps, t_shift)
+        progress_callback = progress_callback or (lambda: None)
+
+        cuda_index = device.index
+        if cuda_index is None:
+            cuda_index = torch.cuda.current_device()
+        rng_context = torch.random.fork_rng(devices=[cuda_index])
+        with rng_context:
+            torch.manual_seed(int(seed))
+            with torch.cuda.device(device):
+                torch.cuda.manual_seed(int(seed))
+            chunks = self._interpolate_sequence_impl(
+                source, temporal_factor, schedule, t_cond, device, progress_callback
+            )
+
+        result = torch.cat(chunks, dim=0)
+        expected = (frames.shape[0] - 1) * temporal_factor + 1
+        return result[:expected]
+
+    def _interpolate_sequence_impl(self, source, factor, schedule, t_cond,
+                                   device, progress):
+        total_mask = torch.zeros(source.shape[0] * factor, dtype=torch.bool)
+        total_mask[::factor] = True
+        n_tiles = self.TRAIN_FRAMES // self.TILE_TIME
+        output_tiles = n_tiles - self.CONDITION_TILES
+        outputs = []
+
+        # First chunk.
+        mask = total_mask[:self.TRAIN_FRAMES]
+        input_end = int(mask.sum())
+        dense, dense_mask, condition, encoded_mask = self._prepare_condition(
+            source[:input_end], mask, device
+        )
+        latent = self._sample_free(condition, encoded_mask, schedule, device, progress)
+        latent = latent[:, :output_tiles]
+        previous = latent[:, -self.CONDITION_TILES:]
+        decode_time = output_tiles * self.TILE_TIME
+        outputs.append(self._decode(
+            latent, dense[:, :, :decode_time], dense_mask[:, :, :decode_time],
+            source.shape[-2], source.shape[-1],
+        ))
+
+        t0 = output_tiles * self.TILE_TIME
+        stride = (n_tiles - self.CONDITION_TILES * 2) * self.TILE_TIME
+        while t0 + stride <= total_mask.shape[0] - 1:
+            # A future/skip chunk establishes the far-side condition.
+            input_start = t0 + stride - self.CONDITION_TILES * self.TILE_TIME
+            input_end_t = t0 + stride * 2 + self.CONDITION_TILES * self.TILE_TIME
+            mask = total_mask[input_start:input_end_t]
+            source_start = int(total_mask[:input_start].sum())
+            source_end = int(total_mask[:input_end_t].sum())
+            dense, dense_mask, condition, encoded_mask = self._prepare_condition(
+                source[source_start:source_end], mask, device
+            )
+            skip = self._sample_free(condition, encoded_mask, schedule, device, progress)
+            skip = skip[:, self.CONDITION_TILES:-self.CONDITION_TILES]
+            following = skip[:, :self.CONDITION_TILES]
+            previous_next = skip[:, -self.CONDITION_TILES:]
+            start_time = self.CONDITION_TILES * self.TILE_TIME
+            end_time = output_tiles * self.TILE_TIME
+            decoded_skip = self._decode(
+                skip, dense[:, :, start_time:end_time], dense_mask[:, :, start_time:end_time],
+                source.shape[-2], source.shape[-1],
+            )
+
+            # Fill the gap between the preceding and skip chunks.
+            input_start = t0 - self.CONDITION_TILES * self.TILE_TIME
+            input_end_t = t0 + stride + self.CONDITION_TILES * self.TILE_TIME
+            mask = total_mask[input_start:input_end_t]
+            source_start = int(total_mask[:input_start].sum())
+            source_end = int(total_mask[:input_end_t].sum())
+            dense, dense_mask, condition, encoded_mask = self._prepare_condition(
+                source[source_start:source_end], mask, device
+            )
+            middle = self._sample_between(
+                previous, following, condition, encoded_mask,
+                schedule, t_cond, device, progress,
+            )
+            decoded_middle = self._decode(
+                middle, dense[:, :, start_time:end_time], dense_mask[:, :, start_time:end_time],
+                source.shape[-2], source.shape[-1],
+            )
+            outputs.extend((decoded_middle, decoded_skip))
+            previous = previous_next
+            t0 += stride * 2
+
+        # Remaining tail.
+        if t0 < total_mask.shape[0]:
+            input_start = t0 - self.CONDITION_TILES * self.TILE_TIME
+            mask = total_mask[input_start:]
+            source_start = int(total_mask[:input_start].sum())
+            dense, dense_mask, condition, encoded_mask = self._prepare_condition(
+                source[source_start:], mask, device
+            )
+            tail = self._sample_tail(
+                previous, condition, encoded_mask, schedule,
+                t_cond, device, progress,
+            )
+            start_time = self.CONDITION_TILES * self.TILE_TIME
+            outputs.append(self._decode(
+                tail, dense[:, :, start_time:], dense_mask[:, :, start_time:],
+                source.shape[-2], source.shape[-1],
+            ))
+
+        return outputs
@@ -1,7 +1,7 @@
 [project]
 name = "comfyui-tween"
-description = "Video frame interpolation nodes for ComfyUI using BIM-VFI, EMA-VFI, SGM-VFI, and GIMM-VFI. Designed for long videos with thousands of frames."
-version = "1.1.0"
+description = "Video frame interpolation nodes for ComfyUI using BIM-VFI, EMA-VFI, SGM-VFI, GIMM-VFI, SPEED, and LDF-VFI."
+version = "1.2.0"
 license = "Apache-2.0"
 requires-python = ">=3.10"
 dependencies = [
@@ -12,6 +12,9 @@ dependencies = [
    "easydict",
    "einops",
    "huggingface_hub",
+    "diffusers>=0.33.1,<0.40",
+    "accelerate>=1.5,<2",
+    "safetensors",
 ]

 [project.urls]
@@ -5,3 +5,6 @@ yacs
 easydict
 einops
 huggingface_hub
+diffusers>=0.33.1,<0.40
+accelerate>=1.5,<2
+safetensors
@@ -0,0 +1,158 @@
+"""ComfyUI inference adapter for the official SPEED runtime."""
+
+from __future__ import annotations
+
+from contextlib import nullcontext
+import importlib
+import logging
+from pathlib import Path
+import sys
+import types
+
+import torch
+
+
+logger = logging.getLogger("Tween")
+_SPEED_NAMESPACE = "_tween_speed_upstream"
+
+
+def _cuda_bf16_supported(device: torch.device) -> bool:
+    if device.type != "cuda":
+        return False
+    with torch.cuda.device(device):
+        return torch.cuda.is_bf16_supported()
+
+
+def _namespace_package(name: str, path: Path) -> None:
+    if name in sys.modules:
+        return
+    package = types.ModuleType(name)
+    package.__path__ = [str(path)]
+    package.__package__ = name
+    sys.modules[name] = package
+
+
+def load_speed_model_class(source_root: str):
+    """Import SpeedDiT without adding the upstream repository to sys.path."""
+    root = Path(source_root).resolve()
+    model_file = root / "src" / "models" / "model.py"
+    if not model_file.is_file():
+        raise RuntimeError(f"Invalid SPEED source directory: missing {model_file}")
+
+    _namespace_package(_SPEED_NAMESPACE, root)
+    _namespace_package(f"{_SPEED_NAMESPACE}.src", root / "src")
+    _namespace_package(f"{_SPEED_NAMESPACE}.src.models", root / "src" / "models")
+    module = importlib.import_module(f"{_SPEED_NAMESPACE}.src.models.model")
+    return module.SpeedDiT
+
+
+class SpeedVFIModel:
+    """Midpoint interpolation wrapper around the official SPEED SpeedDiT."""
+
+    def __init__(self, checkpoint_path: str, source_root: str,
+                 precision: str = "auto", device: str = "cpu"):
+        SpeedDiT = load_speed_model_class(source_root)
+        self.model = SpeedDiT(
+            hidden_dim=768,
+            head_dim=64,
+            depths=(2, 6, 4),
+            patch_sizes=(64, 32, 16),
+        )
+        self.precision = precision
+        self.device = str(device)
+        self._seed = 0
+        self._generator = None
+        self._generator_device = None
+        self._load_checkpoint(checkpoint_path)
+        self.model.requires_grad_(False).eval()
+        self.to(device)
+
+    def _load_checkpoint(self, checkpoint_path: str) -> None:
+        checkpoint = torch.load(checkpoint_path, map_location="cpu", weights_only=False)
+        state_dict = checkpoint.get("model", checkpoint) if isinstance(checkpoint, dict) else checkpoint
+        if not isinstance(state_dict, dict):
+            raise TypeError(f"SPEED checkpoint does not contain a state dict: {checkpoint_path}")
+        if state_dict and all(key.startswith("module.") for key in state_dict):
+            state_dict = {key[len("module."):]: value for key, value in state_dict.items()}
+        self.model.load_state_dict(state_dict, strict=True)
+
+    def _autocast_dtype(self, device: torch.device):
+        if self.precision == "fp32" or device.type != "cuda":
+            return None
+        if self.precision == "fp16":
+            return torch.float16
+        if self.precision == "bf16":
+            if not _cuda_bf16_supported(device):
+                raise RuntimeError(
+                    "SPEED BF16 precision requires a CUDA GPU with BF16 support; "
+                    "select auto or fp16 on this GPU"
+                )
+            return torch.bfloat16
+        if _cuda_bf16_supported(device):
+            return torch.bfloat16
+        return torch.float16
+
+    def to(self, device):
+        target = torch.device(device)
+        self.device = str(target)
+        # Match the official runtime: retain FP32 weights and use autocast for
+        # CUDA inference. This also avoids mixed-dtype timestep embedding bugs.
+        self.model.to(device=target, dtype=torch.float32)
+        # Keep the generator alive across CPU offloading. Recreating it on every
+        # pair batch would restart the noise stream whenever keep_device=False.
+        # _get_generator replaces it automatically if inference changes device.
+        return self
+
+    def set_seed(self, seed: int) -> None:
+        self._seed = int(seed)
+        self._generator = None
+        self._generator_device = None
+
+    def reset_seed(self) -> None:
+        self.set_seed(self._seed)
+
+    def clear_cache(self) -> None:
+        rope = getattr(self.model, "rope_embedder", None)
+        cache = getattr(rope, "rope_cache", None)
+        if cache is not None:
+            cache.clear()
+
+    def _get_generator(self, device: torch.device) -> torch.Generator:
+        device_name = str(device)
+        if self._generator is None or self._generator_device != device_name:
+            self._generator = torch.Generator(device=device)
+            self._generator.manual_seed(self._seed)
+            self._generator_device = device_name
+        return self._generator
+
+    @torch.no_grad()
+    def interpolate_batch(self, frames0, frames1, time_step=0.5):
+        if abs(float(time_step) - 0.5) > 1e-6:
+            raise ValueError("SPEED's released checkpoint supports midpoint interpolation only")
+
+        device = next(self.model.parameters()).device
+        frame0 = frames0.to(device=device, dtype=torch.float32, non_blocking=True).mul(2).sub(1)
+        frame1 = frames1.to(device=device, dtype=torch.float32, non_blocking=True).mul(2).sub(1)
+        cond_frames = torch.cat((frame0, frame1), dim=0)
+        noisy_frames = torch.randn(
+            frame0.shape,
+            generator=self._get_generator(device),
+            device=device,
+            dtype=torch.float32,
+        )
+        timestep = torch.full(
+            (frame0.shape[0],), 1000.0, device=device, dtype=torch.float32
+        )
+        autocast_dtype = self._autocast_dtype(device)
+        autocast = (
+            torch.autocast(device_type="cuda", dtype=autocast_dtype)
+            if autocast_dtype is not None
+            else nullcontext()
+        )
+        with autocast:
+            prediction = self.model(
+                noisy_frames=noisy_frames,
+                cond_frames=cond_frames,
+                timestep=timestep,
+            )
+        return prediction.div(2).add(0.5).clamp_(0, 1).float()