model releaseApple

Apple releases FastFrames-L, a 14B two-step video model distilled from Wan2.1-14B for interactive 480p generation

TL;DR

Apple has released FastFrames-L (14B) and FastFrames-S (1.3B), autoregressive two-step video diffusion models that generate 480p video at interactive latency. FastFrames-L is distilled from Wan2.1-14B, and the repository lists 35.4 ms total latency for FastFrames-S on one GPU.

3 min read
0

Apple has released FastFrames-L, a 14B-parameter autoregressive, two-step video diffusion model that generates 480p video at interactive latency. The release, published on Hugging Face with an official implementation, also includes a smaller sibling, FastFrames-S (1.3B). Both models accompany the paper "FastFrames: Video Generation at Interactive Latency" by Gene Chou, Guandao Yang, Fujun Luan, Lars Mescheder, Wei Dong, Stephan R. Richter and Vladlen Koltun.

Model details

  • FastFrames-L: 14B parameters, autoregressive, two-step video diffusion.
  • FastFrames-S: 1.3B parameters, same architecture family.
  • Output: 832 × 480 video, 81 frames at 16 fps.
  • Base model: Wan-AI/Wan2.1-T2V-14B, per the Hugging Face model tree.
  • Pricing and context window: not applicable; no hosted pricing is listed, and the model is not deployed by any Inference Provider.

Training approach

FastFrames-L starts by distilling Wan2.1-14B into a two-step causal model, following the Causal Forcing++ recipe. Apple then trains a pixel decoder and a KV refinement network on top, the same procedure used for FastFrames-S. For FastFrames-S, the two-step frame-wise Causal Forcing++ generator stays frozen while the surrounding networks are trained.

Apple says it is releasing all checkpoints:

  1. A causal teacher trained on latents generated by Wan2.1-14B.
  2. The consistency-distilled model.
  3. The DMD-distilled model.
  4. The final generator, pixel decoder and KV refinement network.

The authors thank the creators of Wan2.1, Self-Forcing and Causal Forcing++, on which the FastFrames code and models build.

Latency

The repository's example timings cover FastFrames-S at 480p on 1 GPU with a full KV cache:

Component Calls Latency per call Total latency
DiT evaluation, t=1000 1 16.1 ms 16.1 ms
DiT evaluation, t=833 1 16.2 ms 16.2 ms
KV refinement 1 2.2 ms 2.2 ms
Pixel decoder 1 0.9 ms 0.9 ms
Total 35.4 ms (paper: 36 ms)

A benchmark script for FastFrames-L accepts a --gpus 8 flag, and the demo can run on 1 or 8 GPUs. Latency figures for FastFrames-L are not listed in the example table. VBench Quality, Semantic and Total scores (Table 1 of the paper) can be reproduced with the repository's evaluation guide; the scores themselves are not listed on the model page.

LingBot optimization

Apple also releases an optimized version of LingBot 1.3B that it says is 16x faster than the official model and allows fluid camera-conditioned control. An interactive demo with LingBot text-to-video generation runs locally at http://localhost:8000.

Requirements and license

The code requires Linux x86-64, an NVIDIA GPU and a CUDA 12.8-compatible driver. LingBot-World v2 code and weights follow their upstream CC BY-NC-SA 4.0 license. All other FastFrames code is licensed under the Apple Sample Code License.

What this means

FastFrames-L applies a two-step causal distillation recipe to a 14B-parameter base, Wan2.1-14B, rather than only to a small model. The 35.4 ms figure for FastFrames-S on one GPU shows the latency target the project is built around, and the 1.3B and 14B pair gives users two sizes to choose from. Because Apple publishes the teacher, the consistency-distilled and DMD-distilled checkpoints along with the final generator, researchers can inspect each distillation stage. Anyone planning to use the LingBot components should note their CC BY-NC-SA 4.0 license.

Related Articles

model release

Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license

Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.

model release

Bria releases RMBG-3.0, a 4-step background remover built on Fibo 2.0, with open weights for non-commercial use

Bria AI has released RMBG-3.0, the third generation of its background removal models and the first built on a generative foundation. It is Fibo 2.0, fine-tuned and distilled to produce an alpha matte in 4 steps. Bria claims it ranks first among six systems in its own benchmark of 234 hard images, with an average rank of 1.51.

model release

Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context

Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.

model release

SenseNova-Loop-9B released: 9B parameters, 56 layer executions, 512K context

SenseNova has released SenseNova-Loop-9B, a roughly 9B-parameter multimodal model that runs 56 decoder-layer executions per forward pass from 32 distinct layers. It supports up to 512K tokens of context and, per the model card, needs about 15 GiB of GPU memory at 256K context with 4-bit weights and an 8-bit KV cache.

Comments

Loading...