Diffusers 0.40.0 ships open video-with-sound, animation, and music
The image library now runs MiniMax-H3 video+audio, LTX-2.5, Wan-Animate-2 and Stable Audio 3 — all open, with Modular Diffusers now stable.
Copy markdown
MiniMax-H3: text-to-video with a native soundtrack
A single 33B transformer denoises video and stereo audio together — 24fps, 5–15s clips, no separate vocoder. With int8 quantization and block offloading it runs on a 24–32GB GPU (it does want ~75GB of host RAM); an 80GB Hopper card with FlashAttention-3 is ~3x faster.
LTX-2.5 lands with built-in prompt enhancement
Lightricks' video model slots into the existing LTX2 pipeline with a Gemma 4 text encoder, adding duration prediction and prompt enhancement so a short prompt expands into a usable clip.
Wan-Animate-2 retargets motion onto your character
Feed it a driving video and Wan-Animate-2 animates your reference character in fixed-length segments, with base and distilled presets to trade quality for speed.
Stable Audio 3 and MiniMax Music 3 round out the audio stack
Stable Audio 3 generates 44.1kHz stereo audio from text across three pipeline variants, and MiniMax Music 3 writes full songs up to five minutes from lyrics — an open soundtrack layer to sit next to the video models.
Modular Diffusers is stable; tensor-parallel inference arrives
Modular Diffusers drops its experimental flag (the new models ship as modular blocks), and TensorParallelConfig brings tensor-parallel inference to CUDA and AWS Neuron, so you can split these large models across GPUs. It's one `pip install -U diffusers` to pull all of it.