Compatibility Matrix¶

This page summarizes common model + optimization combinations.

For the canonical, code-level list of model IDs recognized by VideoGenerator.from_pretrained(...), see the registrations in fastvideo/registry.py (register_configs(...) entries).

Registered Model IDs¶

Every Hugging Face model ID registered in fastvideo/registry.py on main (commit 8d89f30d), grouped by family. Any ID below can be passed to VideoGenerator.from_pretrained(...); FastVideo resolves the matching pipeline and sampling defaults. The Family column is a documentation grouping: it follows each registration's declared model_family, except black-forest-labs/FLUX.1-dev, which declares none and is listed under flux for readability. The Workloads column shows each registration's declared workload_types; — means the entry is registered without a UI workload option but is still loadable by ID. The Example column links a runnable script in examples/inference/basic/ where one exists.

Note (stable_audio): the Stable Audio Open pipelines generate audio (StableAudioT2AConfig / StableAudioOpenSmallConfig); they are registered under the generic T2V workload option in the registry.

Note (MMAudio): the registered Hugging Face model ID is reserved but not yet public. Follow the MMAudio inference guide to convert the official weights locally and set MMAUDIO_MODEL_PATH.

Note (Cosmos Predict2.5 distilled): the released 2B Text2World student is supported through local checkpoint conversion, but no public converted model ID is registered yet. Follow the validation guide in tests/local_tests/cosmos25/README.md, then pass the converted directory to basic_cosmos2_5_distilled_t2w.py --model.

The public DFD Video2World student is also supported through local DCP conversion. It uses exactly one conditioning image, 4 steps, 81 frames at 704x1280 and 24 FPS. Use basic_cosmos2_5_dfd_i2w.py; a public converted model ID is not registered yet.

Note (MiniMax H3): T2VA, FL2VA, and Ref2VA all generate video with stereo audio. Use the Ref2VA example when passing ordered image, video, or audio references. Distilled Ref2VA PDD students (eight transformer forwards, with reference videos as sparse VSA regions) run through basic_fasth3_omniref_pdd.py; see FastH3 distilled checkpoint schedules.

Note (Kandinsky 6): the kandinsky6 IDs (Pro 30.1B and Lite 3.2B, each base and distilled) generate video with audio from text, optionally plus an image (see the T2IVA guide); the kandinsky6_sr IDs upscale an existing video (see the Video SR guide).

Note (LTX-2.5): T2V and I2V produce synchronized video and audio. The official Lightricks/LTX-2.5 repository uses gated, split component files, so it must first be converted to FastVideo's component layout. See the LTX-2.5 inference guide.

Note (Wan-VACE): not currently supported — no VACE pipeline or registered model ID exists on main (#1435). The closest supported path is the Wan2.1-Fun control pipeline (IRMChen/Wan2.1-Fun-1.3B-Control-Diffusers).

The symbols used have the following meanings:

• ✅ = Full compatibility
• ❌ = No compatibility
• ⭕ = Does not apply to this model

Models x Optimization¶

The HuggingFace Model ID can be passed directly to from_pretrained(). FastVideo then uses model-specific default settings for pipeline initialization and sampling.

Registered models absent from this table have not been validated against these optimizations: absence means untested, not incompatible.

Apple Silicon native runtime¶

Apple Silicon uses the native MLX runtime. FastMetal-QAD is the packaged Wan release. FastH3 V1 and FastH3 V2 use a source checkout and local DiT conversion. CUDA FastWan-QAD (FastVideo/FastWan-QAD-1.3B, FastVideo/FastWan-QAD-FP8-1.3B) is the NVIDIA release. See the MLX install guide and the FastMetal-QAD blog.

Note: Wan2.2 TI2V 5B has some quality issues when performing I2V generation. We are working on fixing this issue.

***Lucy Edit Dev uses a non-commercial model license. FastVideo support is focused on inference integration for video editing workflows.

****Wan2.2 S2V is audio-driven: it takes a reference image plus a speech/audio track and animates the subject in sync with it. The official checkpoint ships in native Wan format; scripts/checkpoint_conversion/wan_s2v_to_diffusers.py repackages it into the Diffusers layout this repo id refers to (bundling the wav2vec2-large-xlsr-53-english audio encoder). num_frames follows the usual 4k+1 contract (default 81): one clip generates 84 frames and shows 81, and longer requests chain clips that carry the previous 73 frames forward as motion context, as the official runner does. The source audio is muxed into the saved MP4. Sequence parallelism is not yet wired up for this pipeline.

Sliding Tile Attn (Legacy Branch) entries refer to the archived sta_do_not_delete branch workflow, not active main inference wiring.

Canonical Supported IDs¶

The authoritative source for model-ID recognition is fastvideo/registry.py. If a model ID is registered there, FastVideo can resolve default pipeline and sampling configuration for it.

Note (GEN3C): The official nvidia/GEN3C-Cosmos-7B repo provides a raw model.pt checkpoint. Use a Diffusers-format repo (for example, FastVideo/GEN3C-Cosmos-7B-Diffusers) or convert locally with scripts/checkpoint_conversion/convert_gen3c_to_fastvideo.py.

Hardware and OS¶

Per the installation guides:

• NVIDIA GPU (x86_64) — CUDA 12.6 or 13.0; see the GPU install guide.
• NVIDIA DGX Spark (GB10, aarch64) — CUDA 13, from-source kernel build; see the DGX Spark install guide. Two Sparks over QSFP use Ray sequence parallel; see Pair two NVIDIA DGX Sparks.
• Apple silicon — macOS 14 or newer; MLX runtime for FastMetal-QAD, FastH3 V1, and FastH3 V2. See the MLX install guide. The older basic_mps.py demo is PyTorch MPS only.

Optimization-specific hardware constraints (e.g. STA requiring Hopper) are listed under Special requirements.

Special requirements¶

Sliding Tile Attention¶

• Full STA pipeline usage is on the archived branch: https://github.com/hao-ai-lab/FastVideo/tree/sta_do_not_delete
• STA currently requires Hopper GPUs (H100s).

TurboWan2.1 (TurboDiffusion)¶

• Uses TurboDiffusionPipeline with RCM scheduler for 1-4 step generation
• Requires SLA attention backend: export FASTVIDEO_ATTENTION_BACKEND=SLA_ATTN
• Uses guidance_scale=1.0 (no classifier-free guidance)

Matrix Game 2.0¶

• Image-to-video game world models with keyboard/mouse control input
• Three variants available: Base (universal), GTA, and TempleRun
• Each variant has different keyboard dimensions for control inputs