Video Models

Comfy Headless supports multiple video generation models, each with curated presets that configure resolution, frame count, and inference steps. This page covers what is available, what hardware each model needs, and how to choose the right one.

Supported models

Model	VRAM	Quality	Speed	Best for
LTX-Video 2	12 GB+	Excellent	Fast	General use, RTX 3080+
Hunyuan 1.5	14 GB+	Best	Slow	High quality, RTX 4080+
Wan 2.1/2.2	6-16 GB	Great	Medium	Budget GPUs, efficiency
Mochi	12 GB+	Excellent	Slow	Text adherence
AnimateDiff	6 GB+	Good	Fast	Quick previews
SVD	8 GB+	Good	Medium	Image-to-video
CogVideoX	10 GB+	Good	Slow	Legacy support

Models in bold are the recommended first choices for new users. The others remain fully supported for specific use cases.

Video presets

Every model has one or more presets — pre-configured combinations of resolution, frame count, and inference steps that produce good results out of the box.

LTX-Video 2

LTX-Video 2 is the best general-purpose choice. It balances quality and speed well and runs on any GPU with 12 GB or more of VRAM.

Preset	Resolution	Frames	Steps	Use case
ltx_quick	768x512	25	20	Fast previews and iteration
ltx_standard	1280x720	49	25	General-purpose generation
ltx_quality	1280x720	97	30	High frame count, best quality

Hunyuan 1.5

Hunyuan 1.5 produces the highest visual quality but requires more VRAM and runs slower. Best suited for final renders rather than iteration.

Preset	Resolution	Frames	Notes
hunyuan15_720p	1280x720	121	High quality at 720p
hunyuan15_1080p	1920x1080	—	Uses super-resolution upscaling

Wan

The Wan family offers excellent efficiency. The 1.3B parameter variant runs on GPUs with as little as 6 GB of VRAM, making it the most accessible option for budget hardware.

Preset	Resolution	Frames	VRAM	Notes
wan_1.3b	720x480	49	6 GB	Lightweight, budget GPUs
wan_14b	1280x720	81	12 GB	Full quality Wan model

Choosing a model

Use the built-in recommendation function to let the library choose based on your GPU:

from comfy_headless import get_recommended_preset

preset = get_recommended_preset(vram_gb=16)  # Returns "hunyuan15_720p"
preset = get_recommended_preset(vram_gb=8)   # Returns "wan_1.3b"
preset = get_recommended_preset(vram_gb=12)  # Returns "ltx_standard"

Decision guide

• 6 GB VRAM — Use Wan 1.3B (wan_1.3b). It is the only model that fits comfortably.
• 8-10 GB VRAM — AnimateDiff for quick previews, Wan 1.3B for better quality.
• 12 GB VRAM — LTX-Video 2 is the sweet spot. Fast and high quality.
• 14-16 GB VRAM — Hunyuan 1.5 at 720p for the best visual quality, or LTX-Video 2 for faster iteration.
• 16 GB+ VRAM — Hunyuan 1.5 at 1080p with super-resolution for production renders.

Working with presets in code

from comfy_headless import VIDEO_PRESETS, ComfyClient

# Inspect all presets
for name, preset in VIDEO_PRESETS.items():
    print(f"{name}: {preset}")

# Generate with a specific preset
client = ComfyClient()
result = client.generate_video(
    prompt="a timelapse of clouds over a valley",
    preset="ltx_quality"
)

The VIDEO_PRESETS dictionary maps preset names to their full configuration. You can inspect it to see exactly what settings each preset uses.

ComfyUI node requirements

Each video model requires specific custom nodes installed in your ComfyUI instance. See the Architecture page for the full list of required nodes per model.

Next steps

• Configure WebSocket progress tracking for long video generations in Configuration.
• See the VideoSettings and VideoModel classes in API Reference.