Picture by Creator
# Lights, Digital camera…
With the launch of Veo and Sora, video era has reached a brand new excessive. Creators are experimenting extensively, and groups are integrating these instruments into their advertising workflows. Nonetheless, there’s a downside: most closed techniques accumulate your information and apply seen or invisible watermarks that label outputs as AI-generated. In case you worth privateness, management, and on-device workflows, open supply fashions are your best choice, and a number of other now rival the outcomes of Veo.
On this article, we are going to evaluate the highest 5 video era fashions, offering technical information and a demo video that will help you assess their video era capabilities. Each mannequin is out there on Hugging Face and may run domestically through ComfyUI or your most popular desktop AI purposes.
# 1. Wan 2.2 A14B
Wan 2.2 upgrades its diffusion spine with a Combination-of-Specialists (MoE) structure that splits denoising throughout timesteps into specialised specialists, rising efficient capability with no compute penalty. The staff additionally curated aesthetic labels (e.g. lighting, composition, distinction, shade tone) to make “cinematic” appears extra controllable. In comparison with Wan 2.1, coaching scaled considerably (+65.6% photographs, +83.2% movies), enhancing movement, semantics, and aesthetics.
Wan 2.2 studies top-tier efficiency amongst each open and closed techniques. You’ll be able to discover the text-to-video and image-to-video A14B repositories on Hugging Face: Wan-AI/Wan2.2-T2V-A14B and Wan-AI/Wan2.2-I2V-A14B
# 2. Hunyuan Video
HunyuanVideo is a 13B-parameter open video basis mannequin educated in a spatial–temporal latent area through a causal 3D variational autoencoder (VAE). Its transformer makes use of a “dual-stream to single-stream” design: textual content and video tokens are first processed independently with full consideration after which fused, whereas a decoder-only multimodal LLM serves because the textual content encoder to enhance instruction following and element seize.
The open supply ecosystem consists of code, weights, single- and multi-GPU inference (xDiT), FP8 weights, Diffusers and ComfyUI integrations, a Gradio demo, and the Penguin Video Benchmark.
# 3. Mochi 1
Mochi 1 is a 10B Uneven Diffusion Transformer (AsymmDiT) educated from scratch, launched underneath Apache 2.0. It {couples} with an Uneven VAE that compresses movies 8×8 spatially and 6x temporally right into a 12-channel latent, prioritizing visible capability over textual content whereas utilizing a single T5-XXL encoder.
In preliminary evaluations, the Genmo staff positions Mochi 1 as a state-of-the-art open mannequin with high-fidelity movement and robust immediate adherence, aiming to shut the hole with closed techniques.
# 4. LTX Video
LTX-Video is a DiT-based (Diffusion Transformer) image-to-video generator constructed for pace: it produces 30 fps movies at 1216×704 quicker than actual time, educated on a big, numerous dataset to stability movement and visible high quality.
The lineup spans a number of variants: 13B dev, 13B distilled, 2B distilled, and FP8 quantized builds, plus spatial and temporal upscalers and ready-to-use ComfyUI workflows. In case you are optimizing for quick iterations and crisp movement from a single picture or quick conditioning sequence, LTX is a compelling selection.
# 5. CogVideoX-5B
CogVideoX-5B is the higher-fidelity sibling to the 2B baseline, educated in bfloat16 and advisable to run in bfloat16. It generates 6-second clips at 8 fps with a hard and fast 720×480 decision and helps English prompts as much as 226 tokens.
The mannequin’s documentation exhibits anticipated Video Random Entry Reminiscence (VRAM) for single- and multi-GPU inference, typical runtimes (e.g. round 90 seconds for 50 steps on a single H100), and the way Diffusers optimizations like CPU offload and VAE tiling/slicing have an effect on reminiscence and pace.
# Selecting a Video Technology Mannequin
Listed below are some high-level takeaways for serving to select the correct video era mannequin on your wants.
- If you need cinema-friendly appears and 720p/24 on a single 4090: Wan 2.2 (A14B for core duties; the 5B hybrid TI2V for environment friendly 720p/24)
- In case you want a big, general-purpose T2V/I2V basis with robust movement and a full open supply software program (OSS) toolchain: HunyuanVideo (13B, xDiT parallelism, FP8 weights, Diffusers/ComfyUI)
- If you need a permissive, hackable state-of-the-art (SOTA) preview with trendy movement and a transparent analysis roadmap: Mochi 1 (10B AsymmDiT + AsymmVAE, Apache 2.0)
- In case you care about real-time I2V and editability with upscalers and ComfyUI workflows: LTX-Video (30 fps at 1216×704, a number of 13B/2B and FP8 variants)
- In case you want environment friendly 6s 720×480 T2V, stable Diffusers help, and quantization all the way down to small VRAM: CogVideoX-5B
Abid Ali Awan (@1abidaliawan) is an authorized information scientist skilled who loves constructing machine studying fashions. At the moment, he’s specializing in content material creation and writing technical blogs on machine studying and information science applied sciences. Abid holds a Grasp’s diploma in know-how administration and a bachelor’s diploma in telecommunication engineering. His imaginative and prescient is to construct an AI product utilizing a graph neural community for college students battling psychological sickness.
Picture by Creator
# Lights, Digital camera…
With the launch of Veo and Sora, video era has reached a brand new excessive. Creators are experimenting extensively, and groups are integrating these instruments into their advertising workflows. Nonetheless, there’s a downside: most closed techniques accumulate your information and apply seen or invisible watermarks that label outputs as AI-generated. In case you worth privateness, management, and on-device workflows, open supply fashions are your best choice, and a number of other now rival the outcomes of Veo.
On this article, we are going to evaluate the highest 5 video era fashions, offering technical information and a demo video that will help you assess their video era capabilities. Each mannequin is out there on Hugging Face and may run domestically through ComfyUI or your most popular desktop AI purposes.
# 1. Wan 2.2 A14B
Wan 2.2 upgrades its diffusion spine with a Combination-of-Specialists (MoE) structure that splits denoising throughout timesteps into specialised specialists, rising efficient capability with no compute penalty. The staff additionally curated aesthetic labels (e.g. lighting, composition, distinction, shade tone) to make “cinematic” appears extra controllable. In comparison with Wan 2.1, coaching scaled considerably (+65.6% photographs, +83.2% movies), enhancing movement, semantics, and aesthetics.
Wan 2.2 studies top-tier efficiency amongst each open and closed techniques. You’ll be able to discover the text-to-video and image-to-video A14B repositories on Hugging Face: Wan-AI/Wan2.2-T2V-A14B and Wan-AI/Wan2.2-I2V-A14B
# 2. Hunyuan Video
HunyuanVideo is a 13B-parameter open video basis mannequin educated in a spatial–temporal latent area through a causal 3D variational autoencoder (VAE). Its transformer makes use of a “dual-stream to single-stream” design: textual content and video tokens are first processed independently with full consideration after which fused, whereas a decoder-only multimodal LLM serves because the textual content encoder to enhance instruction following and element seize.
The open supply ecosystem consists of code, weights, single- and multi-GPU inference (xDiT), FP8 weights, Diffusers and ComfyUI integrations, a Gradio demo, and the Penguin Video Benchmark.
# 3. Mochi 1
Mochi 1 is a 10B Uneven Diffusion Transformer (AsymmDiT) educated from scratch, launched underneath Apache 2.0. It {couples} with an Uneven VAE that compresses movies 8×8 spatially and 6x temporally right into a 12-channel latent, prioritizing visible capability over textual content whereas utilizing a single T5-XXL encoder.
In preliminary evaluations, the Genmo staff positions Mochi 1 as a state-of-the-art open mannequin with high-fidelity movement and robust immediate adherence, aiming to shut the hole with closed techniques.
# 4. LTX Video
LTX-Video is a DiT-based (Diffusion Transformer) image-to-video generator constructed for pace: it produces 30 fps movies at 1216×704 quicker than actual time, educated on a big, numerous dataset to stability movement and visible high quality.
The lineup spans a number of variants: 13B dev, 13B distilled, 2B distilled, and FP8 quantized builds, plus spatial and temporal upscalers and ready-to-use ComfyUI workflows. In case you are optimizing for quick iterations and crisp movement from a single picture or quick conditioning sequence, LTX is a compelling selection.
# 5. CogVideoX-5B
CogVideoX-5B is the higher-fidelity sibling to the 2B baseline, educated in bfloat16 and advisable to run in bfloat16. It generates 6-second clips at 8 fps with a hard and fast 720×480 decision and helps English prompts as much as 226 tokens.
The mannequin’s documentation exhibits anticipated Video Random Entry Reminiscence (VRAM) for single- and multi-GPU inference, typical runtimes (e.g. round 90 seconds for 50 steps on a single H100), and the way Diffusers optimizations like CPU offload and VAE tiling/slicing have an effect on reminiscence and pace.
# Selecting a Video Technology Mannequin
Listed below are some high-level takeaways for serving to select the correct video era mannequin on your wants.
- If you need cinema-friendly appears and 720p/24 on a single 4090: Wan 2.2 (A14B for core duties; the 5B hybrid TI2V for environment friendly 720p/24)
- In case you want a big, general-purpose T2V/I2V basis with robust movement and a full open supply software program (OSS) toolchain: HunyuanVideo (13B, xDiT parallelism, FP8 weights, Diffusers/ComfyUI)
- If you need a permissive, hackable state-of-the-art (SOTA) preview with trendy movement and a transparent analysis roadmap: Mochi 1 (10B AsymmDiT + AsymmVAE, Apache 2.0)
- In case you care about real-time I2V and editability with upscalers and ComfyUI workflows: LTX-Video (30 fps at 1216×704, a number of 13B/2B and FP8 variants)
- In case you want environment friendly 6s 720×480 T2V, stable Diffusers help, and quantization all the way down to small VRAM: CogVideoX-5B
Abid Ali Awan (@1abidaliawan) is an authorized information scientist skilled who loves constructing machine studying fashions. At the moment, he’s specializing in content material creation and writing technical blogs on machine studying and information science applied sciences. Abid holds a Grasp’s diploma in know-how administration and a bachelor’s diploma in telecommunication engineering. His imaginative and prescient is to construct an AI product utilizing a graph neural community for college students battling psychological sickness.
