CogVideoX

CogVideoX is an open-source video generation model series developed by Zhipu AI and researchers at Tsinghua University's THUDM lab. First released in August 2024 and accepted as a conference paper at ICLR 2025, it generates video from text prompts or source images without any subscription or per-clip fees. The model weights are freely downloadable from HuggingFace and the codebase is MIT/Apache 2.0 licensed on GitHub, where the project has accumulated over 12,700 stars. As of April 2026 the repository remains actively maintained, with the CogKit fine-tuning framework added in March 2025 and community wrappers serving hundreds of ComfyUI workflows.

The model family spans five variants covering different hardware tiers and use cases: the lightweight 2B-parameter version runs on GPUs as old as a GTX 1080 Ti; the 5B flagship handles detailed text-to-video generation with strong prompt adherence; and the 5B-I2V variant is consistently ranked the best image-to-video model accessible to consumer GPU owners. CogVideoX1.5 (November 2024) pushed resolution to 1360x768 at 16fps and extended clip length to 10 seconds. The architecture uses a 3D Variational Autoencoder for spatial-temporal compression paired with an Expert Transformer, with no traditional cross-attention module. Full diffusers integration means one-line pipeline calls, and LoRA fine-tuning is supported out of the box via CogKit.

What CogVideoX outputs in April 2026

The current model lineup as of April 2026 covers two generations. The original CogVideoX-5B produces 720x480 video at 8 fps for 6-second clips. It runs in BF16 precision with inference around 180 seconds on an A100 or 90 seconds on an H100. With CPU offloading and INT8 quantization the VRAM floor drops to approximately 4.4GB, making it one of the most accessible local video models available.

CogVideoX1.5-5B, released November 8, 2024, is a meaningful step up: 1360x768 resolution, 16 fps, and 5-10 second durations. The image-to-video companion (CogVideoX1.5-5B-I2V) accepts any aspect ratio and outputs up to 81 frames. Both are accessible through the Hugging Face diffusers library with the same pipeline interface as the base models. On the Qingying platform, Zhipu AI also demonstrated the models paired with their CogSound audio model to produce synchronized audio and 4K ultra-high-definition output in batch, though the open-weight releases remain at the specifications above.

ComfyUI support is first-class. The kijai/ComfyUI-CogVideoXWrapper has 1,500+ stars and 341 commits, supporting text-to-video, image-to-video, LoRA, GGUF quantization, ControlNet pose control, temporal tiling for extended clips, and experimental motion guidance via Go-with-the-Flow. Stable Diffusion WebUI Forge integration also exists.

"The 5B model absolutely blows it out of the water. Not only is it more detailed in terms of the scenery, it also did a much better job in terms of prompt adherence.". Bijan Bowen, Ominous Industries, August 2024

Where CogVideoX sits versus HunyuanVideo and Mochi 1

Three models dominate the conversation when self-hosters compare open-source video generation: HunyuanVideo from Tencent, Mochi 1 from Genmo, and CogVideoX. They are built on different architectural bets and serve meaningfully different hardware profiles.

HunyuanVideo (Tencent) uses a 13-billion-parameter MoE (Mixture-of-Experts) architecture with cross-frame text guidance modules and mixed-resolution training, where temporal dynamics are learned at lower resolution while spatial details are trained at higher resolution. HunyuanVideo 1.5 (November 2025) trimmed this to 8.3B parameters and dropped the VRAM floor to 14GB minimum. In benchmarks it scores 9.1/10 for temporal consistency against CogVideoX's 7.9/10, and it is clearly the superior choice for anything involving photorealistic human faces or multi-character scenes. The tradeoff is hardware: at full quality on the original model, you are looking at 60-80GB VRAM at 720p, which requires datacenter GPU access. On-demand API pricing runs approximately $11/minute of output. For most self-hosters CogVideoX is the accessible alternative when HunyuanVideo's quality is desired but the hardware is not available.

Mochi 1 (Genmo) uses an AsymmDiT (Asymmetric Diffusion Transformer) architecture with 10 billion parameters, specifically tuned for fluid motion at 30 fps. That frame rate advantage is real: Mochi 1's output looks noticeably smoother than base CogVideoX at 8 fps. However, Mochi 1 is capped at 480p resolution with no HD upgrade released, requires 16GB VRAM minimum, and Genmo did not ship a meaningful successor that could keep pace with either HunyuanVideo 1.5 or CogVideoX1.5. The community largely views Mochi 1 as historically important for demonstrating that fluid motion was achievable at open-source, but practically superseded for most workflows. CogVideoX1.5 at 16fps and 1360x768 produces higher-resolution output than Mochi 1, though the motion smoothness still favors Mochi on fast movement.

CogVideoX at 5 billion parameters has the lowest hardware floor of the three. It runs from approximately 5GB VRAM with quantization, generates acceptable video on RTX 3060-class cards, and offers the most accessible path to local image-to-video animation. The tradeoff is photorealism: CogVideoX produces more stylized, illustrative output than HunyuanVideo, scoring 7.4/10 versus HunyuanVideo's 8.7/10 on photorealistic human subjects in third-party benchmarks.

"CogVideoX has the best image-to-video mode among open models, allowing you to generate a hero image with Flux or SDXL, then animate it with CogVideoX, with 3D Causal VAE technology delivering strong detail preservation.". InsiderLLM, Local AI Video Generation Guide, 2025

The real cost of generating video with CogVideoX

The model weights cost nothing. Hugging Face hosting is free. The actual cost is electricity, hardware depreciation, and time.

On an RTX 4090, expect 5-7 clips per hour for the 5B model at standard settings. On an RTX 3060 (12GB), generation time extends to 9-15 minutes per clip. An H100 cuts the 5B inference to approximately 90 seconds. If you do not own suitable hardware, cloud GPU providers (Vast.ai, RunPod, Modal) charge roughly $0.50-$3.00 per hour of H100/A100 time, putting a single clip at cents to low dollars depending on settings and the provider.

Third-party API wrappers such as Replicate and SiliconFlow host quantized CogVideoX endpoints with per-run pricing that varies by provider, typically in the $0.02-$0.10 range per clip. These are independent businesses, not Zhipu AI products. The official Qingying platform from Zhipu AI offers cloud access with its own pricing, but the open-weight models themselves remain free.

For fine-tuning with CogKit, the compute budget scales with dataset size and target resolution. LoRA fine-tuning on CogVideoX-5B typically completes in hours on an 80GB A100 for small datasets. The January 2025 update reduced memory requirements for LoRA training, making fine-tuning more tractable on consumer 24GB cards.

Where CogVideoX consistently breaks

Slow generation is the most common complaint. GitHub issue #545 documented CogVideoX1.5-5B running at 0/50 inference steps for over an hour on some configurations. More typically, users on consumer GPUs report GPU utilization around 60% during inference, indicating pipeline bottlenecks in CPU-to-GPU data movement. Community solutions include SageAttention (reduces VRAM and speeds compute), torch.compile, and the xDiT parallel framework (claimed 7.75x speedup across multiple GPUs). None of these are plug-and-play.

Motion instability on complex scenes is a documented limitation. Research papers published in 2025 specifically cite CogVideoX as showing "identity drift and physically implausible dynamics" during high-dynamic motion such as fast dancing or competitive sports. The issue is more pronounced on the base 5B model than on 1.5, and it worsens as clip length increases beyond 6 seconds.

The 8fps output on base CogVideoX-2B and 5B looks noticeably choppy when played back. CogVideoX1.5 improved this to 16fps, which is better but still below Mochi 1's 30fps or commercial tools that target 24fps.

English-only prompts with a hard 226-token limit mean non-English users must translate first, and long, complex prompt descriptions get truncated. There is no native multilingual capability in the current weights.

Breaking changes in tooling are an ongoing friction point. The kijai ComfyUI wrapper's Update 8 refactor broke existing workflows, requiring users to rebuild their node graphs. This is expected for a fast-moving open-source project but inconvenient for anyone who built production pipelines on older configurations.

Best use cases versus skip-this scenarios

CogVideoX works best when:

• You want to animate a static image. The I2V pipeline, particularly with FLUX or SDXL as the source image generator, is the most praised real-world workflow for CogVideoX. Detail preservation is strong, and the 5B-I2V model consistently outranks alternatives in its VRAM tier for this specific task.

• You need locally-run, zero-subscription video generation on consumer hardware. From GTX 1080 Ti (2B) to RTX 3060 (5B), CogVideoX has the lowest hardware floor of any comparable quality-tier model.

• Prompt adherence matters for structured scenes. Technical visualizations, architectural walkthroughs, product demonstrations, and scientific content all benefit from CogVideoX's text-following precision, which independent testers have rated higher than other open models for instruction-following on detailed descriptions.

• You want to fine-tune on custom style data. CogKit and diffusers LoRA support make the model a practical fine-tuning base for studio-specific looks.

Skip CogVideoX when:

• Your content centers on photorealistic human faces, bodies, or multi-character scenes. HunyuanVideo scores two full points higher on photorealistic human subjects in third-party benchmarks, and the quality difference is visible to non-technical reviewers.

• Speed matters. If you need a clip in under 3 minutes on consumer hardware, LTX-Video is the community recommendation. CogVideoX at 9-15 minutes per clip on an RTX 3060 will frustrate any workflow requiring iteration.

• You work in non-English languages. The 226-token English-only constraint is not a setting that can be adjusted; it is baked into the model training.

• You do not have access to a modern Nvidia GPU. CPU-only inference is technically possible but impractically slow for any production use.