Stable Video Diffusion (SVD) is Stability AI's video extension of the Stable Diffusion U-Net, released in November 2023. It adds temporal convolution and attention layers to the SD 2.1 architecture, enabling image-to-video generation from a single still frame.
This guide covers how to run SVD in ComfyUI, what hardware it requires, and where it fits now that newer DiT-based models have surpassed it on motion quality. AnimateDiff is also covered as an alternative for text-to-video generation.
Takeaways
- SVD is not Stable Diffusion but rather a separate Stability AI model for image-to-video generation, built on SD 2.1 with temporal layers added.
- ComfyUI is the recommended interface for SVD.
- SVD-XT needs a minimum of 8GB VRAM, but 16GB+ is recommended to avoid pipeline OOM errors.
- Newer models have moved ahead: Wan 2.2 supports T2V and I2V from 8GB VRAM; LTX-2.3 adds native audio and 4K output.
- Thunder Compute's one-click ComfyUI template runs SVD on an RTX A6000 48GB at $0.35/hr, with per-minute billing and no local setup required.
Can Stable Diffusion Generate Video?
No, Stable Diffusion only generates images. Video generation in the Stable Diffusion ecosystem uses different approaches, each with different tooling and output characteristics.
Stable Diffusion (Image) vs Stable Video Diffusion (Video)
Stable Video Diffusion (SVD) is a separate model from Stable Diffusion, though it shares the SD 2.1 image backbone. To create it Stability AI extended the SD 2.1 U-Net with temporal convolution and attention layers, trained it on a large video dataset, and fine-tuned it on a curated high-quality subset. SVD "understands" how pixels move over time, not just how they look in a single frame.
SVD is an image-to-video (I2V) model. The model needs a starting image to generate a short clip from it. SVD does not accept text prompts; motion is inferred from the image content and a handful of parameters.
Text-to-Video vs Image-to-Video
Image-to-video (I2V) animates a still image you supply. SVD is an I2V model: give it a portrait, a landscape, or a product shot, and it produces a clip with natural motion. Text-to-video (T2V) generates video from a text prompt, without a starting frame. SVD does not support T2V natively.
Modern DiT-based models like Wan 2.2 handle both modes natively in a single model.
The Stable-Diffusion-Native Ways to Make Video
Stable Video Diffusion (SVD)
- SVD (14 frames, 576x1024)
- SVD-XT (25 frames, 576x1024).
SVD-XT is the standard choice. Both are open-source with weights on Hugging Face. ComfyUI is the recommended way to run SVD; Stability AI's earlier Streamlit demo is no longer reliably maintained and frequently breaks on current dependency versions.
AnimateDiff
AnimateDiff is a motion module that plugs into existing Stable Diffusion 1.5 checkpoints rather than replacing them. You add a temporal layer to an SD image model you already have, including community fine-tunes and LoRAs. The result is text-prompted animation: write a prompt, choose a motion preset, and get a short looping clip in your checkpoint's visual style. AnimateDiff runs in both ComfyUI and Forge Neo, and fits better than SVD when you want text-driven motion within a specific fine-tuned aesthetic.
How to Run SVD in ComfyUI
ComfyUI is the standard interface for SVD in 2026. Its node-based pipeline exposes every step of the generation process, makes workflows portable as JSON files, and handles complex multi-model setups that single-tab UIs cannot match. The Thunder Compute ComfyUI template launches a pre-configured instance with ComfyUI already running, with no local installation required.
Image-to-Video Workflow
The SVD workflow in ComfyUI connects five nodes: an image loader, the SVD model checkpoint, a video KSampler, a VAE decoder, and a video combine node. SVD-XT weights are available on Hugging Face. Once the model is loaded, select the SVD-XT checkpoint in the Image Only Checkpoint Loader node, drop in your source image, and queue the prompt.
SVD outputs a WebP animation by default. To export MP4, add a Video Combine node with the appropriate codec settings, or use a community workflow that includes this step.
Key Parameters: Motion, FPS, Frames, and Augmentation
Four parameters control SVD output:
- Motion Bucket ID sets the amount of motion in the clip. The range is 0-255, and the default of 127 produces moderate motion. Higher values increase movement but can introduce instability.
- FPS sets the playback frame rate, not the number of frames generated. SVD-XT generates 25 frames; FPS 10 produces a 2.5-second clip, FPS 25 produces a 1-second clip.
- Frames should stay at 25 for SVD-XT and 14 for the base SVD model. These match the training values; changing them degrades output quality.
- Augmentation Level adds noise to the input image before encoding. Increase it when source image dimensions differ from SVD's native 576x1024, or to have the output deviate more from the starting frame.
Learn how to run ComfyUI on Thunder Compute.
Stable Video Diffusion VRAM Requirements
SVD-XT requires roughly 8-10GB of VRAM at standard 576x1024 resolution, making it lighter than most 2026 video models. In practice, users on 16GB-24GB cards might hit CUDA out-of-memory errors due to poorly optimized workflows. The issue is usually pipeline overhead, not raw model weight.
| Method | Min VRAM | Recommended VRAM | Notes |
|---|---|---|---|
| SVD-XT in ComfyUI | 8 GB | 16 GB+ | ComfyUI's dynamic VRAM management offloads unused models between steps |
| SVD (14 frames) | 8 GB | 12 GB+ | Lighter than SVD-XT; fewer frames means less temporal memory pressure |
| AnimateDiff in ComfyUI | 6 GB | 12 GB+ | Depends on the SD checkpoint used; SD 1.5 is lighter than SDXL |
| Requirements vary by resolution, frame count, and enabled extensions. | |||
Fixing CUDA Out-of-Memory Errors
Most SVD CUDA OOM errors have a fixable cause before the problem becomes a hardware one. In ComfyUI, enable --lowvram in the launch flags to offload model components aggressively between steps. Generate at SVD's native 576x1024 resolution; non-standard sizes increase memory pressure. Unload other models before queuing the SVD workflow, and restart ComfyUI to clear memory leaks from long sessions.
If OOM persists after these steps, the GPU lacks sufficient VRAM headroom. The practical fix is more VRAM.
Run SVD on a Cloud GPU
Thunder Compute's one-click ComfyUI template is the fastest way to run SVD without OOM errors. The template launches a pre-configured ComfyUI instance on a cloud GPU with no local installation required.
One-Click ComfyUI on Thunder Compute
Three commands get you running:
tnr create --template comfy-ui- Launch an instance creation wizard.tnr connect 0- Connect to the created instancestart comfyui- Start ComfyUI on the remote server. This will take a minute or two.
ComfyUI starts and prints a public URL in the terminal. Load the SVD workflow, add the SVD-XT model weights, and start generating. Snapshots keep downloaded models available across sessions, so you are not re-downloading weights each time.
The RTX A6000 48GB at $0.35/hr is the right GPU for SVD. At 48GB GDDR6 and 768 GB/s of memory bandwidth, there is no memory pressure for SVD-XT at standard resolution. The same instance handles AnimateDiff and Wan 2.2's 5B variant without reconfiguration. If you plan to test all three models make sure to create an instance with sufficient disk space for the weights.
Local vs Cloud: Cost and Break-Even
Cloud GPU is almost always cheaper for occasional use or time-boxed projects. A consumer RTX 4090 costs over $3,800 up front, while an RTX A6000 is a professional workstation card that can cost up to $6,800. At $0.35/hr on Thunder Compute, you pay only for active generation time with no upfront hardware cost and no recurring electricity bill.
The RTX A6000 48GB handles SVD-XT at standard resolution without memory pressure, making it a capable choice for iterative SVD work.
Is SVD Still Worth Using in 2026?
SVD still works well for I2V at 576x1024. SVD-XT produces stable, natural motion with predictable parameters and solid ComfyUI support. If you have an existing SD workflow and want to animate outputs without switching model families, SVD is a reasonable path.
For new projects, DiT-based models have moved ahead. Wan 2.2 handles both T2V and I2V in a single model, runs on 8GB VRAM with the 5B GGUF variant, and produces higher motion realism on most subjects. LTX-2.3 adds native audio and 4K output at 50fps.
Last Thoughts on Stable Diffusion Video Generation
SVD-XT is a capable img2vid model with predictable parameters and solid ComfyUI support. It runs on 8GB VRAM in theory, but 16GB+ is the comfortable floor for stable sessions. For new projects where motion quality is the priority, DiT-based models like Wan 2.2 have moved ahead. Regardless, SVD remains a practical choice if you are already working within the Stable Diffusion ecosystem.
FAQ
Can Stable Diffusion generate video, or only images?
Stable Diffusion is an image model, but two approaches add video output to the ecosystem. Stable Video Diffusion (SVD) is a standalone Stability AI model trained on video data; it takes a still image and produces a short clip. AnimateDiff is a motion module that plugs into existing SD 1.5 checkpoints without changing the image weights, and accepts text prompts. SVD produces better motion quality; AnimateDiff offers more style control. Neither is part of the base Stable Diffusion model.
What VRAM do I need for Stable Video Diffusion?
SVD-XT in ComfyUI runs on 8GB minimum, with 16GB recommended. OOM errors on 16-24GB cards are usually caused by pipeline overhead, not model size. Enable --lowvram in ComfyUI launch flags and run SVD as a standalone session to resolve most issues.
Why do I keep getting CUDA out-of-memory errors with Stable Video Diffusion?
SVD OOM errors are usually caused by pipeline overhead, not the model weight. In ComfyUI, add --lowvram to launch flags, unload other models first, and generate at SVD's native 576x1024 resolution. If errors persist, the GPU lacks sufficient VRAM headroom.
Should I use Stable Video Diffusion or AnimateDiff?
SVD animates a still image using a dedicated video model; use it in ComfyUI for I2V. AnimateDiff adds motion to an SD 1.5 checkpoint and supports text prompts; it runs in ComfyUI and Forge Neo.
Is Stable Video Diffusion free to use?
SVD and SVD-XT weights are open-source and free on Hugging Face. ComfyUI is also free. The cost is compute. Running SVD on a capable cloud GPU starts at $0.35/hr on Thunder Compute with per-minute billing.
How do I export Stable Video Diffusion output as MP4?
SVD outputs WebP animation by default in ComfyUI. To export MP4, add a Video Combine node with the appropriate codec settings, or load a community workflow that includes this step.