How to run Stable Video Diffusion in ComfyUI on a cloud GPU

Q: How much VRAM does the SVD model card publish?

The ComfyUI video examples page publishes no GB figure. The model card says the 14-frame model generates 14 frames at 576x1024 and the 25-frame model generates 25 frames at that resolution. It says SVD takes ~100s and SVD-XT takes ~180s on an A100 80GB card with the default settings at release, and that optimizations can trade quality, memory, and speed for inference on lower VRAM cards. That sentence prints no GB target. This page did not measure it. simple_video_sample.py says decoding_t eats most VRAM and to decrease it if you run out. Its default is 14. That file prints no GB figure. Every vram_gb in https://cloudgpu.app/gpu-prices.json (updated_at 2026-11-13T01:07:50.488Z) is 8 or higher. The catalog row A100 80G is 80. This page does not mark a row as enough. The catalog template named ComfyUI has id flux-1-dev, recommended_gpu RTX 4090 48G, and min_vram_gb 24. Those are template fields. Which listing you can start is on https://cloudgpu.app/deploy.

Q: Where do the checkpoints go, and where is the video saved?

The examples page says to put svd.safetensors and svd_xt.safetensors in ComfyUI/models/checkpoints. The 14-frame workflow JSON loads svd.safetensors. The 25-frame workflow JSON loads svd_xt.safetensors. SaveAnimatedWEBP writes a webp file under the output directory. The code names the file with the prefix, a counter, and the .webp suffix. simple_video_sample.py writes a jpg and an mp4 under outputs/simple_video_sample/svd/ or outputs/simple_video_sample/svd_xt/. This page did not run either path.

Q: How do I open a page for this model?

ComfyUI is the page for the examples workflow. The install guide documents 127.0.0.1 port 8188 and the SSH tunnel. The generative-models README prints python -m scripts.demo.gradio_app and streamlit run scripts/demo/video_sampling.py. gradio_app.py calls demo.launch(share=True) and does not pass a port. The streamlit command does not pass a port. This page did not launch either demo and does not name a port for them.

Q: What license is on the model card?

LICENSE.md is the Stability AI Community License Agreement, last updated July 5, 2024. The file says this Agreement is intended to allow research, non-commercial, and limited commercial uses of the Models free of charge. It says this Agreement preserves free access to the Models for people or organizations generating annual revenue of less than US $1,000,000 (or local currency equivalent). It also says that if at any time You or Your Affiliate(s), either individually or in aggregate, generate more than USD $1,000,000 in annual revenue (or the equivalent thereof in Your local currency), regardless of whether that revenue is generated directly or indirectly from the Stability AI Materials or Derivative Works, any licenses granted to You under this Agreement shall terminate as of such date. Those sentences are license text, not a CloudGPU price and not a storage allowance. The November 21, 2023 README bullet still says the release is for research purposes. This page does not merge that bullet with the license file.

Rent a GPU on Deploy, then open a terminal in JupyterLab or over SSH. Install ComfyUI from the official repository, download the two checkpoints the official video examples name, and put them in ComfyUI/models/checkpoints. Load the image-to-video workflow from that examples page, run it, and copy the output folder off the instance before you destroy it. Destroying the instance deletes its disk.

Other video pages: How to run AnimateDiff in ComfyUI on a cloud GPU, How to run CogVideoX on a cloud GPU, How to run LTX-Video on a cloud GPU, and How to run Wan 2.2 video generation on a cloud GPU. If the process runs out of memory: How to fix CUDA out of memory on a cloud GPU. The tunnel for ComfyUI is How to use SSH port forwarding on a cloud GPU.

The examples page links stabilityai/stable-video-diffusion-img2vid file svd.safetensors for the 14-frame checkpoint and stabilityai/stable-video-diffusion-img2vid-xt file svd_xt.safetensors for the 25-frame checkpoint. Both model cards, read 2026-11-13, had gated false, license other, license_name stable-video-diffusion-community, and license_link LICENSE.md. lastModified was 2024-07-10T11:41:17.000Z for the 14-frame card and 2024-07-10T11:43:18.000Z for the 25-frame card. LICENSE.md is the same file on both cards: the Stability AI Community License Agreement, last updated July 5, 2024. The revenue sentences in that file are quoted in the FAQ. The November 21, 2023 bullet in the generative-models README still says the release is for research purposes. This page does not merge that bullet with the later license file. Read the license before you use the weights.

The commands below are copied from the upstream files named in each step. They were not run on a CloudGPU instance for this page. Anything this page adds around them is marked.

What the model card says about memory

The ComfyUI video examples page does not print a GPU GB figure. The 14-frame model card says the model was trained to generate 14 frames at resolution 576x1024 given a context frame of the same size. The 25-frame card says it was trained to generate 25 frames at that resolution, finetuned from the 14-frame model. The limitations section says the generated videos are rather short (<= 4sec). The appendix says: “With the default settings at the time of release, SVD takes ~100s for generation, and SVD-XT takes ~180s on an A100 80GB card. Several optimizations to trade off quality / memory / speed can be done to perform faster inference or inference on lower VRAM cards.” Those are the card’s sentences. This page did not measure them and does not treat them as a result on a CloudGPU listing. The sentence about lower VRAM cards prints no GB figure.

scripts/sampling/simple_video_sample.py in Stability-AI/generative-models, a different repository from the ComfyUI examples, sets decoding_t default 14 with the comment “Number of frames decoded at a time! This eats most VRAM. Reduce if necessary.” The docstring says: “If you run out of VRAM, try decreasing decoding_t.” That file prints no GB figure. The November 21, 2023 SVD bullet names that script. The --encoding_t=1 and --decoding_t=1 lines later in the same README are under the SV4D sections. This page does not copy those flags onto SVD.

GET /v1/public/compute-capabilities does not list GPUs. It returns feature flags. The memory column is vram_gb from gpu-prices.json, updated_at 2026-11-13T01:07:50.488Z. GET /v1/public/gpu-catalog the same day (generated_at 2026-11-13T01:07:55.359Z) listed the same names and the same vram_gb values, plus Ascend 910B at 64, which is not in gpu-prices.json and is not listed here. The catalog response also includes an availability field. This page does not copy it. Every prices row has gpu_count 1. The catalog row named A100 80G has vram_gb 80. That is the catalog number only. This page does not mark a row as enough, because neither the examples page nor the model card publishes a GB target to compare. Which listing you can start is on Deploy.

The catalog template named ComfyUI has id flux-1-dev, category image-gen, recommended_gpu RTX 4090 48G, and min_vram_gb 24. Those are template fields. The template list had no Stable Video Diffusion template. The ComfyUI README’s own memory clause is on the install guide linked above. This page does not treat that clause as an SVD figure.

Catalog name	vram_gb
RTX 2080	8
RTX 3080	10
RTX 2080 Ti	11
RTX 3080 Ti	12
V100 16G	16
A30	24
RTX 3090	24
RTX 4090	24
RTX 4090D	24
RTX 4080 SUPER 32G	32
RTX 5090	32
RTX 5090D	32
V100 32G	32
A100 40G	40
A800 40G	40
A40	48
RTX 4090 48G	48
RTX 4090D 48G	48
RTX A6000	48
NVIDIA RTX PRO 5000 72G	72
RTX PRO 5000 72G	72
A100 80G	80
A800 80G	80
H800	80
RTX 6000D	83
H20 96G	96
RTX PRO 6000 96G	96
RTX PRO 6000 Max-Q	96
H20-3e 141G	141

Checkpoints and the ComfyUI workflow

On Deploy, choose Raw GPU if you are installing ComfyUI yourself. The deploy page says “SSH access and root within a dedicated container”. The ComfyUI template is a different option. Its id and min_vram_gb are the template fields above, not a line in the video examples. The ComfyUI install is How to install ComfyUI and ComfyUI-Manager on a cloud GPU. This page does not repeat that install.

The examples page says to put svd.safetensors and svd_xt.safetensors in the ComfyUI/models/checkpoints folder. It does not print a download command. The links it prints are the Hub blob pages for those two files. The 25-frame card also links svd_xt_image_decoder.safetensors, and the 14-frame card links svd_image_decoder.safetensors. The examples page does not name those decoder files. This page did not download any of them.

How to download a Hugging Face model is the token reference. The examples page does not print an HF_ENDPOINT line. Both cards were gated false on the day above.

The examples page says you can download image_to_video.webp or drag it onto ComfyUI to get the workflow that uses the 14-frame model. workflow_image_to_video.json is linked beside it. workflow_txt_to_img_to_video.json is the workflow that uses SDXL to make an initial image and then the 25-frame model. This page did not load either file.

workflow_image_to_video.json sets ImageOnlyCheckpointLoader to svd.safetensors and SVD_img2vid_Conditioning widgets to 1024, 576, 14, 127, 6, 0. workflow_txt_to_img_to_video.json sets ImageOnlyCheckpointLoader to svd_xt.safetensors and those widgets to 1024, 576, 25, 127, 6, 0. ComfyUI comfy_extras/nodes_video_model.py on commit c1eece62cbf1805d4f64286d7f5592e6a215e6ba (2026-11-12T20:46:52Z) defines those inputs in this order: width default 1024, height default 576, video_frames default 14, motion_bucket_id default 127, fps default 6, augmentation_level default 0.0. The model card’s resolution line is 576x1024. This page does not rewrite one number into the other.

The examples page explains video_frames, motion_bucket_id, fps, and augmentation level in words and does not print those widget numbers. It says VideoLinearCFGGuidance linearly scales the cfg across frames, and that in the example the first frame is cfg 1.0 (the min_cfg in the node), the middle frame is 1.75, and the last frame is 2.5 (the cfg set in the sampler). The 14-frame workflow JSON sets the VideoLinearCFGGuidance widget to 1 and the KSampler cfg widget to 2.5, with 20 steps, sampler name euler, and scheduler karras. This page did not run the sampler.

The commits API for video/README.md listed 8a21245c6934fd5875a2e694b362686e7c023022, committed 2023-11-28T19:49:34Z, message “Add SDXL turbo example.” The examples repository HEAD the same day was f9431bb000ce792094ff345446e22cac1ea6cef3, committed 2025-11-26T20:33:58Z. The top-level directory list included video. It had no separate directory named for this model.

The research repository and its demos

The model card says to get started at https://github.com/Stability-AI/generative-models. The commits API on 2026-11-13 listed e8cd657656fa5d61688191730d0e03242bf4ed44 on main, committed 2025-12-16T08:36:15Z, message “Merge pull request #467 from Stability-AI/deprecate”.

The November 21, 2023 bullet says you can run the community-build Gradio demo locally with:

python -m scripts.demo.gradio_app

The spelling community-build is the README’s. scripts/demo/gradio_app.py on that commit calls demo.launch(share=True) and does not pass a port. This page did not launch it and does not name a port for it.

The same bullet says the repository provides a streamlit demo scripts/demo/video_sampling.py and scripts/sampling/simple_video_sample.py. A later line in the README says, for SVD or SV3D:

streamlit run scripts/demo/video_sampling.py

That command does not pass a port. video_sampling.py on that commit does not set one. This page did not run it and does not name a port for it.

simple_video_sample.py on that commit defaults version to svd. For svd it sets num_frames 14, num_steps 25, and output_folder outputs/simple_video_sample/svd/. For svd_xt it sets num_frames 25, num_steps 30, and output_folder outputs/simple_video_sample/svd_xt/. It writes {base_count:06d}.jpg and {base_count:06d}.mp4 in that folder. The November 21 bullet does not print an install command. This page does not copy the dependency block that sits under the SV4D 2.0 section. This page did not run the script. The route this page walks through is the ComfyUI example.

ComfyUI is the web UI for that route. The install guide documents 127.0.0.1 port 8188 and the tunnel. This page does not add a port command.

Where the ComfyUI video is saved

workflow_image_to_video.json and workflow_txt_to_img_to_video.json both use SaveAnimatedWEBP with widgets ComfyUI, 10, false, 85, default. comfy_extras/nodes_images.py on the ComfyUI commit above calls UI.ImageSaveHelper.get_save_animated_webp_ui. comfy_api/latest/_ui.py on that commit saves through folder_paths.get_save_image_path on the output directory when the folder type is output, and the filename is {filename}_{counter:05}_.webp. The node schema defaults are filename_prefix ComfyUI, fps 6.0, lossless true, and quality 80. The workflow JSON is the override above. folder_paths.py sets output_directory to base_path joined with output. base_path is --base-directory when that argument is set, otherwise the directory of folder_paths.py. The latest release tag fetched the same day was v0.45.0, published 2026-11-10T21:42:42Z.

When it fails

CUDA out of memory. The examples page does not print an out-of-memory flag or a GB target. The sampling script’s own line is to decrease decoding_t. That flag is in the research script, not in the ComfyUI workflow JSON this page read. If the process raises that error, use how to fix CUDA out of memory on a cloud GPU.

“CUDA error: no kernel image is available for execution on the device” on an RTX 5090. The examples page does not pin torch and does not name compute capability 12.0. The RTX 5090 guide says that card is compute capability 12.0 (sm_120) and that PyTorch 2.7 and later, on a build for CUDA 12.8 or newer, is the line that announced support. This page did not install a torch wheel and did not re-test this workflow on an RTX 5090. Catalog RTX 5090 and RTX 5090D are 32. That is the catalog number only.

Save the video, then destroy

Copy the ComfyUI output directory, including the .webp files from SaveAnimatedWEBP, off the instance before you destroy it. If you also ran simple_video_sample.py, copy the outputs/simple_video_sample/ directory that script defaults to. Save your work before you destroy a cloud GPU is the checklist. The terms say data on local instance storage is permanently deleted when the instance is destroyed.

The pricing page FAQ states the billing rule in these words:

Per-minute options have a 1-minute minimum; unused prepaid minutes are refunded when you destroy the instance. Some configurations use whole-hour billing with a 1-hour minimum; unused time within a charged hour is not refunded on voluntary stop. The billing period, minimum charge and initial prepayment are shown before confirmation. No subscription.

Prices are not copied here.

Frequently asked questions

How much VRAM does the SVD model card publish?

The ComfyUI video examples page publishes no GB figure. The model card says the 14-frame model generates 14 frames at 576x1024 and the 25-frame model generates 25 frames at that resolution. It says SVD takes ~100s and SVD-XT takes ~180s on an A100 80GB card with the default settings at release, and that optimizations can trade quality, memory, and speed for inference on lower VRAM cards. That sentence prints no GB target. This page did not measure it. simple_video_sample.py says decoding_t eats most VRAM and to decrease it if you run out. Its default is 14. That file prints no GB figure. Every vram_gb in https://cloudgpu.app/gpu-prices.json (updated_at 2026-11-13T01:07:50.488Z) is 8 or higher. The catalog row A100 80G is 80. This page does not mark a row as enough. The catalog template named ComfyUI has id flux-1-dev, recommended_gpu RTX 4090 48G, and min_vram_gb 24. Those are template fields. Which listing you can start is on https://cloudgpu.app/deploy.

Where do the checkpoints go, and where is the video saved?

The examples page says to put svd.safetensors and svd_xt.safetensors in ComfyUI/models/checkpoints. The 14-frame workflow JSON loads svd.safetensors. The 25-frame workflow JSON loads svd_xt.safetensors. SaveAnimatedWEBP writes a webp file under the output directory. The code names the file with the prefix, a counter, and the .webp suffix. simple_video_sample.py writes a jpg and an mp4 under outputs/simple_video_sample/svd/ or outputs/simple_video_sample/svd_xt/. This page did not run either path.

How do I open a page for this model?

ComfyUI is the page for the examples workflow. The install guide documents 127.0.0.1 port 8188 and the SSH tunnel. The generative-models README prints python -m scripts.demo.gradio_app and streamlit run scripts/demo/video_sampling.py. gradio_app.py calls demo.launch(share=True) and does not pass a port. The streamlit command does not pass a port. This page did not launch either demo and does not name a port for them.

What license is on the model card?

LICENSE.md is the Stability AI Community License Agreement, last updated July 5, 2024. The file says this Agreement is intended to allow research, non-commercial, and limited commercial uses of the Models free of charge. It says this Agreement preserves free access to the Models for people or organizations generating annual revenue of less than US $1,000,000 (or local currency equivalent). It also says that if at any time You or Your Affiliate(s), either individually or in aggregate, generate more than USD $1,000,000 in annual revenue (or the equivalent thereof in Your local currency), regardless of whether that revenue is generated directly or indirectly from the Stability AI Materials or Derivative Works, any licenses granted to You under this Agreement shall terminate as of such date. Those sentences are license text, not a CloudGPU price and not a storage allowance. The November 21, 2023 README bullet still says the release is for research purposes. This page does not merge that bullet with the license file.

Sources (read 2026-11-13, Asia/Shanghai)

ComfyUI video examples (checkpoint folder, 14-frame and 25-frame Hub files, parameter names, cfg 1.0 / 1.75 / 2.5): https://github.com/comfyanonymous/ComfyUI_examples/blob/master/video/README.md commit 8a21245c6934fd5875a2e694b362686e7c023022; examples HEAD f9431bb000ce792094ff345446e22cac1ea6cef3
workflow_image_to_video.json and workflow_txt_to_img_to_video.json (checkpoint names, conditioning widgets, SaveAnimatedWEBP widgets): that examples directory
SVD and SVD-XT model cards (14 and 25 frames, 576x1024, <= 4sec, ~100s and ~180s on an A100 80GB card, lower VRAM sentence with no GB figure) and LICENSE.md (Stability AI Community License Agreement, July 5, 2024): https://huggingface.co/stabilityai/stable-video-diffusion-img2vid and https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt
generative-models README (November 21, 2023 bullet, Gradio and streamlit commands) and scripts/demo/gradio_app.py (demo.launch(share=True)), scripts/sampling/simple_video_sample.py (decoding_t 14, output folders, mp4): commit e8cd657656fa5d61688191730d0e03242bf4ed44
ComfyUI comfy_extras/nodes_video_model.py (conditioning defaults), comfy_extras/nodes_images.py and comfy_api/latest/_ui.py (animated webp under the output directory), folder_paths.py: commit c1eece62cbf1805d4f64286d7f5592e6a215e6ba; release v0.45.0
Public GPU memory: https://cloudgpu.app/gpu-prices.json (updated_at 2026-11-13T01:07:50.488Z) and https://cloudgpu.app/v1/public/gpu-catalog (generated_at 2026-11-13T01:07:55.359Z). GET /v1/public/compute-capabilities returned feature flags and no GPU list. The catalog template flux-1-dev is named ComfyUI.

Cards you can start are on the Deploy page. Prices are on https://cloudgpu.app/pricing and are not repeated here.

Last updated: 2026-11-13 (Asia/Shanghai).