How to run AnimateDiff in ComfyUI on a cloud GPU

Rent a GPU on Deploy, then open a terminal in JupyterLab or over SSH. Install ComfyUI from the official repository, then install ComfyUI-AnimateDiff-Evolved with ComfyUI-Manager or a git clone into custom_nodes, and download a motion module into one of the two folders that README names. Load a workflow image from that repository’s samples page, run it, and copy the output folder off the instance before you destroy it. Destroying the instance deletes its disk.

The node README points at the guoyww/AnimateDiff repository for how the motion module works. Other ComfyUI pages: How to install ComfyUI and ComfyUI-Manager on a cloud GPU, How to run FLUX.1-dev in ComfyUI on a cloud GPU, How to run Wan 2.2 video generation on a cloud GPU. How published FLUX and SDXL weight sizes are read: How much VRAM do I need for FLUX and SDXL. Image conditioning on that same ComfyUI checkout: How to use IPAdapter in ComfyUI on a cloud GPU.

The Evolved LICENSE file is the Apache License, Version 2.0, January 2004, and its appendix copyright line is “Copyright 2023 Jedrzej Kosinski” (read 2026-11-12). The GitHub license field the same day was Apache-2.0. pyproject.toml sets version to 1.6.0 and requires-comfyui to >=0.3.68. The guoyww LICENSE.txt is the same Apache license text, and its appendix copyright line is still the unfilled template Copyright [yyyy] [name of copyright owner]. The guoyww README disclaimer says “This project is released for academic use.” That sentence and the Apache file are both in the repository. This page does not merge them into a third license. Read both before you redistribute the program.

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 READMEs say about memory

The Evolved README feature list says “fp8 support; requires newest ComfyUI and torch >= 2.1 (decreases VRAM usage, but changes outputs)”. It does not print a GB figure. Its FP16 bullet says those files take up less storage space, “but uses up the same amount of VRAM as ComfyUI loads models in fp16 by default”. That is a comparison with no GB number. The nodes document says Context Options work by diffusing portions of the animation at a time, “effectively limiting VRAM usage to be equivalent to be context_length latents.” View Options “does NOT decrease VRAM usage”. Neither sentence prints a GB figure. This page did not re-measure them.

The guoyww README’s SDXL-beta section says high resolution videos, “i.e., 1024x1024x16 frames with various aspect ratios”, “could be produced with/without personalized models. Inference usually requires ~13GB VRAM and tuned hyperparameters (e.g., sampling steps), depending on the chosen personalized models.” That sentence is about the sdxl branch it names in the next line. It is not a figure in the Evolved README. The guoyww model card fetched the same day had license apache-2.0 and no GB figure. The Evolved Hugging Face motion-module pages are linked from the README; this page does not treat a file-size column as VRAM.

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 2026-11-12T23:03:50.976Z. GET /v1/public/gpu-catalog the same day (generated_at 2026-11-12T23:03:51.922Z) 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. RTX 2080 is 8, RTX 3080 is 10, RTX 2080 Ti is 11, and RTX 3080 Ti is 12, which are below 13 as numbers. No catalog row is 13. V100 16G is 16, which is above 13 as a number. This page does not mark a row as enough for the SDXL-beta sentence, and it does not apply that sentence to the Evolved node. 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 AnimateDiff template. This page installs the node into a ComfyUI checkout. Installing ComfyUI is the guide linked above.

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

Install the ComfyUI node

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 AnimateDiff README. The ComfyUI install, including Manager, is How to install ComfyUI and ComfyUI-Manager on a cloud GPU. This page does not repeat that install.

The Evolved README’s Manager section says:

  1. Look for AnimateDiff Evolved, and be sure the author is Kosinkadink. Install it.

The manual section says “Clone this repo into custom_nodes folder.” It does not print a git command. This page adds the clone, and it was not run:

cd ComfyUI/custom_nodes
git clone https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved

The commits API on 2026-11-12 listed 9257651221002dcba0a12f9cff37e1944e58fb60 on the default branch main, committed 2026-07-28T11:39:28Z, message “Merge pull request #579 from Kosinkadink/chore/bump-version-1.6.0”. The README has no project-status heading.

The README also names other node packs that workflows often use, including ComfyUI-VideoHelperSuite for loading videos and combining images into videos. It does not print a clone command for that pack. This page does not add one.

Motion modules and Motion LoRAs

The Evolved README says you need at least one motion module. It names the original modules mm_sd_v14, mm_sd_v15, mm_sd_v15_v2, and v3_sd15_mm, and links a Hugging Face tree, a drive folder, and a CivitAI page. This page does not repeat the drive hostname. The same bullet lists stabilized and higher-resolution finetunes by the names mm-Stabilized_mid, mm-Stabilized_high, mm-p_0.5.pth, mm-p_0.75.pth, and temporaldiff-v1-animatediff, each with the link the README prints. Place the files in one of these locations. The README says you can rename them:

  • ComfyUI/custom_nodes/ComfyUI-AnimateDiff-Evolved/models
  • ComfyUI/models/animatediff_models

animatediff/utils_model.py sets Folders.ANIMATEDIFF_MODELS to animatediff_models and registers both the repository models directory and models/animatediff_models under the ComfyUI models directory. Motion LoRAs go in one of these locations:

  • ComfyUI/custom_nodes/ComfyUI-AnimateDiff-Evolved/motion_lora
  • ComfyUI/models/animatediff_motion_lora

The same file sets Folders.MOTION_LORA to animatediff_motion_lora and registers both of those directories. The README says the folder ids in extra_model_paths.yaml are animatediff_models and animatediff_motion_lora. The guoyww v2 section names eight MotionLoRA files: v2_lora_ZoomIn.ckpt, v2_lora_ZoomOut.ckpt, v2_lora_PanLeft.ckpt, v2_lora_PanRight.ckpt, v2_lora_TiltUp.ckpt, v2_lora_TiltDown.ckpt, v2_lora_RollingClockwise.ckpt, and v2_lora_RollingAnticlockwise.ckpt. The Evolved README says the original LoRAs “really only work on v2-based motion models” such as mm_sd_v15_v2, mm-p_0.5.pth, and mm-p_0.75.pth.

The guoyww README says the main branch is for Stable Diffusion V1.5 and the sdxl-beta branch is for Stable Diffusion XL. The Evolved README’s HotshotXL note says the file is hsxl_temporal_layers.safetensors, that you need an SDXL checkpoint, and that “the sweetspot for context_length or total frames (when not using context) is 8 frames”. Its AnimateDiff-SDXL note says to use the autoselect or linear (AnimateDiff-SDXL) beta schedule. Its AnimateLCM note names autoselect, lcm, and lcm[100_ots]. Those are schedule names in the README, not measured timings from this page.

Load an example workflow

The samples document begins “download or drag images of the workflows into ComfyUI to instantly load the corresponding workflows.” Its updated sections are titled txt2vid, txt2vid with prompt travel, “txt2vid - 32 frame animation with 16 context_length”, and the same 32-frame setup with ContextRef. An older section is titled “48 frame animation with 16 context_length (uniform)”. The nodes document’s Standard Static caption uses latent count 64, context_length 16, context_overlap 4, and total steps 20. Those captions are examples in the documents. This page did not run them.

The ComfyUI examples top-level directory list on 2026-11-12 had no directory named for this node. The latest commit on that repository was f9431bb000ce792094ff345446e22cac1ea6cef3, committed 2025-11-26T20:33:58Z. This page does not treat another example directory as an AnimateDiff install.

Context options, as the nodes document names them

The nodes document’s “Context Options and View Options” section names two kinds, Standard and Looped, and says “I recommend using Standard Static at first when not wanting looped outputs.” The Context Options nodes it documents are Standard Static, Standard Uniform, Looped Uniform, and Views Only. The inputs it names on those nodes include context_length, context_overlap, context_stride, closed_loop, fuse_method, use_on_equal_length, start_percent, guarantee_steps, prev_context, and view_options. View Options use view_length, view_overlap, view_stride, and closed_loop. closed_loop is described as adding windows to enhance looping when True. The same document says Gen2’s Use Evolved Sampling node can use Context Options and Sample Settings without a motion model. This page does not add values beyond the captions above.

Where frames and videos are saved

The Evolved README says ComfyUI-VideoHelperSuite is for combining images into videos. The VHS README’s Video Combine section says save_output chooses “the output directory or the temp directory”, and that filename_prefix is the base file name, including a subfolder such as subfolder/video. It also says the frame rate “should usually be kept to 8 for AnimateDiff”. videohelpersuite/nodes.py on commit 4d907bee61e92c2e65af3bd6383a4e4d356126d1 (2026-09-02T06:48:06Z) sets the directory with folder_paths.get_output_directory() when save_output is true, and folder_paths.get_temp_directory() otherwise. The GitHub license field for that repository the same day was GPL-3.0.

ComfyUI folder_paths.py on commit c1eece62cbf1805d4f64286d7f5592e6a215e6ba (2026-11-12T20:46:52Z) sets output_directory to base_path joined with output, and temp_directory to base_path joined with temp. 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.

animatediff/nodes_deprecated.py still defines ADE_AnimateDiffCombine. Its display name says “DEPRECATED, Use Video Combine (VHS) Instead”. The execute method warns “Do not use AnimateDiff Combine node, it is deprecated.” When save_image is true, which is the default, it writes under folder_paths.get_output_directory() with filename_prefix default AnimateDiff. Video formats in the repository video_formats directory include h264-mp4.json, whose extension field is mp4. This page follows the README and uses the VHS node the warning names. It does not treat the deprecated node as the install.

The guoyww README, for the official repository rather than the ComfyUI node, says “The generated samples can be found in samples/ folder.”

Official guoyww repository

This block is the official repository. It is not the ComfyUI node install above. The README quick start prints:

git clone https://github.com/guoyww/AnimateDiff.git
cd AnimateDiff
pip install -r requirements.txt
python -m scripts.animate --config configs/prompts/1_animate/1_1_animate_RealisticVision.yaml

The same section prints six more --config paths under configs/prompts/1_animate/, plus MotionLoRA and SparseCtrl configs. requirements.txt pins torch==2.3.1, torchvision==0.18.1, diffusers==0.11.1, transformers==4.25.1, xformers==0.0.27, and gradio==3.36.1. It does not pin a Python version. The commits API on 2026-11-12 listed e92bd5671ba62c0d774a32951453e328018b7c5b on main, committed 2024-07-17T08:19:47Z, message “update readme”.

The Gradio section says the demo runs at localhost:7860 and prints python -u app.py. app.py calls demo.launch(share=True) and does not pass server_name or server_port. The Gradio 3.36.1 Blocks.launch docstring says a None server name uses 127.0.0.1, a None server port searches for an available port starting at 7860, and share=True creates a publicly shareable link. This page does not claim the search stays on 7860. It does not edit app.py, so the public link is what that call requests. How to use SSH port forwarding is the tunnel for the local address, and that guide’s share-link section is the note on public Gradio links. This page did not run app.py.

The README also says AnimateDiff is officially supported by Diffusers and links a tutorial. This page does not copy that tutorial.

To keep the process after the SSH session closes, this page wraps the README command in tmux. The README does not print tmux:

tmux new -s animatediff
cd AnimateDiff
python -u app.py

When it fails

CUDA out of memory. The Evolved nodes document says Context Options limit sampling to context_length latents and does not print a further procedure. The guoyww README does not print an out-of-memory flag. 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 Evolved README’s torch line is torch >= 2.1 for fp8. The guoyww requirements.txt pins torch==2.3.1. Neither file names compute capability 12.0. The RTX 5090 guide says that card is compute capability 12.0 (sm_120) and needs a PyTorch build for CUDA 12.8 or newer. This page did not install a torch wheel and did not re-test either repository on an RTX 5090. Catalog RTX 5090 and RTX 5090D are 32. That is the catalog number only.

No motion module. animatediff/__init__.py logs “No motion models found” when the model folder list is empty, and the message includes the folder paths from folder_paths.get_folder_paths. Put one file in a folder the motion-module section names.

Save the files, then destroy

Copy the ComfyUI output directory, and samples/ if you ran the official repository scripts, off the instance before you destroy it. 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 do the AnimateDiff READMEs publish?

The ComfyUI-AnimateDiff-Evolved README and its nodes document publish no GB figure. The fp8 line says it decreases VRAM usage and names torch >= 2.1, with no GB number. The guoyww README says SDXL-beta inference usually requires ~13GB VRAM for high resolution videos, and it gives 1024x1024x16 frames as the example. That sentence is not the Evolved node requirement. Every vram_gb in https://cloudgpu.app/gpu-prices.json (updated_at 2026-11-12T23:03:50.976Z) is 8 or higher. RTX 2080 is 8, RTX 3080 is 10, RTX 2080 Ti is 11, and RTX 3080 Ti is 12, which are below 13 as numbers. No catalog row is 13. V100 16G is 16, which is above 13 as a number. 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 motion modules go, and where are videos saved?

The Evolved README says to place motion modules in ComfyUI/custom_nodes/ComfyUI-AnimateDiff-Evolved/models or ComfyUI/models/animatediff_models. utils_model.py registers those two folders. Motion LoRAs go in the repository motion_lora directory or ComfyUI/models/animatediff_motion_lora. The README says ComfyUI-VideoHelperSuite combines images into videos. The VHS README says save_output chooses the output directory or the temp directory. videohelpersuite/nodes.py uses folder_paths.get_output_directory() when save_output is true. ComfyUI folder_paths.py joins base_path with output. The deprecated AnimateDiff Combine node also writes under get_output_directory() with filename_prefix AnimateDiff, and its warning says to use Video Combine from VideoHelperSuite instead. The guoyww README says the sampling script writes samples/.

How do I open the official AnimateDiff Gradio demo?

The guoyww README says to run python -u app.py and that the demo will run at localhost:7860. app.py calls demo.launch(share=True) and does not pass server_name or server_port. requirements.txt pins gradio==3.36.1. That version launch docstring says a None server name uses 127.0.0.1, a None port searches for an available port starting at 7860, and share=True creates a publicly shareable link. This page does not claim the search stays on 7860 and does not edit app.py. The SSH tunnel uses the local address. This page did not run the demo.

Do the videos stay on the instance after I destroy it?

No. Copy the ComfyUI output directory, and samples/ if you used the official repository scripts, off the instance before you destroy it. The terms say data on local instance storage is permanently deleted when the instance is destroyed.

Sources (read 2026-11-12, Asia/Shanghai)

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

Last updated: 2026-11-12 (Asia/Shanghai).