How to run Stable Diffusion 3.5 in ComfyUI on a cloud GPU
Rent a GPU on Deploy, then open the ComfyUI template. Follow the official ComfyUI SD 3.5 example: put the checkpoint it names in ComfyUI/models/checkpoints/, add the text encoders that page names, drag its example image into ComfyUI, and copy the output folder off the instance before you destroy it. Destroying the instance deletes its disk.
The checkpoint links on that example point at a gated model card. A model API read on 2026-11-12 for each of the three checkpoint repositories reported gated: auto, cardData.license other, cardData.license_name stabilityai-ai-community, and license_link LICENSE.md. This page does not paste those Hub ids. Copy the link from the ComfyUI SD 3.5 example. The LICENSE.md fetched from the large-checkpoint repository the same day is titled STABILITY AI COMMUNITY LICENSE AGREEMENT, last updated July 5, 2024. The introduction says the agreement is intended to allow research, non-commercial, and limited commercial uses, and that it preserves access for people or organizations generating annual revenue of less than US $1,000,000 (or local currency equivalent). Section II grants research and non-commercial use as that section defines those purposes. Section III grants commercial use and says you must register at https://stability.ai/community-license if you use the materials for a commercial purpose. The same section says the licenses terminate if you or your affiliates generate more than USD $1,000,000 in annual revenue, and that you must request a license at https://stability.ai/enterprise. Section IV(a) requires this attribution notice when you distribute the materials or a derivative: “This Stability AI Model is licensed under the Stability AI Community License, Copyright © Stability AI Ltd. All Rights Reserved”. Section IV(b) says you will not use the materials, derivative works, or their outputs to create or improve any foundational generative AI model, excluding the Models or Derivative Works. Section IV©(iii) says that, as between you and the licensor, you own outputs to the extent permitted by applicable law. Read the file. This page does not restate the rest of it. Those dollar figures are the license’s revenue condition. They are not a CloudGPU price.
The commands below are copied from the upstream pages 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 example says about memory
The ComfyUI SD 3.5 example says the family contains a large 8B model and a medium 2.5B model (read 2026-11-12). Those are the README’s parameter counts, not a VRAM figure. The same page publishes no GPU GB figure. It says to use t5xxl_fp16.safetensors if you have more than 32GB ram, and t5xxl_fp8_e4m3fn_scaled.safetensors if you do not. That sentence is system RAM, not a catalog vram_gb.
The model API safetensors.total the same day was 8146280768 for the large repository and for the large-turbo repository, and 2469663936 for the medium repository. Those integers are parameter counts on the model API, not VRAM. The ComfyUI blog post dated 2024-10-22 says the large model has 8 billion parameters, the medium model has 2.6 billion parameters, and the large turbo model produces images in 4 steps. This page does not reconcile the example’s “medium 2.5B”, the API total 2469663936, and the blog’s 2.6 billion. The blog calls the fp8 scaled workflow a low vram option and gives no GB figure. It also says the large model is for a 1-megapixel resolution and the medium model is for resolutions from 0.25 to 2 megapixels. Those are the blog’s resolution notes, not 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 2026-11-12T18:08:32.252Z. GET /v1/public/gpu-catalog the same day (generated_at 2026-11-12T18:08:33.832Z) listed the same names and the same vram_gb values, plus Ascend 910B, 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. There is no upstream GPU GB cutoff to match these values against. Which listing you can start is on Deploy. This page does not say which listings are available.
| 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 |
The catalog template whose name is ComfyUI has id flux-1-dev, recommended_gpu RTX 4090 48G, and min_vram_gb 24 (read 2026-11-12). Those three fields are the template record. They are not a figure from the SD 3.5 example or the model card.
Another checkpoint on the same template: How to run FLUX.1-dev in ComfyUI on a cloud GPU. How published FLUX and SDXL weight sizes are read: How much VRAM do I need for FLUX and SDXL. The card page for the template: How to choose a rental GPU for ComfyUI.
Install on the ComfyUI template
On Deploy, choose the template named ComfyUI. The FLUX.1-dev guide says the template installs ComfyUI only and does not include FLUX weights, citing CloudGPU’s docs as read on 2026-11-03. This page did not re-read those docs. Read the template description on Deploy before you rely on that sentence.
The examples page does not print a shell command that finds the ComfyUI directory. This page adds the same lookup the FLUX.1-dev guide uses. It was not run for this page.
COMFY=$(find / -maxdepth 4 -type d -name ComfyUI 2>/dev/null | head -n 1); echo "$COMFY"
Files the example names
Text encoders, copied from the example: clip_l.safetensors, clip_g.safetensors, and a t5xxl file, in ComfyUI/models/text_encoders/ if they are not already there. The example recommends t5xxl_fp16.safetensors if you have more than 32GB ram, and t5xxl_fp8_e4m3fn_scaled.safetensors if you do not. Those three links on the example point at Comfy-Org/stable-diffusion-3.5-fp8. A GET of https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8 on 2026-11-12 ended at https://huggingface.co/Comfy-Org/stable-diffusion-3.5. A model API request for Comfy-Org/stable-diffusion-3.5-fp8 the same day returned id Comfy-Org/stable-diffusion-3.5, gated: false, and license_name stabilityai-ai-community. The file listing reported these sizes, which are file sizes, not VRAM:
| File | Bytes |
|---|---|
| text_encoders/clip_l.safetensors | 246144152 |
| text_encoders/clip_g.safetensors | 1389382176 |
| text_encoders/t5xxl_fp16.safetensors | 9787841024 |
| text_encoders/t5xxl_fp8_e4m3fn_scaled.safetensors | 5157348688 |
| text_encoders/t5xxl_fp8_e4m3fn.safetensors | 4893934904 |
| sd3.5_large_fp8_scaled.safetensors | 14934922866 |
| sd3.5_medium_incl_clips_t5xxlfp8scaled.safetensors | 11638004202 |
The example says sd3.5_large.safetensors and sd3.5_medium.safetensors go in ComfyUI/models/checkpoints/. It says those files do not contain text encoder weights, so the encoders above are loaded separately, as in the example image sd3.5_text_encoders_example.png. The Hub ids on those two links are not copied here. Copy them from the example. The model API file listing reported sd3.5_large.safetensors at 16460379262 bytes and sd3.5_medium.safetensors at 5107104286 bytes. Those are file sizes, not VRAM. Both repositories reported gated: auto. Accept the license on the model page the example links, then download with a Hugging Face read token. How to download a Hugging Face model is the token reference. The example does not print an hf download command. cloudgpu-net is not on that page. The download speed guide says to leave it off during pip, then turn it on for a weight download.
For sd3.5_large_turbo.safetensors, the example says to put it in ComfyUI/models/checkpoints/ and to use the text-encoder example with steps set to 4 and cfg set to 1.2. The Hub id on that link is not copied here. The model API file listing reported that file at 16460374454 bytes. That is a file size, not VRAM. The repository reported gated: auto and the same license_name. This page did not compare those bytes with sd3.5_large.safetensors.
The all-in-one files, copied from the example: sd3.5_large_fp8_scaled.safetensors and sd3.5_medium_incl_clips_t5xxlfp8scaled.safetensors, both in ComfyUI/models/checkpoints/. The example says the large fp8 file can be used in the default workflow like any other checkpoint. The byte sizes are in the table above. The Comfy-Org README shows the same two checkpoint names under models/checkpoints/ and the text encoders under models/text_encoders/. The 2024-10-22 ComfyUI blog says to put clip_g.safetensors, clip_l.safetensors, and t5xxl_fp16.safetensors in models/clip. This page did not check which folder the template actually loads. The example README is the page this guide follows.
Workflow
Save the example images from the SD 3.5 section and drag them into ComfyUI. The example says that loads the workflow. The text-encoder image is sd3.5_text_encoders_example.png. The single-checkpoint image is sd3.5_simple_example.png.
The example’s controlnet section names sd3.5_large_controlnet_canny.safetensors, sd3.5_large_controlnet_depth.safetensors, and sd3.5_large_controlnet_blur.safetensors, says they go in ComfyUI/models/controlnet, and says they are meant to be used with SD 3.5 large. The Hub id on that link is not copied here. Copy it from the example. This page did not download those files.
The example does not ship a separate Gradio app. ComfyUI is the UI on this route. The FLUX.1-dev guide says an SSH tunnel to port 8188 keeps the page private, and that the template’s public URL adds no authentication. This page did not re-check that port. How to use SSH port forwarding on a cloud GPU is the tunnel page.
When it fails
CUDA out of memory. The SD 3.5 example publishes no GPU GB cutoff. The 2024-10-22 blog says a crash can mean insufficient RAM, which is the sentence next to its fp8 scaled note, and it does not give a number. Then use how to fix CUDA out of memory on a cloud GPU.
“CUDA error: no kernel image is available” on an RTX 5090. Catalog RTX 5090 and RTX 5090D are 32. That is the catalog value, not an SD 3.5 requirement. The RTX 5090 guide says that card needs a PyTorch build for CUDA 12.8 or newer. This page did not re-test the ComfyUI template on an RTX 5090.
Gated download. The three checkpoint repositories reported gated: auto. Accept the license on the model page the example links, create a read token, and download again. The Comfy-Org repack reported gated: false. The license section above still applies to those weights. The Comfy-Org card’s license_link points at the large repository’s LICENSE.md.
Save the images, then destroy
ComfyUI writes images under the output directory inside the ComfyUI folder. Copy that directory, and any workflow you exported, before you destroy the instance. 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. Billing is per minute, with a 1-minute minimum, and it stops when you destroy the instance. The billing rule is on the pricing page. Prices are not copied here.
Frequently asked questions
How much VRAM does the ComfyUI SD 3.5 example publish?
The official ComfyUI SD 3.5 example publishes no GPU GB figure (read 2026-11-12). It says the family contains a large 8B model and a medium 2.5B model. Those are parameter counts in that README, not VRAM. It says to use t5xxl_fp16.safetensors if you have more than 32GB ram. That sentence is system RAM. The ComfyUI blog post dated 2024-10-22 says the fp8 scaled workflow is a low vram option and gives no GB figure. This page does not mark any catalog vram_gb as enough or not enough. Which listing you can start is on https://cloudgpu.app/deploy.
What license covers the SD 3.5 weights the example links?
The model API for the three checkpoint repositories the example links reported license_name stabilityai-ai-community and license_link LICENSE.md, with gated auto (read 2026-11-12). The LICENSE.md fetched from the large checkpoint repository is titled STABILITY AI COMMUNITY LICENSE AGREEMENT, last updated July 5, 2024. Its introduction says it allows research, non-commercial, and limited commercial uses, and it names an annual-revenue condition of less than US $1,000,000. Section III says a commercial use must be registered at the URL in that section, and that the license terminates above USD $1,000,000 in annual revenue. Section IV©(iii) says you own outputs to the extent permitted by applicable law. Section IV(b) says you will not use the materials or their outputs to create or improve any foundational generative AI model, excluding the Models or Derivative Works. Read the file. This page does not restate the rest of it.
Where does the official ComfyUI example put the SD 3.5 files?
The example says to put sd3.5_large.safetensors or sd3.5_medium.safetensors in ComfyUI/models/checkpoints/, and to load text encoders separately: clip_l.safetensors, clip_g.safetensors, and a t5xxl file, in ComfyUI/models/text_encoders/. It says sd3.5_large_turbo.safetensors also goes in ComfyUI/models/checkpoints/, with steps 4 and cfg 1.2. It says sd3.5_large_fp8_scaled.safetensors can be used in the default workflow like any other checkpoint, in the same checkpoints directory. The checkpoint links for the large, medium, and turbo files are on that page. This page does not paste those Hub ids. Copy them from the example. Drag the example image on that page into ComfyUI to load the workflow.
Do the images stay on the instance after I destroy it?
No. Copy the ComfyUI output folder 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)
- ComfyUI SD 3.5 example (large 8B, medium 2.5B, text-encoder filenames, 32GB ram sentence, checkpoint filenames and folders, turbo steps 4 and cfg 1.2, fp8 all-in-one files, controlnet filenames, drag-in workflow): https://comfyanonymous.github.io/ComfyUI_examples/sd3/ and https://github.com/comfyanonymous/ComfyUI_examples/blob/master/sd3/README.md
- Comfy-Org repack README and model API for
Comfy-Org/stable-diffusion-3.5-fp8(returned idComfy-Org/stable-diffusion-3.5,gated: false,license_namestabilityai-ai-community). A GET of the fp8 URL ended at https://huggingface.co/Comfy-Org/stable-diffusion-3.5. File sizes are in the table above. - Model API for the three checkpoint repositories linked from the example (
gated: auto,license_namestabilityai-ai-community,safetensors.total8146280768, 2469663936, and 8146280768). File listing sizes:sd3.5_large.safetensors16460379262 bytes,sd3.5_medium.safetensors5107104286 bytes,sd3.5_large_turbo.safetensors16460374454 bytes. The Hub ids are not copied here. - LICENSE.md from the large checkpoint repository (STABILITY AI COMMUNITY LICENSE AGREEMENT, last updated July 5, 2024): the resolve URL linked from that repository’s
license_link. Fetched 2026-11-12. - ComfyUI blog, 2024-10-22 (8 billion, 2.6 billion, 4 steps, 1 megapixel, 0.25 to 2 megapixels, fp8 scaled called a low vram option,
models/clip): https://blog.comfy.org/sd3-5-comfyui/ - Public GPU memory: https://cloudgpu.app/gpu-prices.json (
updated_at2026-11-12T18:08:32.252Z) and https://cloudgpu.app/v1/public/gpu-catalog (generated_at2026-11-12T18:08:33.832Z).GET /v1/public/compute-capabilitiesreturned feature flags and no GPU list. The catalog template named ComfyUI has idflux-1-dev.
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).