How do I run FLUX.1-dev in ComfyUI on a cloud GPU?

Short answer: Deploy the ComfyUI template on a card with enough VRAM, then download FLUX.1-dev on the instance in one of two forms: the single FP8 checkpoint flux1-dev-fp8.safetensors (17.25 GB) in ComfyUI/models/checkpoints, or the full 16-bit set (flux1-dev.safetensors in models/diffusion_models, t5xxl_fp16 and clip_l in models/text_encoders, ae.safetensors in models/vae, about 34 GB together). The official ComfyUI FLUX example says to set CFG to 1.0 with the FP8 checkpoint and that “fp8 degrades the quality a bit” (ComfyUI examples, read 2026-11-03).

Step 1: Pick a card by VRAM

Setup Files (sizes on Hugging Face, read 2026-11-03) VRAM guidance
FP8 checkpoint flux1-dev-fp8.safetensors 17.25 GB About 12-16 GB for FLUX.1 weights at FP8 (CloudGPU model page, read 2026-11-03), plus activations
Full 16-bit flux1-dev.safetensors 23.8 GB, t5xxl_fp16 9.79 GB, clip_l 0.25 GB, ae 0.34 GB About 31.4 GB with the whole pipeline loaded in BF16 (Hugging Face Diffusers blog, read 2026-10-23)

VRAM need depends on model, precision and workflow: resolution, batch size, LoRAs and ControlNet add to it, and ComfyUI can offload parts of the pipeline at the cost of speed. For a card-by-card view, see RTX 4090 vs RTX 4090 48G vs RTX 5090 for image generation and how much VRAM FLUX and SDXL need.

Step 2: Deploy the ComfyUI template

On /deploy, choose the ComfyUI template. Once a template is chosen, the page shows only the GPUs that can run it; pick one of those. Read the template description before you deploy: CloudGPU’s /docs (read 2026-11-03) say the ComfyUI template “installs ComfyUI only” and does not include FLUX weights. If the template you choose says it already includes FLUX.1-dev, skip to Step 5.

When the instance is running, open a terminal in JupyterLab or connect over SSH, and find the ComfyUI folder:

COMFY=$(find / -maxdepth 4 -type d -name ComfyUI 2>/dev/null | head -n 1); echo "$COMFY"
nvidia-smi        # GPU name and VRAM
df -h "$COMFY"    # free space where the models will go

Step 3: Accept the licence (full 16-bit files only)

black-forest-labs/FLUX.1-dev is a gated repo on Hugging Face: open it in your browser, accept the terms with your account, then create a read token. The Comfy-Org FP8 repack is not gated (Hugging Face API, read 2026-11-03), but the FLUX.1-dev weights in it are still under Black Forest Labs’ own licence, so read it before you use the outputs. Token handling is covered in how to download a Hugging Face model on a cloud GPU.

Step 4: Download the files on the instance

Download on the instance itself, not through your laptop. Every CloudGPU instance has a cloudgpu-net command that turns on download acceleration for Hugging Face, GitHub and similar sites in the current shell; turn it off again before pip installs, which can be slower while it is on (more in how to speed up model downloads).

cloudgpu-net on
pip install -U "huggingface_hub"

Option A: FP8 single file (loads with the regular Load Checkpoint node):

hf download Comfy-Org/Flux-1-Dev flux1-dev-fp8.safetensors \
  --local-dir "$COMFY/models/checkpoints"

Option B: full 16-bit set (needs the licence from Step 3):

read -rs HF_TOKEN && export HF_TOKEN      # paste your read token; nothing is echoed
hf download black-forest-labs/FLUX.1-dev flux1-dev.safetensors --local-dir "$COMFY/models/diffusion_models"
hf download black-forest-labs/FLUX.1-dev ae.safetensors        --local-dir "$COMFY/models/vae"
hf download comfyanonymous/flux_text_encoders t5xxl_fp16.safetensors clip_l.safetensors \
  --local-dir "$COMFY/models/text_encoders"

To save memory and space, the ComfyUI example offers t5xxl_fp8_e4m3fn_scaled.safetensors (5.16 GB) instead of t5xxl_fp16; it recommends the fp16 encoder only “if you have more than 32GB ram” (system RAM). When the downloads finish:

cloudgpu-net off
ls -lh "$COMFY"/models/{checkpoints,diffusion_models,text_encoders,vae}

Step 5: Build or load the workflow

Refresh ComfyUI in the browser so it sees the new files. The ComfyUI FLUX examples page publishes example images that carry their workflow; drag one into ComfyUI to load it.

  • FP8 checkpoint: Load Checkpoint → flux1-dev-fp8.safetensors. Set CFG to 1.0, as the example says.
  • Full set: Load Diffusion Model → flux1-dev.safetensors; DualCLIPLoader → t5xxl_fp16 and clip_l, type flux; Load VAE → ae.safetensors. Per the example page, setting weight_dtype to fp8 in the Load Diffusion Model node “will lower the memory usage by half”, with a small quality cost. If you run out of memory, the page suggests the FP8 checkpoint instead.

To open the ComfyUI page, use the public HTTPS URL the ComfyUI template shows on the instance row. That URL adds no authentication, so anyone who has it can use your ComfyUI. To keep it private, use an SSH tunnel to port 8188 instead (SSH port forwarding guide).

Run one image and watch peak memory in a second terminal:

watch -n 1 nvidia-smi --query-gpu=memory.used,memory.total --format=csv

Step 6: Save outputs, then destroy

Images are written to $COMFY/output. Data on the instance is deleted when you destroy it, so copy outputs and exported workflows off first, for example rsync -avP -e "ssh -p <port>" <user>@<host>:<ComfyUI path>/output ./flux-output. Billing is per minute, rounded up, with a 1-minute minimum, and billing stops when you destroy the instance; the instance is billed while the models download. See how per-minute billing works.

Frequently asked questions

Which FLUX.1-dev files does ComfyUI need?

Either the single FP8 checkpoint flux1-dev-fp8.safetensors in models/checkpoints, or the full set: flux1-dev.safetensors in models/diffusion_models, t5xxl_fp16.safetensors and clip_l.safetensors in models/text_encoders, and ae.safetensors in models/vae (ComfyUI examples, read 2026-11-03).

FP8 or full 16-bit FLUX.1-dev?

The ComfyUI example says FP8 “degrades the quality a bit” and recommends the official 16-bit version if you have the resources. FP8 needs about 12-16 GB of VRAM for the weights; the full 16-bit pipeline is about 31.4 GB loaded in BF16 (Hugging Face Diffusers blog, read 2026-10-23).

Do I need a Hugging Face token for FLUX.1-dev?

For black-forest-labs/FLUX.1-dev, yes: accept the licence on the model page and use a read token. The Comfy-Org FP8 repack was not gated when read on 2026-11-03, but the FLUX.1-dev licence still applies to the weights.

What CFG should I use with FLUX.1-dev in ComfyUI?

The official ComfyUI example says to set CFG to 1.0 when using the FP8 checkpoint.

Do the FLUX files stay on the instance after I destroy it?

No. Data on the instance is deleted when it is destroyed, so you download the files again on the next instance. Copy your images off before you destroy it.

Sources (read 2026-11-03, Asia/Shanghai)

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

Last updated: 2026-11-03 (Asia/Shanghai).