Run FLUX image models on a cloud GPU

The FLUX.1 [dev] pipeline fits any 48 GB card: Hugging Face measured a 36.166 GB peak for it in BF16 on an H100. FLUX.2 [dev] is a different size class. Hugging Face says inference takes more than 80 GB of VRAM without offloading, about 62 GB with CPU offload and about 20 GB of free VRAM in 4-bit. The license splits the same way: FLUX.2 [dev] and klein 9B are non-commercial, and klein 4B is Apache-2.0.

FLUX memory by model: 16, 34 and 113 GB

The sizes below are the diffusers folders on Hugging Face, summed from the transformer, text encoder and VAE files and checked 1 October 2026. The VRAM statements are quoted from Black Forest Labs' model cards and Hugging Face's FLUX.2 post. The left-over figure is 92% of the card's memory minus the file total, calculated, so it leaves out activations and says nothing about offloading.

On DeployGPU the smallest card is 48 GB, and klein 4B, klein 9B and FLUX.1 all fit on it: about 28 GB is left after klein 4B's files and about 10 GB after FLUX.1's. FLUX.2 [dev] is the one that decides the GPU.

FLUX.2 [dev]: three ways to fit

Hugging Face's post gives the memory for each diffusers setup, and all three are its own statements.

The post adds that a remote text encoder with a 4-bit transformer runs in about 18 GB, and that group offloading works with as little as 8 GB of VRAM if you have 32 GB of free RAM. It also warns that each extra reference image, up to 10, needs more VRAM. CPU offload moves idle components into system RAM, as the 32 GB figure for group offloading shows, so check the offer's RAM as well as its GPU. My pick for a first FLUX.2 [dev] session is an RTX PRO 6000 with CPU offload, because it leaves room for reference images. If the files do not fit, the post's 4-bit route on a 48 GB card is the cheaper way in.

What ComfyUI downloads for FLUX.2

ComfyUI's FLUX.2 Dev tutorial says it uses quantized weights from the Comfy-Org/flux2-dev repository, and its workflow ships in the workflow templates. The files are flux2_dev_fp8mixed.safetensors at 35.46 GB, the Mistral text encoder mistral_3_small_flux2_bf16.safetensors at 35.58 GB and flux2-vae.safetensors at 0.34 GB. The repository also has FP8 and FP4 text encoders at 18.03 GB and 12.28 GB. With the BF16 encoder, the three files total 71.38 GB, so an RTX PRO 6000 can hold all of them with about 17 GB left (calculated), and an 80 GB card holds them with almost nothing spare. For klein 4B, the Comfy-Org/vae-text-encorder-for-flux-klein-4b repository (spelled that way on Hugging Face) has the 7.75 GB model and an 8.04 GB text encoder.

The ComfyUI template installs no models and keeps everything on the instance disk. Models go under /workspace/runpod-slim/ComfyUI/models/ inside the container, in diffusion_models, text_encoders and vae folders, and sudo docker exec deploygpu-workload ls /workspace/runpod-slim/ComfyUI/models confirms the layout before you download. The ComfyUI models guide covers downloading and the ComfyUI template page covers the app link, which cannot be used from code. Stopping deletes the disk, so each session fetches 54 to 71 GB of FLUX.2 files again, depending on the text encoder, and the stop guide covers what to copy off first.

FLUX.1 [dev] and Kontext on one 48 GB card

FLUX.1 [dev] is a 12B model, and Hugging Face's quantization post lists its BF16 pipeline as about 31.447 GB: T5 at 9.52 GB, CLIP at 246 MB, the transformer at 23.8 GB and the VAE at 168 MB. On an H100 the post measured a 36.166 GB peak in BF16, 17.281 GB in 4-bit and 24.432 GB in 8-bit. So a 48 GB card runs the full BF16 pipeline with about 8 GB beyond that peak, calculated. FLUX.1 Kontext [dev] is the same 12B size and does image editing from an instruction. Both are in ComfyUI's tutorials.

The three FLUX licenses

FLUX.2 [dev] and FLUX.1 [dev] weights come under non-commercial licenses, and both cards say generated outputs can be used for personal, scientific and commercial purposes as that license describes. The FLUX.2 [dev] card also requires inference filters or manual review under its terms. FLUX.2 [klein] 9B is non-commercial, and its card says the same about filters. Only klein 4B is Apache-2.0, and its card says the open weights are available for commercial use. Read the license text before you build a paid service on these weights. For training your own style, the LoRA guide covers the workflow, and the Wan 2.2 page covers video on the same template.