Deploy image models on RunPod¶

Use this page to choose an image-generation or image-editing template, deploy it on a suitable NVIDIA GPU and open ComfyUI. The templates contain different model families, so choose one based on the task you want to perform.

Quick recommendation¶

Starters¶

Start with Z-Image Turbo and Base for straightforward text-to-image generation but a lower quality then Krea-2.

Hot¶

Choose Krea-2 Base and Turbo for identity, composition and broader image workflows, also perfect for text-to-image.
Choose Qwen-Image 2.1 for text-to-image generation and image editing, either with direct prompts or with optional prompt enhancement.

Editing¶

Choose FLUX.2 Klein for control or reference-image workflows.
Choose Qwen-Image-Edit 2511 when an existing image is your main input and you need a more mature eco system.

Experimental¶

Choose SenseNova U1.5 for text-to-image and image-to-image workflows.

Choose a GPU¶

Look for the starred GPUs in the RunPod selection screen: these are the recommended options for your template.
The template automatically selects the appropriate model variant based on your GPU and its available VRAM.

Steps for deployment¶

Choose a template.
Check its hardware and storage requirements.
Open the RunPod deployment link and configure the pod.
Wait until the container log reports that the pod is ready.
Open Connect → HTTP Services → ComfyUI (port 8188).

Choose a template¶

Template	Best suited to	RunPod
Krea-2 Base and Turbo	Text-to-image, identity transfer, composition and image-editing workflows	Deploy Krea-2 →
SenseNova U1.5	Text-to-image and image-to-image with the Final BF16 checkpoint and optional distilled 8-step LoRA	Deploy SenseNova U1.5 →
Qwen-Image 2.1 without prompt enhancer	Text-to-image and image editing with INT8 ConvRot models, an alternative Heretic text encoder and direct prompts	Deploy Qwen-Image 2.1 →
Qwen-Image 2.1 with prompt enhancer	Text-to-image and image editing with the same image models plus Qwen3.5 9B prompt enhancers	Deploy Qwen-Image 2.1 + prompt enhancer →
Z-Image Turbo and Base	General and fast text-to-image generation	Deploy Z-Image →
FLUX.2 Dev	High-quality generation and multiple-angle workflows	Deploy FLUX.2 Dev →
FLUX.2 Klein	Control, target and reference-image workflows	Deploy FLUX.2 Klein →
Qwen-Image 2512	Qwen text-to-image generation	Deploy Qwen-Image →
Qwen-Image-Edit 2511	Image editing and multiple viewing angles	Deploy Qwen-Image-Edit →

Each template includes workflows for its model family. After deployment, use the workflow browser in ComfyUI to select an included example.

Before you deploy¶

These are tested minimum configurations. Larger images, batches and additional custom nodes can require more VRAM and system RAM.

Model profile	Tested GPU	Minimum VRAM	Minimum system RAM
Z-Image Turbo	RTX 3090 or RTX A5000	20 GB	50 GB
Krea-2 fp8	RTX A4500	20 GB	65 GB
Krea-2 bf16	RTX 3090/4090	24 GB	65 GB
FLUX.2 Klein	RTX A4500, 3090, 4090	20 GB	50 GB
FLUX.2 Dev fp8	RTX 3090	24 GB	50 GB
FLUX.2 Dev bf16	RTX 6000 Ada	44 GB	80 GB
Qwen-Image fp8	RTX A5000 or RTX 3090	24 GB	45 GB
Qwen-Image bf16	RTX A6000	44 GB	65 GB
Qwen-Image-Edit fp8	RTX A5000 or RTX 3090	24 GB	45 GB
Qwen-Image-Edit bf16	L40S	48 GB	65 GB
Qwen Image 2.1	RTX 4090, PRO 6000 MiG 24GB/48GB, L40S	24 GB	65 GB
SenseNova U1.5	L40S	50 GB	80 GB

See image-model hardware requirements for the complete tested list.

Storage checklist¶

Model family	Minimum persistent volume at `/workspace`
Z-Image	50 GB
ERNIE-Image	50 GB
Krea-2 fp8 / bf16	50 GB / 60 GB
FLUX fp8 / bf16	75 GB / 90 GB
Qwen-Image	70 GB
SenseNova U1.5	80 GB

Keep at least 15 GB of pod storage in addition to the persistent volume. Do not reduce the template defaults unless you know the complete model footprint.

Optional credentials¶

Set PASSWORD if you want a fixed Code-Server password.
Add HF_TOKEN when a gated Hugging Face model or an additional private model requires it.
Add CIVITAI_TOKEN only when downloading resources from CivitAI.

See RunPod configuration for all supported environment variables and secret names.

FLUX.2 Klein 9B license¶

The FLUX.2 Klein 9B weights are released under the FLUX Non-Commercial License. Local use of the model is limited to the non-commercial purposes defined in that license; commercial use of the local weights requires a separate license from Black Forest Labs. Review the license for the separate conditions that apply to generated outputs.

Before starting a template that downloads the 9B model:

Accept access to the distilled and/or base model on Hugging Face.
Create a Hugging Face token with the required read access.
Add that token to the RunPod template as the HF_TOKEN secret.
Deploy the pod and confirm in the logs that the model download succeeds.

Follow the FLUX.2 Klein 9B access guide for the model links and manual download commands.

Deploy and connect¶

Open the deployment link for your chosen model.
Choose an available GPU that meets its minimum VRAM requirement.
Confirm the system RAM, pod storage and persistent volume.
Add any required RunPod secrets or environment variables.
Click Deploy Pod.
Open the pod logs and wait for the final ready message. Initial startup can take several minutes while the container and models download.
Expand the pod, select Connect, and open ComfyUI under HTTP Services.

For screenshots, normal startup checkpoints and troubleshooting, follow the full RunPod pod deployment guide.

Container date tags¶

RunPod templates use dated container tags. Keep the tag supplied by the template unless you are deliberately testing another build. Different model templates can remain on different validated tags when a newer container has not yet been tested with every model family.

Continue after deployment¶

Reference¶

Topic	Guide
GPU, RAM and storage requirements	Image-model hardware requirements
Environment variables and network services	RunPod configuration
Pod startup and troubleshooting	RunPod pod deployment
Model documentation and upstream resources	Image-model resources