How to Run HiDream in ComfyUI on Runpod

This guide covers running HiDream-I1 in ComfyUI on Runpod. HiDream-I1 is a 17-billion-parameter text-to-image model released under the MIT license, which puts it in a small group: an open-weights image model of this size with no commercial restrictions attached.

By the end you'll have ComfyUI running on a pod with a HiDream workflow loaded, the four text encoders wired correctly, and the right sampler settings for whichever variant you picked. That last part matters more with HiDream than with most models, because the three variants need different settings and the wrong ones produce visibly worse output.

What is HiDream-I1?

HiDream-I1 is a text-to-image model open-sourced by HiDream-ai on April 7, 2025. It's released under the MIT license, so personal projects, research, and commercial use are all permitted.

Two architectural choices define it. The first is a hybrid design: a Diffusion Transformer backbone where dual-stream MMDiT modules handle multimodal information and single-stream DiT modules work on global consistency, with a dynamic routing mechanism allocating compute across them. The second is the text encoding stack, and it's unusual. HiDream runs four text encoders in parallel:

• OpenCLIP ViT-bigG and OpenAI CLIP ViT-L for visual semantic alignment
• T5-XXL for long-text parsing
• Llama-3.1-8B-Instruct for instruction understanding

That combination is why HiDream handles compositional prompts, ones specifying colors, counts, and spatial relationships together, better than models with a single encoder. It also handles Chinese prompts well relative to other open models.

HiDream-ai released three variants, distinguished by step count:

What is ComfyUI?

ComfyUI is an open-source, node-based interface for generative AI pipelines. You build workflows by connecting nodes, each one a model loader, sampler, conditioning step, or output, rather than writing inference code.

ComfyUI has native HiDream support, including the QuadrupleCLIPLoader node that HiDream's four-encoder stack requires. Without native support you'd be wiring that yourself, so make sure your ComfyUI build is current before you start.

Why run HiDream on Runpod?

HiDream's VRAM floor is high. The fp8 checkpoints want more than 16 GB, and full precision wants more than 27 GB, before you account for four text encoders loaded alongside the diffusion model. That rules out most consumer hardware for the full-precision path.

On Runpod you pick a card that fits the variant you want, deploy a pre-built ComfyUI template, and pay per second while the pod runs. Testing the Full variant at full precision is a GPU selection, not a hardware purchase.

GPU requirements for HiDream

Numbers below come from the model's published requirements. Add headroom for the text encoders and your resolution target:

• 16–24 GB VRAM (RTX 4090, RTX 3090, A5000): fp8 checkpoints for all three variants
• 40–48 GB VRAM (L40S, A6000, A100 40GB): Full-precision checkpoints, larger resolutions, batch generation
• 80 GB VRAM (A100 80GB, H100): Full precision with room for upscaler chains and multiple workflows queued

GGUF quantizations from city96 and an NF4 build exist if you need to go lower, but they need extra custom nodes, so factor that into your setup time.

Templates for running HiDream on Runpod

ComfyUI (official Runpod template) Runpod's first-party template. Current ComfyUI build on CUDA 12.8, with ComfyUI-Manager, KJNodes, Civicomfy, and ComfyUI-RunpodDirect pre-installed. ComfyUI copies to your workspace on first boot. Wait for [ComfyUI-Manager] All startup tasks have been completed. in the pod logs before opening the interface, then add the HiDream files.

For other options, browse console.runpod.io/hub and search for ComfyUI. The Hub is a live catalog, so check the image version and last-updated date on community templates before deploying.

Step 1: Create a Runpod account

Go to runpod.io and sign up with email, GitHub, or Google. Add credits or a payment method so you can deploy a GPU.

Step 2: Deploy a ComfyUI pod

1. Open the Pods section and select a GPU. An RTX 4090 covers the fp8 checkpoints; an L40S or A100 covers full precision.
2. Click Change Template and select the official ComfyUI template.
3. Set container disk to at least 60 GB. Four text encoders plus a 17B diffusion model adds up quickly, and Llama-3.1-8B alone is a substantial download.
4. Attach a network volume if you'll come back to this setup. The encoder files are reusable across other workflows, so they're worth keeping.
5. Click Deploy.

Step 3: Download the HiDream model files

HiDream needs six files, and four of them are the text encoders shared across all three variants.

Text encoders → ComfyUI/models/text_encoders/

• clip_l_hidream.safetensors
• clip_g_hidream.safetensors
• t5xxl_fp8_e4m3fn_scaled.safetensors
• llama_3.1_8b_instruct_fp8_scaled.safetensors

VAE → ComfyUI/models/vae/

• ae.safetensors (this is Flux's VAE; if you've run a Flux workflow before, you already have it)

Diffusion model → ComfyUI/models/diffusion_models/

Pick one based on your variant and VRAM:

• hidream_i1_full_fp8.safetensors or hidream_i1_full_fp16.safetensors
• hidream_i1_dev_fp8.safetensors or hidream_i1_dev_bf16.safetensors
• hidream_i1_fast_fp8.safetensors or hidream_i1_fast_bf16.safetensors

The fp8 builds want more than 16 GB of VRAM. The full-precision builds want more than 27 GB. The repackaged ComfyUI-ready files live in the Comfy-Org HiDream-I1 repository on Hugging Face.

Two of these files, t5xxl_fp8_e4m3fn_scaled and ae.safetensors, are shared with other workflows. Check what's already in your model directories before downloading.

Step 4: Open ComfyUI and load the workflow

Once the pod shows a running state, click Connect and open port 8188.

Load the HiDream workflow for your chosen variant from the Template Library, or from ComfyUI's HiDream-I1 documentation. Then verify three nodes:

1. Load Diffusion Model points at your chosen HiDream checkpoint
2. QuadrupleCLIPLoader has all four text encoders loaded, in the order listed above
3. Load VAE points at ae.safetensors

If QuadrupleCLIPLoader doesn't exist in your node list, your ComfyUI build predates native HiDream support. Update and restart.

Step 5: Set the sampler correctly for your variant

This is the step that catches people. The three variants need different ModelSamplingSD3 shift values and different sampler settings, and using the wrong ones degrades output in ways that look like a bad model rather than a bad configuration.

HiDream-I1-Full

• shift: 3.0
• steps: 50
• cfg: 5.0
• Optional: sampler lcm, scheduler normal

HiDream-I1-Dev

• shift: 6.0
• steps: 28
• cfg: 1.0 (this one matters)
• Optional: sampler lcm, scheduler normal

HiDream-I1-Fast

• shift: 3.0
• steps: 16
• cfg: 1.0 (this one matters)
• Optional: sampler lcm, scheduler normal

The Dev and Fast variants are distilled and don't use negative prompts, which is why CFG has to be 1.0. Leave it at a text-to-image default like 5.0 or 7.0 and output quality falls off noticeably.

Once the settings match your variant, enter a prompt and click Queue Prompt. Output appears in the interface and saves to /workspace/ComfyUI/output.

Pro tips

Use Fast to find the prompt, Full to render it. Sixteen steps is fast enough to explore composition and phrasing. Swap the diffusion model node to Full and raise steps to 50 once the prompt is right; everything else in the graph stays the same.

Lean on the compositional prompting. HiDream's four-encoder stack handles counts, colors, and spatial relationships better than single-encoder models. "Three red ceramic bowls on the left side of a pale wooden table, one blue bowl on the right" is the kind of prompt worth actually testing here.

Reuse the shared encoder files. t5xxl_fp8_e4m3fn_scaled and ae.safetensors are also used by Flux and several other workflows. Keeping them on a network volume saves a download every time you spin up a new setup.

Attach a network volume. Pod storage is ephemeral. A network volume keeps checkpoints, encoders, and saved workflows across sessions.

Stop your pod when you're done. Runpod bills by the second while a pod runs.

Wrapping up

HiDream-I1 is a strong open model that happens to be MIT licensed, which makes it a practical choice when the license terms on other open-weights image models get in the way. The setup takes a few more steps than most, mainly because of the four encoders, but it's a one-time cost per environment, and a network volume makes it a one-time cost overall.

Ready to try it? Deploy the ComfyUI template and start with the Fast variant.

FAQ

What GPU do I need to run HiDream on Runpod?

The fp8 checkpoints want more than 16 GB of VRAM, so an RTX 4090 or RTX 3090 works. Full-precision checkpoints want more than 27 GB, which puts you on an L40S, A6000, or A100. Add headroom for the four text encoders and your target resolution.

What's the difference between HiDream Full, Dev, and Fast?

Step count and speed. Full runs 50 steps at CFG 5.0 for the highest quality. Dev runs 28 steps at CFG 1.0, balancing quality and speed. Fast runs 16 steps at CFG 1.0 for rapid iteration. Dev and Fast are distilled and don't use negative prompts, which is why they need CFG at 1.0.

Is HiDream free to use commercially?

Yes. HiDream-I1 is released under the MIT license, which permits personal, research, and commercial use.

Why does my HiDream output look washed out or over-baked?

Almost always the CFG setting. If you're running the Dev or Fast variant, CFG must be 1.0. Check the shift value in ModelSamplingSD3 too: 3.0 for Full and Fast, 6.0 for Dev.

What about HiDream-E1?

HiDream-E1 is a separate instruction-based image editing model from the same team, built on HiDream-I1. This guide covers I1, the text-to-image model. E1 uses a different workflow.

Can I run HiDream as an API instead of a UI?

Yes. Containerize your ComfyUI workflow and deploy it to Runpod Serverless. Your application sends generation requests to the endpoint, which scales from zero and bills per second of compute rather than per hour of pod uptime.