hidream-o1 / HiDream-O1-Image-dev / createImage

This is Civitai's cloud node for HiDream-O1-Image-dev, part of the civitai-comfy-nodes pack. The name is the one honest thing about it: it doesn't run anything on your machine. You type a prompt, hit run, and the job goes to Civitai's own orchestration fleet, billed in Buzz, with the finished image dropped back into your graph as a normal IMAGE tensor. No GPU needed, no 20GB checkpoint download, no VAE to source.

That last part is worth pausing on, because HiDream-O1 is architecturally weird in a good way. It's an 8B pixel-space model - there's no VAE and no separate text encoder anywhere in the stack. The diffusion happens directly in pixels. That's why a single transformer handles text-to-image, editing, and long-text rendering, and it's also why the dev checkpoint here refuses to be shy about resolution: width and height default to 2048×2048 and that's not a suggestion, it's the model's comfort zone. At that size, on a cloud fleet, "generate" costs minutes, not seconds. HiDream is famously a better editor than a generator, and the people who built their own nodes for it said so within days of release. Use this node when you specifically want HiDream's look - deep photoreal long-text rendering, no latent round-trip - not when you want a quick draft.

The inputs that actually matter

Only prompt is required. The rest of the useful ones:

• negative_prompt - works as you'd expect, and HiDream is one of the newer models where you can often leave it blank.
• steps and cfg_scale - dev defaults are 28 steps at cfg 1.0. That's the dev-model tell: low-guidance distilled-family behavior. Crank cfg up toward 5–7 and you'll get oversaturated, doubled-up garbage, not better fidelity. Leave it alone unless you know why you're touching it.
• width / height - default 2048, max 2048. If you're on a budget, knocking this down to 1024 changes the Buzz math dramatically, but the model was trained for the big sizes.
• seed and quantity (1–4 per call) - quantity gets you a batch in one billed job.
• checkpoint_model and loras - typed sockets. You fill them by wiring a Civitai Model Selector or Civitai LoRA Selector node, not by typing a string. Leave them empty and the recipe uses a default HiDream checkpoint.

The outputs are the pack standard: images (IMAGE - wire it into Save Image), errors, workflow_id, and raw_json. The last two are your cost-and-debug window; workflow_id and per-job cost also show on the node itself, which is a nice honest touch for something that bills per call.

Install and credentials

Install once and it covers all ~160 nodes in the pack. From ComfyUI Manager, search Civitai Comfy Nodes (publisher civitai) and install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt   # literally just `requests`

Then authenticate - this is where people get stuck. The node needs a Civitai account with Buzz and an API key:

export CIVITAI_API_TOKEN=your_key   # from https://civitai.com/user/account

Or use the Civitai Auth node (token / base URL overrides) or sign in from the Civitai sidebar's connect panel. Headless setups should use the env var - browser OAuth won't fire there.

Where people get burned

Blank-prompt runs fail fast with a local error, which is better than the orchestration 400 you'd otherwise eat. The real surprise is the bill: a 2048×2048 HiDream job is one of the pricier image recipes on the platform, and quantity: 4 is four of them. The pack is early preview - nodes and behavior "may change without notice" - so pin the version if you're building workflows on top of it. And remember the honest framing from the community: this is the right tool for a model you don't want to host, and the wrong default for one you can. HiDream-O1's weights are downloadable; this node just means you never have to.