hidream-o1 / HiDream-O1-Image / editImage

Of the three HiDream-O1 nodes in the Civitai pack, this is the one with no "dev" suffix, and the difference is the whole point. HiDream-O1-Image / editImage is the full checkpoint running instruction-based editing on Civitai's cloud: an image in, a sentence describing the change, a new image out. The dev edit node defaults to 28 steps at cfg 1.0; this one ships at 40 steps, cfg 5.0. That's a model telling you it intends to take its time and actually look at what you gave it.

It's worth knowing what HiDream-O1 is before you spend Buzz on it. It's an 8B pixel-space model - no VAE, no text encoder, diffusion straight in pixels - and its genuinely famous property is that this makes recursive editing cheap: because nothing ever round-trips through a compressed latent space, a chain of sequential edits doesn't compound compression loss the way latent editors do. Community builders who ran it locally called it a very good editor and a mediocre generator. This node is that editor, hosted. If your workflow is "edit, then edit the edit, then edit that," this is the Civitai node for you - and it's the one where the pricier non-dev variant earns its keep.

The inputs that matter

Wire an image in, describe the change, run.

• images - IMAGE input; feed it from LoadImage or anything downstream in your graph. Required, along with prompt (the instruction, e.g. "make it night, add neon reflections").
• steps (40) / cfg_scale (5) - the full-model defaults. If you came from the dev node, don't leave these alone and expect the same behavior; this variant wants real guidance and a real step count.
• width / height - 2048×2048 default, matching the model's native comfort zone. You can go smaller to cut cost, but HiDream feels cramped below ~1024.
• negative_prompt - handy when the edit drifts into artifacts; otherwise skippable.
• seed / quantity (1–4) / output_format / image_metadata - standard.
• checkpoint_model / loras - CIVITAI_AIR and CIVITAI_LORAS sockets, wired from the Civitai Model/LoRA Selector nodes. Empty means the default checkpoint.

Outputs: images (IMAGE), errors, workflow_id, raw_json. The node shows the workflow id and per-job cost inline - this is the priciest HiDream node in the set, so glance at it once and internalize the number.

Install and auth

Same pack-wide setup as the other HiDream nodes. Install Civitai Comfy Nodes from ComfyUI Manager, or:

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

You need a Civitai account with Buzz: set CIVITAI_API_TOKEN, wire a Civitai Auth node, or sign in via the Civitai sidebar's connect panel.

The honest tradeoff

The full model at 40 steps at 2048×2048 is slow and costs more than the dev variant - that's the point, and it's also the trap. If you only need a quick edit, the dev node will do 80% of the job for a fraction of the price. Reach for this one when output quality is the requirement and the recursive-edit advantage matters. And keep the pack's early-preview warning in mind: Civitai ships these nodes generated from their API spec and warns behavior "may change without notice." Don't build a production pipeline on it without pinning the version.