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

Skip the "dev" in the name if you want; the interesting part is editImage. This is Civitai's cloud node for HiDream-O1-Image-dev used as an editor: you hand it a picture and a sentence of what to change, and the change happens. No mask, no inpaint pass, no ControlNet. Same story as the createImage sibling - the job runs on Civitai's fleet and comes back as an IMAGE tensor, billed in Buzz - but the mechanism is why you're here.

HiDream-O1 is a pixel-space model: no VAE, no text encoder, the diffusion happens directly in pixels. That sounds like trivia until you edit recursively. A latent model round-trips your image through a compressed space every pass, and compression loss compounds with each edit. HiDream never leaves pixel space, so a chain of "change the coat, now change the background, now change the light" edits doesn't quietly degrade. That's the property that made people who built their own HiDream nodes call it "a very good editor and a mediocre generator" - and this node is the editor half, with the download problem removed.

The inputs that matter

The graph shape is simple. Wire a LoadImage (or any IMAGE output) into the images input, type your change into prompt - the only required field - and run. The other ones worth touching:

• steps / cfg_scale - dev defaults here are 28 steps at cfg 1.0. Low guidance is the distilled-style default; the full (non-dev) edit node in the same menu defaults to 40 steps at cfg 5, which tells you the dev variant is tuned to be cheaper and faster, not better.
• width / height - default 2048×2048, and HiDream really does want the resolution. If you're editing small crops you can drop it, but expect the model to feel cramped below ~1024.
• negative_prompt - often skippable; include it when the edit is drifting into artifacts.
• checkpoint_model and loras - CIVITAI_AIR / CIVITAI_LORAS sockets, filled from the Civitai Model/LoRA Selector nodes. Empty = default checkpoint.
• seed, quantity (1–4), output_format (jpeg/png/webP), and image_metadata (free-form metadata stored with the output).

Outputs are the pack standard: images (IMAGE), errors, workflow_id, raw_json. The node also displays the workflow id and per-job cost inline - worth glancing at, because editing is where cloud HiDream bills are made.

Install and auth

Shared with every node in civitai-comfy-nodes. Install via ComfyUI Manager (search Civitai Comfy Nodes) or:

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

Then give it a Civitai account with Buzz. Either set the token env var, wire a Civitai Auth node, or sign in from the Civitai sidebar panel:

export CIVITAI_API_TOKEN=your_key

Where people get burned

The typical complaint is that an edit "changes everything, not just the thing I asked." That's the honest structural limit of instruction editing - the model re-emits the whole frame, so pixels you didn't mention can drift. HiDream compounds this less than latent editors across multiple passes, but a single-pass drift is still a thing; if the model keeps rearranging your subject, add the negative prompt or crop tighter. And remember this is the dev model - it's the budget door into HiDream editing. If quality is the goal and you can eat the Buzz, the non-dev HiDream-O1-Image editImage node is the one that gets the full 8B treatment.