HiDream-I1 Fast

HiDream-I1 is the 17-billion-parameter MIT-licensed model that made "the prompt actually works" its entire personality - best-in-class prompt adherence for its era, especially with text-in-image, which is where most models embarrass themselves. It was also, at 17B, too heavy for most local setups. This node (Runware_hidream_i1_fast, AIR runware:97@3) is the Fast rung of that family in Runware's cloud: same family, faster iteration, no VRAM bill.

Note the "Fast," not "Pro." HiDream-I1 ships in tiers - Full and Dev were the careful renderers, Fast is the iteration tier that trades a little ceiling for speed. In the cloud that trade matters double: Fast is cheaper per run and returns quicker, which makes it the right node when you're testing prompt phrasing or checking text-rendering before committing to a slower, pricier pass.

How it works

imageInference with the runware:97@3 AIR. The node sends your prompt, size, and sampler settings up and returns an IMAGE tensor. The model is a sparse diffusion transformer with dual text encoders - which is a fancy way of saying it was built to read prompts carefully, and it shows in long, detailed, multi-clause prompts that make other models drop details. Text in the image - signage, posters, labels - is its party trick. steps defaults to 16, and CFGScale is gated off by default (the model has its own guidance strategy; enable it only if you have a reason).

The inputs that matter

• positivePrompt (required) - this is the node where the prompt is the product. Write the whole sentence; HiDream rewards specificity more than most.
• width / height - optional here (defaults 1024×1024), 128–2048, step 16.
• steps - 16 default, no documented max. Fast runs comfortably in the teens.
• negativePrompt - supported, though HiDream-family models lean on it less than SDXL does.
• lora - the RUNWARE_LORA socket, if you have a style adapter for the family.
• numberResults - up to 20 variations per run, useful given how fast Fast is.

Installing it

Install the pack. ComfyUI Manager → search Runware → install → restart. Manual:

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

Key from runware.ai/api-keys → ComfyUI Settings → Runware API key (or RUNWARE_API_KEY). HiDream-I1 is hosted; no weights download.

Where people get burned

• Fast is the iteration tier. If you're doing final output, don't expect Fast to match the Full/Dev tiers' finish - that's what the higher rungs are for, and they cost more.
• It's not the same model as HiDream-O1. The KB is emphatic about this: I1 (this family) and O1 (the later pixel-space editor) are different generations. If you came here for O1's editing, you're at the wrong node - and the pack doesn't ship O1 here.
• Long prompts are the point, but prompt engineering still applies. HiDream follows better, not mindlessly; the standard clarity rules still help.
• Cloud cost is per run. Fast is one of the cheaper nodes in the pack, but 20-result batches add up - the title bar shows the total.

HiDream-I1 Fast is the node for "I need this prompt to actually land, and I need to try it twelve times." When text-in-image or dense detail matters, it's the pick.