HiDream-I1 Full

HiDream-I1 is the model that topped Artificial Analysis in April 2025 - the first Chinese-developed generative visual model to lead that ranking - and it shipped under MIT, the most permissive licence in the game. This node runs the Full variant on Runware's cloud. "Full" is the complete-precision checkpoint, the one that needs a serious GPU to run locally; here that requirement lives on Runware's side and you just get the results.

The honest framing, because the history is weird: HiDream is a lab that gives away better terms than almost anyone and gets less for it than anyone. I1 was genuinely celebrated, then quietly fell out of the conversation as the community's attention moved to Z-Image and Flux 2 Klein. None of that is the model's fault - I1 still renders with a distinctive, detailed, high-native-resolution look that some people never stopped preferring. And via Runware you can finally meet it without committing 40GB of disk and a weekend of setup.

The inputs that matter

• positivePrompt (required, STRING, multiline): the scene. HiDream reads natural language well - full sentences beat tag soup.
• width / height (128–2048, step 16): up to 2048 native, which is part of I1's whole pitch - it was a native-resolution model before that was fashionable. The 1024 default is fine; the headroom is there when you want it.
• steps (default 30): comfortable default; I1 doesn't need the Flux-style 50-step marathon.
• CFGScale is gated behind a toggle - off by default, meaning the model's own guidance behavior applies. Flip it on if you're chasing specific prompt adherence.
• scheduler (76 options), negativePrompt, seed, numberResults (up to 20): the standard diffusion block.
• Feature sockets: lora, watermark, acceleratorOptions - the usual builders. (I1's LoRA ecosystem is thin, so this node is more about the base model than its add-ons.)

Output is a native image (IMAGE).

How it works

Pack-standard: your prompt and settings ride up as a REST request, HiDream-I1-Full runs on Runware's GPUs, and the image comes back as an IMAGE tensor. "Full" matters here because the cloud runs the real thing - no fp8 or distilled shortcut unless you want it. Cost per run lands on the title bar.

Install and gotchas

Install the pack once - ComfyUI Manager, search Runware - or git clone https://github.com/Runware/ComfyUI-Runware into custom_nodes + pip install -r requirements.txt, restart, and set a Runware API key (Settings → Runware API key, RUNWARE_API_KEY, or runware auth login).

The gotchas are mild. I1's aesthetic leans detailed and "clean," which reads to some as overcooked - if that's not your taste, no setting here will fix it, it's the model. And since it's a paid API, the 2048×2048 max with numberResults: 20 is a spending decision, not a default. But as a way to finally try a leaderboard-topping MIT model without owning its hardware, this node is the smoothest introduction the pack offers.