HidreamFull (fal)

HiDream-I1 is a genuinely strong open model with one nasty catch: it's a 17B diffusion transformer driven by two text encoders (T5-XXL and Llama-3.1-8B-Instruct), and that's exactly why it's too heavy to live comfortably on a consumer card. The whole month after its April 2025 launch, the community spent quantizing it down to fit 16GB. Hidreamfull sidesteps all of that - it runs HiDream-I1 Full, the top-quality variant, on fal's servers and hands you back a native ComfyUI IMAGE. No 24GB card, no offloading gymnastics.

It's from gokayfem's ComfyUI-fal-API pack, so it's an API wrapper: nothing downloads, you pay per image.

How it works

I1's strength is prompt adherence - it follows instructions better than most of its generation, and that came directly from those two big text encoders. The "Full" variant is the highest-quality, slowest one (there are lighter Dev/Fast versions of HiDream elsewhere, but this node is Full). Running it via API means the heavy model sits on fal's hardware and you're just sending a prompt and settings. One thing worth getting right up front: this is HiDream-I1, not the newer HiDream-O1. O1 is a completely different 8B pixel-space architecture from 2026 - nothing about one transfers to the other beyond the company name.

The inputs and outputs that matter

• prompt - lean into it. I1 rewards detailed, structured instructions; that's its whole reason to exist.
• image_size - presets from square_hd and square through portrait_16_9 and landscape_16_9, plus custom.
• width / height - these only matter when image_size is set to custom (512–1440, in steps of 32). On a preset they're ignored.

The rest are safe on defaults: num_inference_steps (28), guidance_scale (3.5), num_images (1–10), safety_tolerance (1 strict → 6 permissive), and an optional seed (-1 = random). Output is a single IMAGE.

How to install it

Via ComfyUI Manager: search ComfyUI-fal-API, install, restart. Or manually:

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

No checkpoints to fetch. Get a fal API key from fal.ai, drop it into config.ini under [API] (copy config.ini.example first) or export FAL_KEY=your_key, restart, and the node appears under FAL.

Common issues & troubleshooting

• Auth error on first run. FAL_KEY isn't set. Fix and restart.
• You set width/height and nothing changed. They're only read when image_size is custom. Switch to custom or use a preset.
• A generation errors out. The pack raises fal's real message - validation, quota, or content policy. Read it; it usually points at the exact fix.
• Cost creep. It's a paid API call per image, and Full is the slow, expensive variant. The pack's result cache makes repeat identical calls free (force_rerun bypasses), and the spend guard caps a session - both worth turning on.
• Windows Portable. If fal-client is missing: python_embeded\python.exe -m pip install fal-client.

Fair warning: HiDream-I1 faded from the local conversation through 2025 as lighter models caught up, and it never grew much of a LoRA ecosystem. But if you specifically want its instruction-following at Full quality and don't want to wrestle 17B onto your GPU, the API route is the sane one.