HiDream-I1

HiDream-I1 is a 17B model that makes crisp images from text fast. It’s open-source under MIT license and free for all uses, also completely uncensored and capable of NSFW image generation.

Visit This Site

Overview

HiDream-I1 is a free AI that turns your words into pictures. It’s built by the folks at vivago.ai and it’s packing 17 billion parameters. That’s a lot of brainpower. It uses this thing called Mixture of Experts setup with DiT blocks to work faster and smarter.

It’s loaded with four strong text encoders like CLIP and Llama 3 so it really gets what you’re saying. That means better images that actually match your prompt. You can check it out on GitHub or try it live on Hugging Face.

There's a ComfyUI Wrapper available already, but no Comfy native version yet.

It is pretty dang good. It beat models like DALL·E 3 and SDXL on some test scores like DPG-Bench and GenEval. These are tests that check how close the images match the prompt and how clean they look. spaces/HiDream-ai/HiDream-I1-Dev

HiDream-I1 Hardware Requirements for HiDream-I1

-> Requires Nvidia GPU.

4Bit Quantized Model (HiDream-I1-Fast-nf4):

GPU Architecture: NVIDIA >= Ampere (e.g. A100, H100, A40, RTX 3090, RTX 4090)
GPU RAM: >= 16 GB
CPU RAM: >= 16 GB

Bit of a side note. text_encoder_4 runs on Meta’s Llama 3.1-8B Instruct which isn't really fully open-source the way people usually expect.

That model follows Meta’s special license so you can’t just use it for anything you want. You gotta agree to their rules and you’re not allowed to share it freely.

So even though HiDream says it’s MIT licensed adding stuff based on Llama 3.1 would likely go against Meta’s rules.

That’s why they can’t just include text_encoder_4 in the repo. You gotta grab it yourself from Meta’s HuggingFace page.

Links

Text-to-Image

Educators and Trainers Creative Professionals Content Creators Media and Film Makers Marketing and Branding Specialists Developers and Tech Creators Nonprofit and Advocacy Creators Small Business Owners Entertainment and Performance Artists Professional Content Creators

This tool is free to use and is offered under MIT License.

HiDream dropped with a loud claim —it’s being called the best open-source image model out there. But while some folks are impressed, others are hitting the brakes.

At first glance it looks solid. Prompt following is tight and tech heads like that it’s built with a Mixture of Experts setup which means it only fires up part of the model at a time. That should make it faster and lighter... in theory. If you quantize it hard enough it’ll even run on mid-range GPUs and work great for local use.

But right now the results aren’t blowing people away. Images have that overly-sharpened “AI made this” feel. Stuff like art style accuracy and subtle detail? Still lacking. It kind of struggles with creativity and anything complex or painterly. One guy called it “roughly at Flux Schnell quality” which ain’t exactly high praise.

Even though it handles some prompts okay the overall look still leans toward that generic overly-optimized vibe. It nails the words but flattens the soul of the picture. If you’re into technical upgrades it’s interesting but if you’re chasing visual magic it’s not quite there yet. Maybe with some fine-tunes and extra work it could turn into something better but for now it’s more promise than payoff.

[ Reddit ]

Prompt:

Medium shot, fish-eye lens. Shallow depth of field creates a focus on the robots taking selfies while at the top of the mountain. All robots huddled together while taking a group selfie picture. Goldy is a red 1950s retro robot monster, slightly rusty. Dolbus is a sleek futuristic humanoid robot with rounded features, black and steel look and taller than the rest. Bingus is a copper steampunk robot with big eyes and a stylish hat posing with crossed arms, chin up, projecting attitude. All robots appear happy, smiling directly at the camera. The mood is jolly and humorous.