Skip to content
CozyGPU

HunyuanImage 3.0

Tencent's large image model: text to image, edits from one picture, and edits that combine two.

Powers
Create Image, Edit Image
Modes
Text to image · Image edit · Multi-image edit
GPU class
96 GB Blackwell (falls back to 80 GB Ada Pro)
01

What it is

HunyuanImage 3.0 is Tencent's open-weight image model, here as its Instruct Distil release in 4-bit (NF4) form so it fits a single large GPU. The same model edits: give it one picture and say what to change, or two pictures and say how to combine them.

It is the heaviest image model in CozyGPU. It runs on a 96 GB Blackwell GPU (or an 80 GB Ada Pro), takes about 45 seconds a picture once the worker is warm, and its first start downloads a 60 GB worker image, so give it about 25 minutes the first time.

Model card and weights: tencent/HunyuanImage-3.0-Instruct-Distil
02

Tools it powers

  • Create ImageoptionA picture from a text prompt, with full model settings and LoRAs.Also: Z-Image Turbo, Qwen Image 2.1, Krea 2 Turbo, Ideogram 4
  • Edit ImageoptionChange a picture with instructions, up to ten reference images.Also: Qwen Image 2.1
03

Modes

  • Text to imagePrompt in, image out. 33 native sizes from 512 to 2048 pixels on a side; default 1024 x 1024.
  • Image editOne picture plus an instruction.
  • Multi-image editTwo pictures combined as the prompt describes; the result follows the first picture's shape.
04

Settings exposed

  • SizeThe model's 33 native sizes (text to image); edits follow the first picture.
  • SeedRandom, or fixed.
  • Steps and guidanceFixed by the distilled model (10 steps).
  • Custom LoRAsNot supported.

Other image models

All models