Run Nano Banana Locally?

We use an MoE LLM as the base model and introduce image understanding and generation capabilities into the model. For image understanding tasks, we utilize a joint feature composed of VAE and ViT as the image input; for image generation tasks, we incorporate Diffusion-based image modeling into the LLM architecture based on the Transfusion approach. Through mixed training of different multimodal tasks, we are able to unleash the potential of the MoE architecture and enhance the model's generalization ability. HunyuanImage 3.0 is the image generation component of this model.

Decoder-Only Transformer (Hunyuan-A13B) — unified architecture for image understanding, language modeling, and image generation.

For multimodal data with intertwined text and images, we have specifically designed attention masks and 2D positional encoding. The attention mask isolates noisy images from clean images, ensuring that different images within a rich-text sequence do not interfere with each other. By employing 2D RoPE compatible with the original 1D RoPE of the LLM base model, we promote more efficient learning and convergence of image structures.

2D RoPE illustration with joint feature — attention mask design ensures text and image tokens do not interfere during training.

Native Chain-of-Thought (CoT) Generation: To leverage the advantages of native multimodal models, we have constructed a Chain-of-Thought process for every image-to-image data point. The model is guided to first analyze the features of the original image, then further decompose the user's instructions. By combining the image feature analysis with the user's intent, the model constructs a more complex and systematic fine-grained editing command. This command includes specific details on how to modify the areas requiring change, as well as which features from the original image must be preserved. Ultimately, the model expands the user's prompt into a complex, detailed, and structured instruction.

Native CoT Generation workflow — the model analyzes input images, decomposes instructions, and constructs structured editing commands before generation.

Model Variants

Quick Start: Transformers

pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128
from transformers import AutoModelForCausalLM

model_id = "./HunyuanImage-3-Instruct"
kwargs = dict(
    attn_implementation="sdpa",
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto",
    moe_impl="eager",  # Use "flashinfer" if installed
    moe_drop_tokens=True,
)
model = AutoModelForCausalLM.from_pretrained(model_id, **kwargs)
model.load_tokenizer(model_id)

prompt = "A futuristic city at sunset with flying vehicles"
cot_text, samples = model.generate_image(
    prompt=prompt, seed=42, image_size="1024x1024",
    use_system_prompt="en_unified",
    bot_task="think_recaption",
    diff_infer_steps=50, verbose=2
)
samples[0].save("output.png")

License

Tencent Hunyuan Community License. Review the repository's current license terms and restrictions before commercial deployment.

Citation

@article{cao2025hunyuanimage,
  title={HunyuanImage 3.0 Technical Report},
  author={Cao, Siyu and Chen, Hangting and Chen, Peng et al.},
  journal={arXiv preprint arXiv:2509.23951},
  year={2025}
}

Introduction

GLM-Image is an image generation model that adopts a hybrid autoregressive + diffusion decoder architecture. According to the official release, it shows significant advantages in text rendering and knowledge-intensive generation, while also supporting text-to-image and rich image-to-image workflows.

Architecture

• Autoregressive Generator: 9B parameter model initialized from GLM-4-9B. Expands 256 compact tokens to 1K-4K visual tokens.
• Diffusion Decoder: 7B parameter single-stream DiT for latent decoding. Features a Glyph Encoder for text rendering.

Inference Cost (H100)

Quick Start: Diffusers

pip install git+https://github.com/huggingface/diffusers.git transformers
import torch
from diffusers import GlmImagePipeline
pipe = GlmImagePipeline.from_pretrained("zai-org/GLM-Image", torch_dtype=torch.bfloat16).to("cuda")
image = pipe(prompt="A futuristic city", width=1024, height=1024).images[0]

Quick Start: SGLang Server

pip install "sglang[diffusion] @ git+https://github.com/sgl-project/sglang.git#subdirectory=python"
sglang serve --model-path zai-org/GLM-Image
# Client
curl http://localhost:30000/v1/images/generations -d '{"prompt": "hello world", "size": "1024x1024"}'

Highlighted Capabilities

• Video with audio: Generates video clips up to 20 seconds with native audio generation.
• Reference-guided generation: Supports image-to-video, video-to-video, visual references, keyframes, and audio/video continuation.
• Multilingual and stylistic range: Handles multilingual dialogue, animation, cinematics, candid footage, and varied aspect ratios.
• Image generation and editing: The official announcement says complex prompt following and multilingual text rendering have improved over earlier FLUX generations.
• Action-aware research: The same backbone is being explored for action prediction and robotics partnerships such as FLUX-mimic.

Release Plan

• FLUX 3 Video: Video and audio generation/editing through APIs and private weight access after Early Access.
• FLUX 3 Image: Image synthesis and editing through APIs and private weight access; image Early Access is expected in the following weeks.
• FLUX 3 Dev: Black Forest Labs lists future open-weight access to a multimodal backbone for content creation and action prediction.

Benchmark Caution

The official comparisons against other video models are preliminary, use an evolving model and evaluation harness, and should not be treated as a stable leaderboard ranking. We will add reproducible results only after public access and fixed evaluation details are available.

Official Source

Read the Black Forest Labs FLUX 3 announcement for the current Early Access status, capability descriptions, evaluations, and launch plan.

Key Features

• Rich Content: Up to 4.5K-token instructions for newspapers, storyboards, exam papers, complex diagrams, and nine-panel knowledge graphics.
• Authentic Details: Official examples emphasize approximately 10px text, formulas, annotations, pores, hair strands, and realistic material textures.
• Deep Knowledge: Native rendering across 12 languages, multiple fonts, 100+ artistic styles, and realistic web, game, and livestream interfaces.
• Spatial Composition: Supports side-by-side content expansion and nested picture-in-picture interfaces in one generated image.

Official Release Status

Qwen-Image 3.0 is currently documented through the official announcement and available as a hosted experience. As of this page update, the official Qwen repositories and model hubs do not provide downloadable 3.0 weights, a confirmed license, a model card, parameter count, benchmark table, or local deployment instructions.

Official Sources

• Qwen official announcement
• Alibaba Cloud Community English version
• Qwen Chat image generation

Images: Qwen-Image 3.0 official announcement examples, mirrored to this site's Cloudflare R2 CDN for efficient delivery.

Project Updates

• 2026.05.08: Open-sourced HiDream-O1-Image (8B) — both undistilled (full) and distilled (Dev) variants, plus the Reasoning-Driven Prompt Agent.
• 2026.05.14: Open-sourced HiDream-O1-Image-Dev-2604 with its prompt refiner, tailored for text-to-image.
• 2026.05: Dev-2604 was released as the distilled text-to-image variant.

Key Features

• Pixel-Level Unified Transformer: One end-to-end model on raw pixels, no VAE, no disjoint text encoder.
• One Model, Many Tasks: Text-to-image, long-text rendering, instruction editing, subject-driven personalization, and storyboard generation in a single architecture.
• Reasoning-Driven Prompt Agent: Built-in "thinking" agent that resolves implicit knowledge, layout, and text rendering before generation.
• Native High Resolution: Direct synthesis up to 2,048 × 2,048 with sharp detail.
• 8B Unified Architecture: One released model family covers generation, editing, and subject-driven workflows.

Model Variants

Installation

git clone https://github.com/HiDream-ai/HiDream-O1-Image.git
cd HiDream-O1-Image
pip install -r requirements.txt
# flash-attn is strongly recommended. Without it, set "use_flash_attn": False
# in models/pipeline.py (line ~341) or inference will fail to import the kernel.

Quick Start: Text-to-Image

# Undistilled full model (best quality)
python inference.py --model_type full --prompt "A serene mountain lake at sunset, cinematic lighting"

# Distilled dev model (faster)
python inference.py --model_type dev --prompt "A serene mountain lake at sunset, cinematic lighting"

Reasoning-Driven Prompt Agent (optional)

# Rewrites a raw instruction into a self-contained English prompt.
# Backend A: local Gemma-4-31B-it   Backend B: external OpenAI-compatible API
python prompt_agent.py --backend api --instruction "做一张猫咪生日party的海报，标题写HAPPY BIRTHDAY"
# Feed the returned "prompt" field into inference.py for best results.

Web Demo

python app.py --model_path /path/to/HiDream-O1-Image --model_type full
# launches the repository's local Flask interface

License

The code and the HiDream-O1-Image models are released under the MIT License — free for personal and commercial use.

Citation

@article{hidreamolimage,
  title={HiDream-O1-Image: A Natively Unified Image Generative Foundation Model with Pixel-level Unified Transformer},
  author={Cai, Qi and Chen, Jingwen and Gao, Chengmin and others},
  journal={arXiv preprint arXiv:2605.11061},
  year={2026}
}

Model Zoo

Decoupled-DMD Algorithm

Core Insight: Decoupling CFG Augmentation (Engine) from Distribution Matching (Shield). This allows 8-step inference with high fidelity.

Community Acceleration (Low VRAM)

• stable-diffusion.cpp: Pure C++ inference. Runs on 4GB VRAM. View Details →
• Cache-DiT: 4x speedup on multi-GPU.
• LeMiCa: Training-free timestep acceleration.

Citation

@article{team2025zimage, title={Z-Image: An Efficient Image Generation Foundation Model}, author={Z-Image Team}, year={2025}}