Google Imagen 4 Sunset Countdown: Can ERNIE-Image Be the Open-Source Alternative?

On August 17, 2026, Google will officially shut down all three Imagen 4 endpoints. For developers and businesses relying on Google's image generation API, the countdown is ticking — less than three weeks remain. For those who need multilingual text rendering, local deployment, and zero API costs, the open-source community has a compelling answer.

Imagen 4: From Google Flagship to Retirement

In April 2026, Google was still actively promoting Imagen 4's text rendering capabilities and high-fidelity image generation. Just three months later, the entire Imagen 4 lineup has entered its retirement countdown.

According to Google's official deprecations page, the following three Imagen 4 model endpoints will shut down on August 17, 2026:

Model ID	Status	Shutdown Date	Recommended Replacement
imagen-4.0-generate-001	Deprecated	2026-08-17	gemini-3.1-flash-image
imagen-4.0-ultra-generate-001	Deprecated	2026-08-17	gemini-3.1-flash-image
imagen-4.0-fast-generate-001	Deprecated	2026-08-17	gemini-3.1-flash-image

This isn't Google's first image model retirement. Imagen 3 was shut down in November 2025. Each model iteration brings API endpoint changes, code rewrites, and migration costs.

Google recommends migrating to Gemini 3.1 Flash Image (community nickname: Nano Banana). But Nano Banana is also a closed-source API model — you can't download it, run it locally, fine-tune it, or train LoRAs on it.

For many developers, this means: yet another API migration, yet more vendor lock-in.

Why Look for an Open-Source Alternative?

Closed-source API models carry three fundamental risks:

1. Vendor lock-in: Google can change pricing, shut down endpoints, or modify terms of service at any time. Imagen 4 users who migrated to the new API in June 2025 now face another migration in August 2026.
2. Privacy and compliance: Every API call means your prompts and generated images pass through Google's servers. For scenarios involving customer data or trade secrets, this may be unacceptable.
3. Unpredictable costs: API calls are billed per-use. At high throughput, costs accumulate rapidly. Self-hosted models require a one-time hardware investment with near-zero marginal cost thereafter.

ERNIE-Image: 8B Parameters, Apache 2.0, Open-Source Ready

On April 15, 2026, Baidu open-sourced ERNIE-Image — an 8B-parameter text-to-image model built on a single-stream Diffusion Transformer (DiT) architecture, released under the Apache 2.0 license.

Key Specifications:

Feature	ERNIE-Image	Imagen 4	Nano Banana
Open Source	✅ Apache 2.0	❌ Closed	❌ Closed
Local Deployment	✅ Supported	❌ Not supported	❌ Not supported
Parameters	8B	Unknown (closed)	Unknown (closed)
Chinese Text Rendering	✅ Best-in-class	❌ English only	✅ Supported
Cost	Zero (self-hosted)	API billing	API billing
Fine-tuning/LoRA	✅ Supported	❌ Not supported	❌ Not supported
Privacy	✅ Fully local	❌ Data to cloud	❌ Data to cloud

Benchmark Comparison

GenEval Complex Instruction Following

On GenEval, ERNIE-Image (w/o PE) leads all open-source models with a score of 0.8856:

Model	Single Obj	Two Obj	Counting	Colors	Position	Att. Binding	Overall
ERNIE-Image (w/o PE)	1.0000	0.9596	0.7781	0.9282	0.8550	0.7925	0.8856
Qwen-Image	0.9900	0.9200	0.8900	0.8800	0.7600	0.7700	0.8683
FLUX.2-klein-9B	0.9313	0.9571	0.8281	0.9149	0.7175	0.7400	0.8481

LongTextBench Text Rendering

Text rendering is ERNIE-Image's strongest advantage. On LongTextBench:

Model	English	Chinese	Average
Seedream 4.5 (closed)	0.9890	0.9873	0.9882
ERNIE-Image (w/ PE)	0.9804	0.9661	0.9733
GLM-Image	0.9524	0.9788	0.9656
Qwen-Image-2512	0.9560	0.9650	0.9605

ERNIE-Image is the strongest text-rendering open-source model, handling both English and Chinese. By contrast, Imagen 4 only supports English prompts and consistently lags behind ERNIE-Image in third-party text rendering evaluations.

Migrating from Imagen 4 to ERNIE-Image: A Three-Step Guide

Step 1: Environment Setup

ERNIE-Image has minimal hardware requirements — a single 24GB GPU (RTX 3090/4090) is sufficient.

Quick Start with Diffusers:

pip install git+https://github.com/huggingface/diffusers

import torch
from diffusers import ErnieImagePipeline
pipe = ErnieImagePipeline.from_pretrained(

"baidu/ERNIE-Image",

torch_dtype=torch.bfloat16,

).to("cuda")
image = pipe(

prompt="A photo of a Japanese-style ramen shop interior, warm lighting, wooden counter, steam rising from the soup",

height=1024,

width=1024,

num_inference_steps=50,

guidance_scale=4.0,

use_pe=True

).images[0]

image.save("output.png")

Step 2: Migrate Imagen API Calls to Local

Before (Imagen API):

from google import genai
client = genai.Client()
response = client.models.generate_images(
    model='imagen-4.0-generate-001',
    prompt='A futuristic cityscape at sunset',
    config={'number_of_images': 1, 'aspect_ratio': '16:9'}
)

After (ERNIE-Image Local):

import torch
from diffusers import ErnieImagePipeline
pipe = ErnieImagePipeline.from_pretrained(

"baidu/ERNIE-Image", torch_dtype=torch.bfloat16

).to("cuda")
image = pipe(

prompt="A futuristic cityscape at sunset, cyberpunk style, neon lights reflecting on wet streets, flying vehicles in the sky",

height=848,

width=1512,

num_inference_steps=50,

guidance_scale=4.0,

use_pe=True

).images[0]

Key differences:

• Imagen 4 uses API key + remote calls; ERNIE-Image uses local models
• ERNIE-Image supports much longer prompts (far exceeding Imagen's 480-token limit)
• ERNIE-Image supports Chinese prompts (Imagen 4 does not)
• ERNIE-Image allows custom inference steps and guidance scale

Step 3: Production Deployment

For production, SGLang deployment is recommended for optimal performance:

# Install SGLang
git clone https://github.com/sgl-project/sglang.git
pip install -e "python[diffusion]"
Start ERNIE-Image service
sglang serve --model-path baidu/ERNIE-Image 

--num-gpus 1 

--performance-mode auto 

--port 30010

Once running, call via OpenAI-compatible API:

curl -X POST http://localhost:30010/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "An oil painting of a mountain lake, impressionist style",
    "height": 1024,
    "width": 1024,
    "num_inference_steps": 50,
    "guidance_scale": 4.0,
    "use_pe": true
  }' \
  --output output.png

You can replace Imagen API calls with your self-hosted endpoint with near-zero code changes — just update the base_url in any language client.

Turbo Mode: 8-Step Inference for Speed

If Imagen 4 Fast's response speed matters to you, ERNIE-Image-Turbo requires only 8 inference steps, generating a 1024×1024 image in 2-3 seconds on an RTX 4090:

pipe = ErnieImagePipeline.from_pretrained(
    "baidu/ERNIE-Image-Turbo", torch_dtype=torch.bfloat16
).to("cuda")
image = pipe(

prompt="A minimalist logo design with geometric shapes",

height=1024,

width=1024,

num_inference_steps=8,

guidance_scale=1.0,

use_pe=True

).images[0]

The Turbo model uses DMD (Distribution Matching Distillation) and RL optimization to deliver 6×+ speedup while maintaining quality.

Cost Comparison

Solution	Initial Cost	Monthly Operating Cost	Cost Per Image	Privacy
Imagen 4 API	$0	~$150 (100K images/mo)	~$0.0015/image	❌
Nano Banana API	$0	~$390 (100K images/mo)	~$0.039/image	❌
ERNIE-Image Self-hosted (RTX 4090)	~$2,000 (one-time)	~$50 (electricity)	~$0.0005/image	✅
ERNIE-Image Cloud GPU (RunPod)	$0	~$100 (A100 40GB)	~$0.001/image	✅

Note: Estimates vary based on usage volume and cloud provider.

Who Should Migrate Immediately?

Strongly recommended to migrate:

• Chinese content creators: If you need images with Chinese text (posters, social media graphics, e-commerce images), ERNIE-Image's Chinese text rendering far exceeds Imagen 4 (which doesn't even support Chinese prompts)
• Privacy-sensitive applications: Healthcare, finance, legal — scenarios where data cannot leave local infrastructure
• High-throughput scenarios: 100K+ images per month — self-hosting offers dramatic cost advantages
• Advanced customization: Users needing LoRA fine-tuning, ControlNet, or custom VAE control

Can wait:

• Users working exclusively in English who don't need local deployment can migrate to Nano Banana
• Users primarily needing image editing — ERNIE-Image's editing model hasn't been released yet (community alternatives like FLUX Kontext or Bernini-R can fill the gap)

The Editing Model: Community Expectations

As of July 30, 2026, the ERNIE-Image editing model (Inpainting/Outpainting) has not been officially released. However, the community has several effective alternatives:

• Simple edits: img2img + mask solution is fully functional
• High-quality edits: FLUX Kontext provides professional-grade editing
• Text + edits: Bernini-R hybrid pipeline preserves text rendering advantage

Based on Baidu's roadmap, the editing model Beta is expected in July-August 2026 via ComfyUI nodes. For users needing editing capabilities, a recommended workflow combines ERNIE-Image for base generation with community alternatives for editing.

Summary: The Era of Open-Source Alternatives Has Arrived

Imagen 4's retirement is not an isolated event — it's the latest example of the inherent risks of closed-source API models. Every model iteration, every endpoint shutdown reminds developers: the convenience of APIs comes with the cost of vendor lock-in.

ERNIE-Image, as the strongest open-source text-to-image model under Apache 2.0, achieves SOTA performance in text rendering, instruction following, and structured generation. In Chinese text rendering, it's the only model that can compete with closed-source flagships.

Migrating from Imagen 4 to ERNIE-Image is more than a technical decision — it's a transition from "renting capability" to "owning capability."

---

Sources: Google Gemini API Deprecations, ERNIE-Image Technical Report (arXiv:2605.25347), HuggingFace baidu/ERNIE-Image, SGLang Documentation, Google Imagen Migration Guide