Bài 11: Midjourney, Flux & Emerging Models

Generative AI: Tạo Hình ảnh & Video với AI

Phần 4: DALL-E, Midjourney & Commercial APIs

xdev.asia

Giới thiệu

Thị trường image generation 2026 rất đa dạng. Ngoài Stable Diffusion và DALL-E, còn có Midjourney (aesthetic tốt nhất), Flux (open-source mạnh), Google Imagen 3, Adobe Firefly (commercial-safe). Bài này so sánh và hướng dẫn tích hợp.

---

1. Midjourney

Đặc điểm:
- Aesthetic quality tốt nhất (đặc biệt art, illustration)
- Chạy qua Discord bot hoặc Web UI
- Closed-source, subscription-based
- Hỗ trợ: text-to-image, image-to-image, vary, upscale

Pricing (2026):
- Basic: $10/month (~200 images)
- Standard: $30/month (~900 images)
- Pro: $60/month (unlimited relaxed)

Midjourney Parameters

/imagine prompt: a dragon flying over mountains --ar 16:9 --v 6 --stylize 750

Parameters:
--ar 16:9      → Aspect ratio
--v 6          → Version
--stylize 750  → Creativity level (0-1000)
--chaos 50     → Variation (0-100)
--quality 2    → Detail level
--no text      → Exclude elements
--tile          → Seamless pattern
--seed 12345   → Reproducibility

---

2. Flux

# Flux: open-source từ Black Forest Labs (ex-Stability AI team)
# Kiến trúc: DiT (Diffusion Transformer) + T5 text encoder
# Quality ngang DALL-E 3, open-source

from diffusers import FluxPipeline
import torch

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

image = pipe(
    prompt="A cat holding a sign that says 'Hello World'",
    num_inference_steps=30,
    guidance_scale=3.5,
    height=1024,
    width=1024,
).images[0]

Flux Variants

Model	License	Quality	Speed
Flux.1 Pro	API only	Best	Fast
Flux.1 Dev	Non-commercial	Great	Medium
Flux.1 Schnell	Apache 2.0	Good	Very fast (4 steps)

---

3. Google Imagen 3

# Google Imagen 3 via Vertex AI
from google.cloud import aiplatform
from vertexai.preview.vision_models import ImageGenerationModel

model = ImageGenerationModel.from_pretrained("imagen-3.0-generate-001")

response = model.generate_images(
    prompt="A peaceful Japanese garden with cherry blossoms",
    number_of_images=4,
    aspect_ratio="16:9",
    safety_filter_level="block_some",
    person_generation="dont_allow",
)

response[0].save("imagen_output.png")

---

4. Adobe Firefly API

# Adobe Firefly: trained on licensed content → commercially safe
import requests

headers = {
    "Authorization": f"Bearer {FIREFLY_TOKEN}",
    "Content-Type": "application/json",
}

response = requests.post(
    "https://firefly-api.adobe.io/v2/images/generate",
    headers=headers,
    json={
        "prompt": "A modern office space with plants",
        "contentClass": "photo",  # photo, art
        "size": {"width": 2048, "height": 2048},
        "n": 1,
        "styles": {"presets": ["photo"]},
    }
)

---

5. So sánh Platforms

Feature	SD/SDXL	DALL-E 3	Midjourney	Flux	Imagen 3	Firefly
Quality	Great	Excellent	Best (art)	Excellent	Excellent	Good
Text in images	Poor	Great	Good	Best	Good	Good
Prompt following	Good	Excellent	Good	Excellent	Excellent	Good
Speed	Fast (local)	~10s	~30s	Medium	~10s	~15s
Cost/image	Free (GPU)	$0.04-0.08	~$0.03	Free/API	~$0.04	~$0.04
Open source	Yes	No	No	Partial	No	No
Commercial use	Yes	Yes	Yes (paid)	Varies	Yes	Yes
Fine-tunable	Yes	No	No	Yes	No	No
Self-hosted	Yes	No	No	Yes	No	No

---

6. Multi-model Orchestration

class ImageGeneratorOrchestrator:
    """Route to best model based on use case"""

    def __init__(self):
        self.dalle = OpenAI()
        self.sd_pipe = StableDiffusionXLPipeline.from_pretrained(...)

    async def generate(self, prompt, use_case="general"):
        if use_case == "text_in_image":
            # Flux/DALL-E best for text rendering
            return await self._dalle_generate(prompt)
        elif use_case == "artistic":
            # Local SD with LoRA for custom styles
            return self._sd_generate(prompt)
        elif use_case == "commercial":
            # Firefly for copyright-safe content
            return await self._firefly_generate(prompt)
        elif use_case == "batch":
            # Local SD for cost efficiency
            return self._sd_generate(prompt)
        else:
            return await self._dalle_generate(prompt)

    async def _dalle_generate(self, prompt):
        response = self.dalle.images.generate(
            model="dall-e-3", prompt=prompt, size="1024x1024"
        )
        return response.data[0].url

    def _sd_generate(self, prompt):
        return self.sd_pipe(prompt, num_inference_steps=25).images[0]

---

7. Chọn Model cho Use Case

📸 Product Photography → DALL-E 3, Imagen 3
🎨 Art & Illustration → Midjourney, SD + LoRA
📝 Text in Image → Flux, DALL-E 3
🏢 Commercial (copyright safe) → Adobe Firefly
🔧 Custom Style → Stable Diffusion + LoRA
💰 Budget / Batch → Stable Diffusion (local)
⚡ Real-time → Flux Schnell, SD Turbo
🔬 Research → Stable Diffusion, Flux Dev

---

Tổng kết

Platform	Best for	Pricing model
DALL-E 3	General, text rendering	Per-image API
Midjourney	Art, aesthetic	Subscription
Flux	Open-source, text	Free/API
Imagen 3	Enterprise, Google Cloud	Per-image
Firefly	Copyright-safe commercial	Per-image
SD/SDXL	Custom, batch, self-hosted	Free (GPU cost)

📌 Bài tiếp theo: Video Generation — Sora, Runway, Kling & Pika.