Google’s Top AI Surpassed by HKUST Open Source? The Image Editing Powerhouse Making Overseas Creators Shout “King Bomb”

Driving Multimodal AI Toward True Generalized Creation

Introduction — The "Photoshop is Dead" Moment

In recent months, "Photoshop is dead" has become a hot topic in the AI creator community.

The explosive growth of image editing and generation models is challenging professional creative software’s long-standing dominance. Multimodal image technology upgrades from Google’s Nano Banana, ByteDance’s Seedream4.0, and Alibaba’s Qwen-Image-Edit-2509 have introduced entirely new capabilities, including:

• OOTD outfit styling
• Text rendering
• Movie storyboard generation

These tools allow creators to focus on controllability, creativity, and product readiness without mastering complex photo-retouching skills.

---

From Innovation to Limitations

While models like Nano Banana combine language understanding, visual recognition, and multimodal instruction control for a natural creative process, some limitations remain:

• Editing prompts can be ambiguous — often needing reference images or extra description
• Generation tasks excel with concrete objects but struggle with abstract concepts (e.g., hairstyles, textures, lighting, artistic styles)

---

DreamOmni2 — A New Benchmark Model

Two weeks ago, Professor Jia Jiaya’s team at HKUST open-sourced DreamOmni2, addressing these weaknesses. Built on FLUX-Kontext, it maintains existing instruction editing and text-to-image abilities, while adding multi-reference image generation editing for greater flexibility.

Key Advantages:

• Superior handling of both concrete and abstract concepts
• Outperforms popular models like Nano Banana in certain scenarios

Example — Instruction-Based Editing:

> Change the hat color of the woman in the first image to match the sweater in the second reference image.

Example — Instruction-Based Generation:

> Place image 1 on the bedroom wall, change the cup in image 3 to have the same material as the plate in image 2, and place it on the table.

DreamOmni2 quickly gained traction, amassing 1.6k GitHub Stars in just two weeks.

Repository: https://github.com/dvlab-research/DreamOmni2

---

Hands-On Testing — Where DreamOmni2 Excels

We tested DreamOmni2’s instruction-based multimodal editing and compared it with GPT-4o and Nano Banana.

---

Test 1 — Background Replacement

Prompt:

"Replace the background of the panda in picture 1 with picture 2 to generate an ID photo."

✅ Result: High-quality, fur detail preserved, professional-level output

---

Test 2 — Style Transfer

Prompt:

"Make the first image have the same style as the second image."

✅ Result: Accurate tone, mood, and stylistic integration into original image

---

Test 3 — Clothing Replacement

Prompt:

"Replace the jacket in the first image with the clothes in the second image."

DreamOmni2:

Preserves posture and face; minimal collar differences.

GPT-4o:

Face overly smoothed; awkward proportion changes.

Nano Banana:

Pose intact, but color/shape altered; logo removed.

---

Test 4 — Instruction-Based Multimodal Generation

Example Prompt:

"Print the logo from image 1 on the object from image 2 and place it on a desk."

✅ Result: Accurate logo placement, realistic light/shadow, desk context recognition

---

Test 5 — Pose Transfer from Sketch

Prompt:

"Anime image 1 adopts the pose of image 2"

✅ Result: Captured pose naturally, smooth transformation from lines to motion

---

Test 6 — Necklace on a Cat

Prompt:

"Put the necklace from image 2 on the cat’s neck from image 1"

DreamOmni2:

GPT-4o:

AI artifacts noticeable

Nano Banana:

---

Technical Innovations

1. Three-Stage Data Construction Paradigm

DreamOmni2 overcomes dataset limitations via:

• Feature Mixing Scheme — Generates high-quality concrete + abstract pairs without resolution loss.
• Instruction-Based Editing Data — Combines real + generated images to create diverse editing examples.
• Instruction-Based Generation Data — Integrates multiple reference images into cohesive datasets.

---

2. Multi-Reference Framework Design

• Uses index encoding with positional offsets to prevent artifacts and confusion between reference images.

---

3. Joint VLM + Generation Model Training

• LoRA modules activate automatically when reference images are detected.
• Improves understanding of non-standard, real-world prompts.

---

Conclusion

DreamOmni2 unifies data, framework, and training to leap forward in multimodal generation/editing.

Advancements vs Previous Models:

• Style transfer and structural reorganization with abstract attribute editing
• Flexible object + concept combination in generation tasks
• Open-source — enabling wide adoption and integration

---

Industry Impact

Platforms like AiToEarn bridge AI creation and monetization.

They integrate:

• AI content generation
• Cross-platform publishing (Douyin, Kwai, Bilibili, Instagram, YouTube, X, etc.)
• Analytics and AI model ranking (AI模型排名)

This fits perfectly with the multimodal creative future enabled by DreamOmni2.

---

References:

• Paper: https://arxiv.org/pdf/2510.06679v1
• GitHub: https://github.com/dvlab-research/DreamOmni2

---

✔ Final Takeaway:

DreamOmni2 isn’t just an upgrade — it’s a new standard for instruction-based multimodal editing & generation, transforming how creators interact with AI.

---

Do you want me to create a side-by-side visual comparison table for DreamOmni2 vs GPT-4o vs Nano Banana results? That would make the performance differences even clearer.