Ultimate AI Showdown: Dall-E 2 vs Stable Diffusion

Ultimate AI Showdown: Dall-E 2 vs Stable Diffusion

Table of Contents:

1. Introduction
2. A Comparison of Stable Diffusion and Dolly 2
   • Training Data
   • Model Size and Parameters
   • Accessibility and Cost
   • Restrictions
   • Generation Time
   • Additional Features
   • Future Developments
3. Image-to-Image Comparison: Prompts and Results 3.1. Comparison of Various Prompts
   • Prompt 1: Dolly 2 vs. Stable Diffusion
   • Prompt 2: Dolly 2 vs. Stable Diffusion
   • Prompt 3: Dolly 2 vs. Stable Diffusion
   • Prompt 4: Dolly 2 vs. Stable Diffusion
   • Prompt 5: Dolly 2 vs. Stable Diffusion
   • Prompt 6: Dolly 2 vs. Stable Diffusion
   • Prompt 7: Dolly 2 vs. Stable Diffusion
   • Prompt 8: Dolly 2 vs. Stable Diffusion
   • Prompt 9: Dolly 2 vs. Stable Diffusion
   • Prompt 10: Dolly 2 vs. Stable Diffusion
4. Conclusion
5. FAQ

Introduction

In this article, we will explore and compare two popular text-to-image AI models: Stable Diffusion and Dolly 2. These models have gained Attention for their ability to generate realistic images Based on textual prompts. We will analyze various factors such as training data, model size, accessibility, cost, restrictions, generation time, and additional features. Furthermore, we will conduct a comprehensive image-to-image comparison, where we will evaluate the performance of both models for various prompts.

A Comparison of Stable Diffusion and Dolly 2

• Training Data:

  • Stable Diffusion: Trained on a dataset of 5 billion+ images known as Lion 5b, which includes Captions and aesthetics ratings.
  • Dolly 2: The exact number of training images is unknown, though estimates suggest several hundred million images.

• Model Size and Parameters:

  • Stable Diffusion: Approximately 800 million parameters (subject to change during development).
  • Dolly 2: 3.5 billion parameters.

• Accessibility and Cost:

  • Stable Diffusion: Open-source model, available for researchers. A web app is planned to be released, expected to cost around $5/month.
  • Dolly 2: Not open-source. Accessible via OpenAI's prompt-based API, with Current pricing at 13 cents per prompt.

• Restrictions:

  • Stable Diffusion: Currently low restrictions within beta testing, except for explicit content. NSFW content generation possible on personal GPUs in the future.
  • Dolly 2: High restrictions, aiming for G-rated output. Certain words and topics are banned.

• Generation Time:

  • Stable Diffusion: Fast, generating images in approximately 5 seconds on the web app.
  • Dolly 2: Varies but generally generates images within 10-15 seconds.

• Additional Features:

  • Stable Diffusion: In-painting and aspect ratio options planned for future updates.
  • Dolly 2: In-painting included, but limited to a 1x1 square aspect ratio.

• Future Developments:

  • Stable Diffusion: Plans for larger models with advanced capabilities and increased VRAM requirements.
  • Dolly 2: No announced successor or larger model at present.

Image-to-Image Comparison: Prompts and Results

1. Prompt 1: 3D octane render of a cute chibi lemon character sipping a Caribbean drink on a tropical beach at sunrise.

   • Stable Diffusion: Good overall, but lacks specific sunrise details.
   • Dolly 2: Coherently represents the prompt, more accurate depiction of sunrise.

2. Prompt 2: Two lemon characters engaged in a heated discussion, with a sinister feel and harsh lighting.

   • Stable Diffusion: Varies in coherency, some images miss sinister aspect.
   • Dolly 2: Coherent representations, captures discussions and lighting.

3. Prompt 3: Ginger cat with a white chest and paws yawning and stretching on a windowsill.

   • Stable Diffusion: Inconsistent, misses aspects of yawning and stretching.
   • Dolly 2: Captures yawning and stretching, but inconsistent coherency.

4. Prompt 4: Movie still of Walter White from Breaking Bad in a lab coat holding a beaker of green liquid.

   • Stable Diffusion: Accurately represents the prompt, clear image of Walter White.
   • Dolly 2: Unable to generate Walter White, produces unrelated images.

5. Prompt 5: Close-up Studio photograph of a tsunami in a jar with swirling Water and dramatic lighting.

   • Stable Diffusion: Represents prompt creatively, follows some aspects.
   • Dolly 2: Inconsistent image quality, does not capture the essence of a tsunami.

6. Prompt 6: Pineapple pizza cupcake food photography.

   • Stable Diffusion: Creative and coherent representations of the prompt.
   • Dolly 2: Consistent and appealing imagery, but not following the prompt exactly.

7. Prompt 7: Low-angle photo of a Shih Tzu on a pirate ship.

   • Stable Diffusion: Sharp, creative, and faithful to the prompt.
   • Dolly 2: Artifacts and lack of clear image quality, fair representation of the prompt.

8. Prompt 8: Walter White cooking an egg.

   • Stable Diffusion: Accurate representation of Walter White, captures the prompt creatively.
   • Dolly 2: Generates unrelated images, unable to depict Walter White.

9. Prompt 9: Low-angle photo of a Shih Tzu on a pirate ship in the middle of the ocean.

   • Stable Diffusion: High-quality and creative representation of the prompt.
   • Dolly 2: Lack of Clarity, lower image quality, fair representation of the prompt.

10. Prompt 10: Photo of a shih tzu made entirely out of glittering Stardust.

    • Stable Diffusion: Creative and coherent representations with sharp focus on stardust.
    • Dolly 2: Inconsistent image quality, lacks sharp focus and clarity.

Conclusion

In conclusion, both Stable Diffusion and Dolly 2 offer impressive text-to-image generation capabilities. Stable Diffusion demonstrates a higher level of coherency, creative interpretations, and closely follows the given prompts. The open-source nature, lower restrictions, and ability to run on personal GPUs make Stable Diffusion an attractive option. Dolly 2, despite limitations, provides consistent and appealing imagery with faster generation times. Choosing between the two depends on specific preferences, prompt requirements, and considerations such as access, cost, restrictions, and future development plans.

FAQ

Q: Can Stable Diffusion generate explicit or NSFW content? A: Stable Diffusion has low restrictions within beta testing, allowing for NSFW content generation on personal GPUs. However, restrictions may be subject to change.

Q: Is Dolly 2 limited to generating G-rated content? A: Yes, Dolly 2 has high restrictions, aiming to generate G-rated output. It restricts certain words and topics to maintain appropriate content.

Q: Are there any plans for future improvements or updates to Stable Diffusion and Dolly 2? A: Stable Diffusion aims to release larger models with advanced capabilities and increased VRAM requirements. Dolly 2's future plans have not been announced yet, but further developments can be expected from OpenAI.

Q: Which model offers more realistic and coherent image generation? A: Stable Diffusion generally produces images that are more realistic and coherent, closely following the given prompts. However, Dolly 2 also provides consistent and appealing imagery. The choice depends on specific prompt requirements and preferences.

Q: Are there any differences in accessibility and cost between Stable Diffusion and Dolly 2? A: Stable Diffusion will be open-source, allowing free access for personal GPU usage. It is expected to cost around $5 per month for access to the web app. Dolly 2 is currently accessible via OpenAI's prompt-based API, with pricing at around 13 cents per prompt.

Q: Can I download and modify Stable Diffusion? A: Yes, Stable Diffusion is open source, which means anyone can download and modify the model. This allows for extensive modifications and the creation of personalized applications.