How to Bloom Compares Two LLMs Painting the Same Prompt
This guide explains how to use Bloom to compare how two different large language models (LLMs) interpret and render the same text prompt as an image. You’ll enter a prompt, choose two models, and see side-by-side outputs so you can evaluate style, accuracy, and creativity. No coding is required, but you’ll need a Bloom account and access to the image generation feature.
What You’ll Need
- A Bloom account (free tier works).
- A web browser (Chrome, Firefox, Safari, or Edge).
- A clear, descriptive text prompt you want to test.
- Optional: a reference image if you want to include it in the prompt.
Step-by-Step
- Log in to Bloom – Go to the Bloom website and sign in with your credentials. If you don’t have an account, create one first.
- Open the Image Generation tool – From the dashboard, click on “Create� or “Image Generation� in the main menu. This opens the prompt input area.
- Enter your prompt – In the text box, type the exact prompt you want to test. Be specific: include subject, style, lighting, and any other details. For example, “a red fox sitting in a snowy forest at dusk, photorealistic, soft lighting.�
- Locate the model selector – Look for a dropdown menu or a settings icon labeled “Model� or “Engine.� It may be near the prompt box or under “Advanced Options.�
- Choose the first model – From the dropdown, select the first LLM you want to test (e.g., “Model A�). Do not generate yet.
- Enable comparison mode – Look for a button or checkbox that says “Compare,� “Side-by-side,� or “A/B test.� Click it. This will reveal a second model selector.
- Choose the second model – In the newly appeared selector, pick the second LLM (e.g., “Model B�).
- Adjust settings (optional) – If available, set image size, aspect ratio, or number of images per model. For a fair comparison, keep these identical for both models.
- Generate the images – Click the “Generate� or “Create� button. Bloom will process the prompt with both models and display the results side by side.
- Review the outputs – Look at each image. Note differences in composition, color, detail, and how well each follows the prompt.
- Save or export – If you want to keep the comparison, use the download button on each image or take a screenshot. Some versions allow you to save the entire comparison as a single image.
Common Mistakes
- Changing settings between models – If you alter the image size or style for one model, the comparison is unfair. Keep all settings identical.
- Using vague prompts – A prompt like “a dog� gives too little information. The models may produce wildly different results for reasons unrelated to the model quality.
- Overlooking the model selector – Some users generate with the default model without noticing the dropdown. Always verify which model is active.
- Forgetting to enable comparison mode – If you only see one output, you likely didn’t turn on the compare feature. Look for the toggle again.
- Ignoring the seed/randomness – If Bloom uses random seeds, each run may differ. For a controlled test, see if you can set a fixed seed (though this may not be available).
Tips
- Use a detailed prompt – The more specific you are, the easier it is to judge which model better follows instructions.
- Test multiple prompts – One comparison may not be representative. Try a few different subjects (e.g., a portrait, a landscape, an abstract concept) to see strengths and weaknesses.
- Take notes – Write down what you observe about each image: adherence to the prompt, artistic quality, and any artifacts.
- Check for hidden biases – Some models may handle certain styles (e.g., anime, watercolor) better than others. Note that in your evaluation.
- Use the same seed if possible – If Bloom offers a “seed� field, using the same seed for both models can reduce random variation, making the comparison more about the models themselves.
Troubleshooting
- No comparison option appears – If you don’t see a “Compare� button, your plan may not support it, or you may need to click “Advanced� first. Check the help docs or upgrade if necessary.
- Only one image is generated – Make sure you selected two different models. If the second selector is empty, try refreshing the page or logging out and back in.
- Images look identical – This can happen if the two models are actually the same (check the names) or if the random seed is fixed to the same value. Try changing the seed or using a different prompt.
- Generation fails or times out – This is often due to server load. Wait a few minutes and try again. If it persists, reduce the image resolution or simplify the prompt.
- Results are not saved – Some versions don’t auto-save comparisons. Download the images immediately after generation to avoid losing them.
Frequently Asked Questions
Can I compare more than two models at once?
Bloom’s built-in comparison currently supports only two models side by side. For more, you can generate separately and arrange the images yourself.
Does the comparison cost extra credits?
Each image generation consumes credits, so a comparison of two models will use twice the credits of a single generation. Check your plan’s credit system for details.
Can I use a reference image in the comparison?
Yes, if Bloom supports image-to-image prompts. Upload the reference image and it will be used for both models, allowing you to compare how each interprets it.
Why do the two outputs look so different even with the same prompt?
Different models have different training data, architectures, and internal randomness. That’s exactly what the comparison is meant to reveal.
Is there a way to get a text description of the differences?
Bloom doesn’t auto-generate a comparison report. You’ll need to visually inspect the images yourself. Some third-party tools may help, but they aren’t integrated.
Can I share the comparison with others?
You can download the images and share them as files, or take a screenshot of the side-by-side view. There’s no direct share link for the comparison itself.
Related
Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.