Getting Reliable Results from Generative Models

This guide is not a list of platform rules. It is a set of practices that make it easier for the model to give you good results. Most failed generations are not caused by something the model CANNOT do — they happen because the model did not receive input that was clear, focused, and ready to act on.

1. Write prompts as creation instructions, not chat requests

Models are best at carrying out clear creative tasks. Try to avoid vague asks like:

• "Fix this image for me"
• "Take a look at what is wrong here"
• "Make this look nicer"
• "Please fix the problems in the storyboard"

Phrasing like this often makes the model reply with a text explanation instead of generating an image.

A stronger version:

Create a new image based on the reference image.
Keep the same character and scene, but fix the duplicate panel number.
Remove the lower panel labeled 9.
Extend panel 10 to fill the empty space.
Do not add extra text, borders, or labels.
Output a clean final image.

Core principle: tell the model what image to Produce, not just what problem you are having.

2. Do one main thing per task

If a single prompt asks the model to change the character, swap the background, fix the text, preserve the style, rearrange the storyboard, enhance the detail, And adjust the composition all at once, it is much more likely to lose focus.

Break the work into smaller steps:

1. Fix composition and layout first
2. Then work on character consistency
3. Then enhance style and detail
4. Finally, do the high-resolution or finishing pass

For Image editing and Video generation especially, going Step by Step is usually more reliable than writing one enormous prompt.

3. Longer prompts are not always better

Long prompts can express complex needs, but going too long creates two problems:

• Some models have a length limit
• The model may ignore or blur the details near the end

A reliable structure:

Goal: the final result you want
Reference: which image/element must stay consistent
Changes: what needs to be modified
Style: visual style, camera, lighting, color
Avoid: what should not appear

For example:

Goal: Create a cinematic 16:9 fantasy illustration.
Reference: Keep the same woman and dragon from the reference image.
Changes: Place them at an airship skyport during golden hour.
Style: Painterly realism, warm light, detailed brass and leather textures.
Avoid: No text, no watermark, no modern objects, no extra characters.

4. With reference images, avoid conflicting inputs

More reference images do not automatically give better results. The model may not know:

• Which image controls the character?
• Which one controls the style?
• Which one controls the composition?
• Which one is just inspiration? State the purpose of each reference:

Use image 1 for character likeness.
Use image 2 for clothing style.
Use image 3 only for background mood.
Do not copy the layout from image 3.

If you only want to edit one image, start with a single primary reference — the success rate is usually higher.

5. Keep video and audio references short and specific

Video and audio models are more sensitive to clip length. A long audio or video file may not be processed in full.

Recommendations:

• Trim audio references down to the most essential moment
• Use short clips for video references
• If you only need a sense of motion or camera movement, do not upload a long video
• If you only need a vocal or sound style, do not upload an entire song or a long narration A better approach: provide just the 5–15 seconds that best represents what you want.

6. When an image is large, edit a lighter version first

Very large images, very high-resolution images, and editing several big images together all raise the failure rate significantly.

Recommendations:

• Confirm the result on a medium-sized image first
• Once you are happy with it, produce the high-resolution version
• If the model supports high-resolution output, you do not necessarily need to upload a huge original
• When editing multiple images, cut any references that are not relevant

7. Avoid asking the model to "read" a complex interface screenshot and then edit it precisely

For example:

• "Look at this storyboard, delete the 9 at the bottom, then enlarge panel 10"
• "There are two numbers in this image — fix the wrong one"

These tasks require the model to first understand a complex layout and then make a precise local edit, which fails more often.

It is better to spell out the operation:

Create a clean corrected storyboard layout.
There should be 10 panels in total.
Panel numbers must be 1 to 10, with no duplicates.
Remove the duplicate lower panel 9.
Make panel 10 wider to fill the remaining space.
Keep the same visual style and characters.

For very precise local edits, pair the prompt with a clear mask, a cropped image, or upload only the region that needs changing.

8. When the model returns text instead of an image

This usually means the model read your input as a question or request for explanation rather than a request to generate an image.

To fix it:

• Add "Create a new image..."
• Add "Output only the final image."
• Avoid questions like "Can you fix..."
• Do not describe the problem the way you would in a chat
• State clearly what the final image should look like Example:

Create a new corrected image based on the reference.
Do not explain the changes.
Output only the final image.

9. How to retry quickly after an error

If a task fails, repeatedly resubmitting the same input is not the best move. It is more effective to:

• Shorten the prompt
• Reduce the number of reference images
• Lower the output resolution
• Trim the audio or video
• Split one complex task into two simpler ones
• Rewrite it as a more direct generation instruction

10. A good prompt template

Create a [image/video] of [main subject].
 
Keep from reference:
- [what must stay the same]
 
Change:
- [what should be changed]
 
Style:
- [visual style, camera, lighting, mood]
 
Constraints:
- No text
- No watermark
- No extra characters
- Keep composition clean
 
Output:
- Generate the final image/video directly.

Good input is not about being more complex — it is about making sure the model clearly knows: what to look at, what to keep, what to change, what to avoid, and what to output.