1. What Are Keyframes and Why Do They Matter? {#section-1}
One sentence explanation: You show the AI two images — what the start looks like and what the end looks like — and the AI figures out how to transition between them.
This is completely different from text-only generation. Text generation is like “painting blind” — the AI guesses what you want. Keyframes are like “painting by reference” — you lock in the start and end, and the AI only fills in the path between them.
Why learn this?
| Benefit | Explanation |
|---|---|
| Precise control | Results are predictable, not RNG-based “gacha” |
| Natural transitions | AI interpolates frames. Much smoother than hard cuts or simple fades |
| Creative freedom | Impossible effects become possible — human to animal, summer to winter |
2. Beginner: Generate a Video from Two Images {#section-2}
This is the foundation. Follow these 6 steps:
Step-by-Step Workflow
Step 1: Enter Keyframe Mode
Open Dream Machine web → Click + icon in bottom-left → Select Keyframe Video
Step 2: Upload Start Frame
In the Start Frame card, upload your first image. Minimum 720p recommended.
Step 3: Upload End Frame
In the End Frame card, upload your second image. Important: Both images must have identical resolution and aspect ratio.
Step 4: Check Alignment (The Step Everyone Skips)
Click Preview Alignment. If red highlights appear on faces or main subjects, your two images have significantly different subject positions. Go back to Photoshop/editor and adjust before re-uploading.
Step 5: Write Your Transition Prompt
Don’t just write anything. Follow this structure: Camera motion + Subject change + Environment detail
✅ Good example: "dolly zoom out slowly, woman turning her head left to right while hair lifts upward, background city lights blur into motion streaks"
❌ Bad example: "a woman turning her head"
Golden rule: Use concrete verbs like rotate, lift, drift, unfurl. Avoid vague words like change or become.
Step 6: Generate
Click Generate. Wait 30 seconds to a few minutes. Watch the magic happen.
3. Intermediate: Extend Your Video to 9 Seconds {#section-3}
The default video length is 4-5 seconds. Not enough? Use Extend to add more.
When to Use Extend
- Your current video isn’t long enough
- You want to break a complex transition into multiple controlled segments
Step-by-Step Workflow
Step 1: Generate a base video first
Use the beginner method to get a 4-5 second initial video.
Step 2: Click the Extend button
Find and click Extend on the generation results page.
Step 3: Upload a new keyframe image
Drag a new image into the popup area. This becomes the new “destination” — your video will transition smoothly from its current ending to this new image.
Step 4: Generate the extended video
Submit and wait. The system renders the extension. Maximum total length: 9 seconds.
Step 5: Preview and save
If satisfied, click the ⋯ in the top-right corner to save or share. If not, extend again.
💡 Tip: You can extend multiple times, but each extension consumes credits. Plan ahead to avoid waste.
4. Advanced: Fix Warping and Distortion with Modify {#section-4}
Generated video has a warped face? Clipping through objects? Don’t regenerate the whole thing. Use Modify for “micro-surgery.”
When to Use Modify
- Face is distorted or unrecognizable
- Body parts are clipping through other objects
- Only part of the video is wrong — the rest is fine
Step-by-Step Workflow
Step 1: Locate the video to fix
Go to Boards → Find the video card under your project.
Step 2: Enter Modify mode
Click ⋯ in the bottom-right corner of the card → Select Modify.
Step 3: Upload anchor frames
In the Modify panel:
- Start Frame: Upload a screenshot of the first frame of your current video
- End Frame: Upload a screenshot of the last frame of your current video
Why screenshots, not the original images? Because the AI needs to know the current state of the video, not what you originally intended. Both screenshots must come from the same video, with timestamps exactly matching frame 0 and the last frame.
Step 4: Adjust the Strength slider
This is the core parameter. Luma AI defines three distinct modes:
| Mode | Slider Position | Effect | Use Case |
|---|---|---|---|
| Adhere | Far left | Tightly follows original shape, edges, details | Texture swap, color change, lighting tweaks |
| Flex | Middle | Follows motion structure but allows creative freedom | Default choice – most fix scenarios |
| Reimagine | Far right | Blurs edges and shapes, freer transformation | Cross-species (human to animal), heavy stylization |
Step 5: Write the fix prompt
Example: "replace the character in image2 with the cat in image1, make the background the living room from image1"
image1 = your character reference imageimage2 = the original image that needs modification
Step 6: Submit modification
The system only regenerates the affected middle frames. Everything else stays the same. This consumes far fewer credits than full regeneration.
💡 Tip: For testing, use 540p resolution first. Confirm your prompt works, then regenerate at 1080p. 1080p consumes significantly more credits.
5. Case Study: Bringing a Photo to Life {#section-5}
Let’s walk through a real example that ties everything together.
Goal: Turn a two-person photo into a video where they slowly turn their heads in a forest, with shifting background light.
Complete Workflow
Preparation Phase:
- Prepare a high-quality photo (clear subjects, even lighting, min 1024×768 resolution)
- Prepare an End Frame — can be a Midjourney-edited variation (same subject positions, only change pose or background)
Generation Phase:
- Enter Keyframe Video mode. Upload both images
- Click Preview Alignment to confirm alignment. If red highlights appear, subject positions are too different — go back and adjust in Photoshop
- Enter prompt:
"two people in forest turning heads slowly left to right, dolly zoom out, sunlight shifting across faces, static camera, no zoom, smooth transition" - Strength setting: Start with Flex mode (slider ~0.5)
- Click Generate. Wait for rendering
Fix Phase (if the result isn’t right):
- If faces are distorted, enter Modify mode
- Extract first and last frame from the generated video. Upload to Start/End Frame slots
- Set Strength to Adhere mode (far left). Use prompt:
"fix faces, keep natural expression, adhere to original facial structure" - Regenerate the corrected version
Extend Phase (if you need more length):
- Once satisfied, click Extend
- Upload another same-style image as the extension target
- Generate the full 9-second version
6. Troubleshooting Common Issues {#section-6
| Problem | Likely Cause | Solution |
|---|---|---|
| Jump cuts, severe subject warping | Subject position/size differs too much between frames | Go back to Photoshop/editor. Adjust so subjects are similarly positioned in the frame |
| Ghosting, flickering, double images | Strength too high OR prompt too vague | Lower Strength to Flex range. Add "stable transition, smooth motion" to prompt |
| Generated video doesn’t start from my first image | Keyframes not properly aligned | Click Preview Alignment. Red areas indicate misalignment. Adjust and re-upload |
| Faces are distorted | Modifier strength too high | Use Modify mode: upload start/end frame screenshots. Set Strength to Adhere mode |
| Extend creates a visible seam | Refine Loop not enabled | Check {loop}. Enable Refine Loop slider to 30-50 range |
| Credits drain too fast | Testing at 1080p resolution | Test at 540p first. Only use 1080p for final renders |
Summary: Your Learning Path
| Stage | Skills | Estimated Time |
|---|---|---|
| Beginner | Basic keyframes (two-image generation) | 1 day |
| Intermediate | Extend, multi-segment transitions | 3-5 days |
| Advanced | Modify for warping, Strength tuning | 1-2 weeks |
| Expert | Character reference for consistency | 1+ month |
Core philosophy: Keyframes are about “certain start + certain end + AI uncertainty in the middle.” The tighter you lock your start and end, the less the AI can go off track.
Master these three things, and you’re already ahead of 90% of users:
- Preview Alignment before generating
- The right Strength mode for your task
- Specific verbs in your prompts
