Add CGI To Any Video With Higgsfield

Start with the source clip, not the prompt. When you are ready to generate, use the exact Gemini Omni Flash (https://higgsfield.ai/gemini-omni-flash) link because this workflow can be hard to find from the main Higgsfield interface. The best AI VFX results come from footage with enough open space for the model to add something without rewriting the whole scene.

Use clips with:

• One readable shot and no cuts.
• A clear subject or simple empty scene.
• Enough background room for the VFX element.
• Foreground and background that are both reasonably visible.
• Simple camera motion and stable lighting.

Avoid tight, cluttered shots for the first pass. A tight shot with an out-of-focus fence can make an effect like a background portal harder to solve because the model may not cleanly understand the depth of the scene. A quick prompt like "put me in the jungle" can create the same problem from another angle: vague prompts and cramped source footage invite the model to remake the scene instead of editing it.

Pro tip: If you have a 10-second clip, trim a 2-second test first. AI video generation is usually billed by duration, so short tests are cheaper while you are still finding the right prompt.

Do not start with a tiny vibe prompt and hope the model guesses correctly. In the demo, the weak first prompt was:

Giant donut monster running past the guy in the warehouse, acid trip style.

That prompt produced a fun result, but it also changed too much of the scene. It added rainbow lighting and extra style changes that were not part of a controlled edit.

Before generating, use ChatGPT to expand the idea into a more specific video-edit prompt. The useful instruction is:

Help me improve this prompt for an AI special effects generator. I want to be more specific about exactly what to add into the scene and what not to change.

Source clip: [describe the shot]
Effect idea: [describe the effect]

Return a concise prompt that adds one effect and preserves the original subject, camera movement, framing, timing, background, and lighting.

The stronger spaceship prompt worked because it named the object, the quantity, the position, the motion, and the preservation rules:

Add one large photorealistic sci-fi craft hovering far in the background outside the parking garage. It should move slowly and feel distant, heavy, and physically present in the scene.

Keep everything else the same: same person, same camera movement, same framing, same timing, same background/environment, and same lighting direction.

Single continuous shot. No scene cuts. No new characters. No face changes. No body warping. No extra explosions or camera moves.

That specificity is the difference between "make this epic" and "add one thing without breaking the shot."

Open Higgsfield Gemini Omni Flash (https://higgsfield.ai/gemini-omni-flash), then click Start generating. In the logged-in workspace, use the video workflow that accepts source media. The visible route may be https://higgsfield.ai/ai/video?model=gemini-omni or it may normalize to https://higgsfield.ai/ai/video.

You are in the right place when the panel has an upload area for media, a prompt box, and the model set to Gemini Omni Flash.

Upload the source clip, paste the improved prompt, and generate one pass. Higgsfield may take a few minutes to generate the edit. If live generation stalls, use a prepared output and continue the walkthrough from the review step.

If Gemini Omni Flash is not available, use Higgsfield AI Video Editor (https://higgsfield.ai/ai-video-editor) or https://higgsfield.ai/ai/video/edit, but only if the editor accepts an uploaded video plus a text prompt. A fallback model such as Kling 3.0 Omni Edit or Kling O1 Video Edit is acceptable. Do not switch to a generic text-to-video or image-to-video flow and pretend it is the same workflow.

After the generation finishes, compare the AI edit against the original clip. Do not only ask whether it looks cool. Ask whether it still behaves like the same shot.

Check:

• Subject integrity: did the person, face, hands, or body change?
• Camera continuity: did the camera motion, framing, or timing drift?
• Effect integration: does the object sit in the background instead of floating on top?
• Lighting match: do the shadows, glow, color temperature, and haze fit the source?
• Source preservation: did Higgsfield change anything unrelated?

In the strong spaceship example, the prompt kept the main shot intact because it clearly said the object should be one distant sci-fi craft outside the parking garage and repeated the preservation constraints.

In the weaker examples, the model either added too much style, struggled with depth, or stacked a new spaceship in front of the old one. Those failures are useful because they tell you exactly what to fix in the next prompt.

Use the review pass to fix one issue at a time. Do not rewrite the whole concept unless the first result is unusable.

Good revision targets:

• Make the object smaller or farther away.
• Remove unwanted color effects.
• Preserve the subject more strictly.
• Match the lighting better.
• Reduce edge flicker.
• Reframe for vertical video.

Use a revision prompt like:

Revise the previous edit. Keep the same sci-fi craft idea, but make the craft feel farther away and more physically integrated into the background.

Preserve the original person, camera motion, framing, timing, background, and lighting. Keep everything else the same.

One useful workaround is to download the generated clip and upload that result back into Gemini Omni Flash as the new source if an edit gets stuck or a direct revision does not behave. Then ask for a narrower correction, such as removing unwanted rainbow light effects while keeping the main character or object.

A useful tool we found while recording was Higgsfield's Ratio and Crop Adjuster. When you edit a video, it can automatically crop and change the aspect ratio of your video.

Use this as a separate workflow. First get the VFX edit looking right. Then use Ratio and Crop Adjuster to test vertical versions for TikTok, Shorts, or Reels. If you ask for too many creative changes at the same time, it becomes harder to know which instruction caused the failure.