AI Image & Video Editor

The Image Editor/Generator lets you edit and generate images with AI powered by OpenAI GPT Image. Use it to transform style, change objects or clothing, edit text in images, swap backgrounds, or keep a character consistent across multiple edits.

Character consistency — preserve the same person/character identity over multiple edits.

Main features

• Style transfer — convert photos to watercolor, oil painting, sketch, etc.
• Object & clothing edits — change hairstyles, add/remove accessories, recolor garments.
• Text editing — replace words on signs, posters and labels while preserving layout.
• Background swapping — change environments while keeping subjects intact.

• Image generation & editing — a set of image-generation models tuned for style transfer, inpainting, and composition control.
• Text → Video and Image → Video — powered by the Wan 2.2 t2v/itv model (used for generating video from text or images).
• Video → Audio — powered by MMAudio V2 (Replicate deployment). This model generates context-aware audio that matches visuals (environmental sounds, action-to-sound mapping, ambient audio, etc.).
• Implementation note for power users: video audio synthesis is integrated via Replicate APIs (dynamic model version selection, file upload to Replicate file storage, predictions endpoint).

Image generation / editor operations — typically 1 credits for a single full image generation.

Background remover — typically 1 credit per removal.

Text → Video and Image → Video — typically 5–10 credits, depending on resolution, number of frames, and extras (higher resolution or more frames cost more).

Add audio to video / audio synthesis — typically 2 credits (may vary by video length & complexity).

What affects cost:

• Output resolution (e.g., 480p vs 720p)
• Number of frames and FPS (more frames/higher FPS = higher cost)
• Extra model options (LoRA, upsampling, seed control)
• “Go Fast” or other acceleration options may change resource use

Tips to save credits:

• Re-use seeds or small edits instead of regenerating entire images.
• Start with lower resolution / fewer frames to test prompts.
• Use previews and small crops before committing to full-resolution runs.

Be specific: include exact colors, objects, textures, and styles.

Bad: “Make it better.”

Good: “Make the photo into a watercolor portrait of a woman with short black hair wearing a red coat, light blue background, soft wet-on-wet brushstrokes.”

Name subjects: “the man in the plaid jacket” is better than “him.”

Preserve intentionally: if you want some parts unchanged, say so:

“Change the background to a beach while keeping the person in the exact same position and preserving facial features.”

Text replacements: use quotes and exact replacement text:

“Replace ‘SALE’ with ‘SUMMER 2025’ in the same font and size.”

Style transfer: be precise — name the movement and the visual traits:

“Renaissance portrait with warm chiaroscuro lighting and visible brushstrokes” instead of just “make it classical.”

Break complex edits into steps:

1. Remove background.
2. Edit clothing or pose.
3. Apply final style transfer.

This yields more controlled, reliable results.

Style transfer:

• “Convert this image to a 1960s pop-art poster: bold halftone dots, saturated cyan/magenta/yellow, strong black outlines.”
• Object/clothing change:
• “Change the woman’s jacket to a navy leather biker jacket with silver zippers; keep her hairstyle unchanged.”

Background swap while preserving subject:

• “Replace the background with a sunny beach at golden hour; keep the subject at the exact same scale and center position.”

Text editing:

• “Replace ‘CLOSED’ on the storefront sign with ‘OPEN’ using a similar block sans-serif font and matching kerning.”

Character consistency for a series:

• “Create five variations of this character with different hats but keep facial features, skin tone, and body proportion identical across images.”

Quick video prompt (text → video):

• “Tracking shot of a neon-lit city street at night, rain-slick pavement, camera steadily moves from left to right, cinematic 16:9, moody synth soundtrack mood.”

• Aspect ratio — choose match_input_image to preserve the original crop or set a custom ratio for specific compositions.
• Seed — set a seed for reproducible results; leave blank for randomness.
• Prompt upsampling / improvement — toggle if you want the platform to refine your prompt automatically.
• Safety tolerance — slider controls content safety strictness (0 strict → 6 permissive). When editing real photos, a maximum of 2 may be enforced for safety.
• Output format — PNG or JPG; choose PNG for transparency and lossless quality.
• LoRA / transformer weights — advanced option to apply style or model weights (use only when you know the model and weights you want to apply).

Practical tips:

• If the first result fails, change one variable at a time (prompt wording → seed → style strength).
• Use masks or selection tools when available to limit edits to specific areas.
• For text edits, try to keep the replacement length similar to preserve layout.

Main capabilities

• Generate short videos from text (Text → Video).
• Convert single images into short videos (Image → Video).
• Add audio to existing videos using MMAudio V2 (visual-to-audio synthesis).
• Model: Wan 2.2 t2v/itv for visuals; MMAudio V2 for audio synthesis (Replicate).

Key controls

• Resolution (e.g., 480p vs 720p) — higher resolution costs more credits.
• Frames per second (FPS) & number of frames — affects smoothness and duration.
• Sample shift / seed — controls sampling behavior.
• “Go Fast” — speeds up generation, with possible small quality tradeoffs.
• LoRA weights — optional model tweaks for style/quality.

Video tips:

For long or complex scenes, split into segments and stitch them together in post-production.

Start with 480p and fewer frames to prototype the motion.

Use cinematic language: camera angles, movement verbs (“tracking shot”, “dolly in”), lighting, and frame composition.

When asking for synchronized audio, provide cues such as “car passes left-to-right at 00:01–00:02; add muffled engine and tire hiss.”

What it does: generates environmental sounds and action-related audio that aligns temporally with visual events.

Good uses: ambient background, subtle action effects, accessibility (adding sound to silent videos), mock Foley for prototypes.

Limitations:

• More processing time for longer videos or noisy scenes.
• Works best when visuals are clear and actions are distinct.
• Some specialized or unique sound effects may still need manual sound design.

• Processing time and quality depend on input clarity, resolution, and scene complexity.
• Content safety filters apply (safety tolerance slider affects this).
• Respect copyright and the rights of subjects: do not upload private photos of people without permission.
• Always review outputs for accuracy (especially logos, trademarks, or text).

If output looks off:

• Make the prompt more specific (color, camera distance, texture).
• Use a mask to isolate problem areas.
• Try a different seed or lower/higher safety tolerance as appropriate.

For iterative creative work:

1. Prototype with low-res / few frames.
2. Lock composition and seed after you like the result.
3. Re-run at higher quality or apply fancy transforms (LoRA or upsampler).

For repeated character edits, save the seed and prompt that produced the best result.

You may use outputs commercially. Use outputs in apps, marketing, or product design. Always follow copyright and intellectual property rules for any source images you upload (don’t infringe on others’ copyrighted works).

Keep or preserve: “Keep facial features and pose unchanged.”

Replace text: “Replace ‘OLD’ with ‘NEW’” (include quotes).

Style: “In the style of [movement], with [traits] — e.g., ‘Renaissance, warm chiaroscuro, visible brushstrokes.’”

Camera: “Close-up portrait, 50mm lens, shallow depth of field, soft rim light.”

Motion: “Tracking shot left to right, 16:9, slight handheld shake.”

Final Tips

• Prototype first to conserve credits.
• Start simple, iterate, and keep one variable change per attempt.
• Use explicit nouns, numbers, and reference points (colors, positions, camera distances).
• Use seeds for reproducibility.