fix-my-look

fix-my-look

Edit the source's first usable frame with gpt-image-2 from the user's prompt, then propagate that look across the clip with kling reference-video while locking the original face, motion and audio via the original video + audio as references. All prep happens in one normalize_video call for short clips, or one normalize call per segment for longer clips. The output ratio uses the normalized clip's closest supported output ratio; this skill does NOT reframe the source video.

Inputs

• <source> — path or URL to a video file with audio
• <change_prompt> — what to change (e.g. "make it night with neon lights", "change my shirt to a leather jacket", "put me on a beach in Hawaii")

Empty-args menu

1. "What's the source video path?"
2. "What do you want to change? (e.g. 'put me on a beach', 'make it night')"

Workflow

Working dir: ~/Downloads/fix-my-look/<run-id>/.

Step 0 — Cost, timer and task IDs

Every tool named below is a Pika MCP tool, written by its bare name. Call it under whatever prefix your session exposes for the Pika MCP.

Start a timer when the source and change prompt are known. Before paid generation, call estimate_cost for the planned generate_image_edit, generate_reference_video, any multi-segment edit_concat, and any optional audio/lipsync repair call. If cost is not surfaced by the host, say Cost not surfaced by this harness in the final report instead of guessing. When any tool returns a task_id, copy the exact value into the run notes and reuse it verbatim; do not hand-type long JWT-style task IDs.

Step 1 — Prepare the clip

Local file? upload_asset it first; an HTTPS media URL passes directly. Decide the source windows before normalizing: use one 14.8s window for sources <=15s, and split longer sources into ordered 14.8s windows. Call normalize_video(video_url=<source>, start_s=<offset>, max_duration_s=14.8, extract_audio=true, extract_face_frame=true) once per window. Use the first window's face_frame_url for the edited still; use each window's video_url as that segment's motion/identity reference. For multi-window clips, also call extract_audio_from_video(video_url=<source>) so the final merged output can be restored to one continuous source audio track.

Wire the result into the rest: face_frame_url is the Step 2 edit target; each normalized video_url is Kling's reference for that segment in Step 4; set aspect_ratio = result.aspect_ratio ?? result.closest_aspect_ratio for each normalize result, then carry that local aspect_ratio through the image and video calls. If neither field is present, stop and report that normalization output is missing an aspect label. Compute duration = max(4, min(15, round(duration_s))) per segment, and use resolution="720p" unless the user asked for high res. If face_found is false, no clear face was found and face_frame_url fell back to the t=0 frame — proceed but warn identity may drift, or re-run with a start_s at a section where the subject faces camera.

Reference-video providers can reject oversized reference assets. If the normalize result or the downstream provider error shows a normalized video is over the provider limit, retry normalize_video once with crf=28 and the same start_s, max_duration_s, extract_audio, and extract_face_frame values. If the reference is still too large, stop before another paid video attempt and report that normalize_video needs a worker-side 1080-edge / reference-size cap. Do not patch this with local shell media commands.

Step 2 — Edit the frame with gpt-image-2 (the "change" stage)

generate_image_edit with provider="gpt-image-2", aspect_ratio=<aspect_ratio>, resolution="2K", images=[<face_frame_url>], quality="high", prompt:

"Modify the reference photograph as follows: <change_prompt>. Keep the person's face, identity, hair, body and pose EXACTLY as in the reference. CRITICAL: preserve every object the subject is holding or touching — phones, products, drinks, bags, props, jewelry — in the exact same hand, position, orientation and scale; never remove, replace or restyle them. Change only the requested scene, background, clothing, lighting or environment, not who the person is."

Keep the "preserve held objects" clause verbatim on every re-render — without it gpt-image-2 silently drops products/phones the subject is holding.

If gpt-image-2 returns a content-policy false positive for fashion, glam, or beauty prompts, retry once with the same intent but a modest / editorial wording such as "polished event styling, opaque clothing, natural pose, non-sexual fashion portrait". For makeup prompts, explicitly preserve the original eye shape, eyelids, iris color and gaze; heavy eyeliner/eye shadow is a high-risk identity-drift source.

Step 3 — Show the edited frame and wait for approval

Surface the edited frame and STOP. Ask "Approve for video generation, or tweak and re-render?" Do NOT call video generation until approved. For tweaks, re-run Step 2 (locked clauses verbatim) and loop.

Step 4 — Propagate via Kling reference-video

For each normalized segment, call generate_reference_video with provider="kling", reference_videos=[<segment video_url>], reference_images=[<edited_frame_url>], aspect_ratio=<aspect_ratio>, duration=<segment duration>, sound=false, video_keep_sounds=[true], prompt:

"Apply the change shown in <<<image_1>>> to <<<video_1>>>. Keep the person in <<<video_1>>> with the EXACT same face, identity, expressions, motion and timing; preserve the original video's kept sound track. The new scene/background/clothing/lighting should match <<<image_1>>>. CRITICAL: preserve every object the subject is holding or touching in <<<video_1>>> — phones, products, drinks, bags, props — in the same hand and orientation every frame. Keep mouth motion active through the final frame when the person is speaking. Do not alter the person's identity."

Append any extra creative direction (e.g. "very cinematic, soft golden light") after the locked text — never replace it.

Do not pass sound=true to Kling with a video input. Kling rejects that combination with error:1201 sound on is not supported with video input; use sound=false plus video_keep_sounds=[true] to keep the source video's audio.

If the source was split into multiple windows, call edit_concat(video_urls=[<segment outputs in order>]). After concat, run edit_audio_replace(video_url=<concat_url>, audio_url=<full_source_audio_url>, duration_policy="video") when the merged output audio is missing, drifted, or discontinuous.

Only try Seedance if the user explicitly asks for it, or if Kling fails and a second provider attempt is useful. Use the same segmenting rule and record the provider error plainly if Seedance rejects the input or drops speech/action.

Async handling: if any call returns a {task_id, status} envelope, poll task_status({task_id}) in a tight loop until terminal.

Step 5 — Audio, duration and identity QA

Before reporting success, verify the generated video against the source:

• Duration must not be meaningfully cut off. If output duration differs from the intended source window or merged source duration by more than 0.5s, mark the run as failed / needs follow-up.
• If the source has speech, audio must be present through the tail and mouth movement must not freeze before the spoken content ends. If words are missing, garbled, silent, or visibly out of sync, do not call the run PASS.
• The approved frame corrections must persist into the video. If the provider reintroduces a removed artifact such as eyeglass glare, mark it as a propagation caveat or re-render from a stronger approved frame.
• Compare identity at start, middle, segment boundaries, and end. If Kling preserved motion but changed the face, call that out as a provider limitation instead of a pass.

If the video is visually acceptable but speech audio is missing, incomplete, or drifted, offer one paid repair pass:

1. edit_audio_replace(video_url=<generated_video_url>, audio_url=<full_source_audio_url or segment_audio_url>, duration_policy="video")
2. edit_lipsync(video_url=<audio_restored_url>, audio_url=<full_source_audio_url or segment_audio_url>, variant="v2-pro")

If the model froze the mouth near the end, do not keep escalating to sync-3 automatically; lip-sync cannot reliably recover a face track with no mouth motion. Offer trim / regenerate instead.

Step 6 — Download + return

Download the result to ~/Downloads/fix-my-look/<run-id>/result.mp4 and return that path plus the final report fields: source, edited frame URL, final video URL, provider, job/task IDs, cost estimate or not surfaced, elapsed time, QA notes, and follow-up issue.

Failure modes

Symptom	Cause	Fix
Output face drifts from the original	gpt-image-2 over-edited the face OR the provider under-weighted the source video	Re-run Step 2 with a stronger "keep the face the same" clause; soften change_prompt.
Output looks like the original (no change)	Edited image too similar, OR you passed the raw frame not the edited output	Re-run Step 2 with a more dramatic prompt; confirm the edited frame URL.
Output aspect doesn't match source	Source aspect not in {16:9, 9:16, 1:1, 4:3, 3:4}	Step 1 returns aspect_ratio, or closest_aspect_ratio on older worker payloads; use it as the closest supported output label and ask the user for exotic aspects.
Provider rejects the normalized video as too large	normalize output can remain too large for 4K/iPhone sources	Retry normalize once with crf=28; if still too large, stop and file worker follow-up for a 1080-edge / reference-size cap.
Long source only returns the first short window	The caller normalized once with max_duration_s=14.8 and skipped segmenting	Split into 14.8s windows, generate each segment, then edit_concat in order and restore full source audio if needed.
Speaking clip loses sound, drops words, or freezes mouth at the tail	Provider regenerated speech/audio instead of preserving the source, or the face track has no mouth motion to drive	Mark as not pass. Offer one edit_audio_replace + edit_lipsync repair pass; if tail mouth motion is frozen, offer trim/regenerate instead.
Approved frame fix disappears in the video	Provider propagation reintroduced the original artifact	Re-render from a stronger approved frame or mark provider propagation caveat; do not claim the frame correction shipped.
Kling rejects with error:1201 sound on is not supported with video input	sound=true was passed with a video reference	Retry the Kling call with sound=false and video_keep_sounds=[true]; do not use reference_audio for Kling video input.
Kling output is shorter than the normalized source	Provider returned a shorter render, or the caller accidentally passed a trimmed reference	Do not mark pass. Compare output duration to the normalized source, then regenerate that segment or ask the user for a shorter window.