# Ultra-Realistic AI Ads: A Partial Production Workflow Source: [Ultra-Realistic AI Ads: A Partial Production Workflow]() Content index: [llms.txt]() Last updated: 2026-10-28 Ad filmmaker An ultra-realistic AI ad needs stable product, character and location references, visible state continuity, motion tests and shot-level take review before editing and sound finishing. This is a **partial production workflow**: ChatCut’s historical headphone-ad run generated assets and 25 production takes, then stopped before office, packshot, assembly, mix or export. Finished-ad realism and campaign performance remain unverified. ## Verified facts and their scope **Last verified: October 28, 2026.** Facts below come from the **ChatCut A Ultrarealistic Headphones Ad acceptance report, October 2026**, a historical local integration run. The internal report is not publicly hosted; current deployed production was not retested. | Layer | Recorded scope | Evidence and limitation | | --- | --- | --- | | Native model | No independent native-capability or cross-platform quality test in this case. | A report; consult the separately sourced [model matrix](https://pr-693-preview.chatcut.dev/docs/model-capability-matrix). | | ChatCut configured route | GPT Image 2 and Sunburst stills, Seedance 2.0 tests/takes, Mureka 9.5 song. | A historical local run; no current-production availability claim. | | Preparation | 161 images, ten motion tests and nine compiled production prompt blocks. | A report; required kitchen fit edit and parts of the moment coverage were missing. | | Production | Kitchen and stadium batches plus four unselected street takes: 25 production takes total. | A report; all used 1080p, 16:9, 15-second Seedance 2.0 settings. | | Delivery | No office, packshot, assembled ad, mix or export. | A acceptance failed; final identity, product crops, pacing and sound unassessed. | ## What must be locked before generating? Write the intended scene order and the moments each scene must show. For a wearable product, decide when the viewer first sees it, when the person puts it on, and which gesture connects one location to the next. Specify the required camera family and product visibility for each moment. Prompt and shot totals alone do not establish coverage. Approve one product sheet and separate character and location references. In the headphone study, this meant a fictional unbranded product, hero and boss sheets, a dancer reference, and four location plates. Keep the original approvals as the comparison target; an attractive generated take can still introduce a face, prop or product change that should not become the new reference. ## How do motion tests reduce uncertainty? A tested preparation stage used four eight-second kitchen cells and two ten-second clips per later location, ten clips and 92 seconds in total. Comparing candidates in motion helped choose a hero. Treat this as a test design example, not a required spending recipe or evidence that all later actions will succeed. Use the smallest useful test to expose a specific risk: face stability while turning, hands touching the product, an outfit moving, or whether the location leaves room for the action. Set an acceptance question before submitting the batch. Check the live quote and decide whether to revise the reference, action or framing when a test fails. ## How are wardrobe and product states tracked? Keep a state list for each scene. The A study prepared athletic and street outfits, dry/wet character states, headphones and supporting props. A dry-to-wet transition is a continuity decision: it should happen at the agreed beat and should not migrate into earlier cuts. Check strap count, colors, product position and the hand used for a repeated gesture. The [source wardrobe demonstration](https://www.youtube.com/watch?v=3rDs6FhFoUQ&t=708s) separates the original face region from edited clothing. It is a method reference, not proof of automatic masking or exact pixel preservation in ChatCut. The A run also treated the tube dancer as cast rather than following the intended prop-candidate count. State preparation needs explicit roles as well as images. ## How are prompts and takes reviewed? 1 #### Map the story moments Record scene order, framing, product visibility and gesture joins before compiling generation prompts. 2 #### Bind approved references Attach stable product, cast, location and state references to each scene; check the intended action in motion. 3 #### Generate bounded batches Approve a live quote and keep settings consistent while comparing takes for the same prompt. 4 #### Harvest by moment Inspect usable source windows for each required action, identity, props and framing; keep a record of missing coverage. 5 #### Finish and verify delivery Assemble selected windows, inspect joins, place music and dialogue, then review the actual export for picture and sound. Nine prompt blocks and 33 planned CUTs preceded A’s production takes. The report still found missing source framing families and inserts. Compare every required moment to an actual source window rather than treating the compiled count as a quality check. Kitchen and stadium raw selections showed some consistent appearances, but final blind identity and product-crop checks were not completed. Street samples showed phone props and a duplicated dancer; the boss lock lacked required glasses. Product geometry, costume continuity and recognizable faces need separate review. A successful generation job is only a candidate take. ## Which stages were actually completed? A completed asset preparation, a song and derived audio excerpts, prompt compilation, kitchen and stadium production, and four street outputs. It stopped after the last street import when its project stopped responding. The report did not establish the originating runtime cause. Street review and locking remained unfinished. Whole-take kitchen and stadium choices were not harvested scene edits. There was no office or packshot footage and no assembled timeline or export. Read the [source-attributed partial case report](https://pr-693-preview.chatcut.dev/blog/recreate-higgsfield-headphones-ad) for the historical media ledger and precise stopping point. ## What remains before a finished export? The intended deliverable was a 48–60-second five-scene ad. A continuation would first review existing street takes, repair missing casting and moment coverage, then generate office and packshot material. Selected windows would become scene edits and an assembled timeline, with matching gesture direction and a deliberate final product view. Music was generated but never placed. The continuous bed, its entry, dialogue ducking and return, unwanted score and final loudness still needed listening checks. A transcript cannot establish a clean mix. Verify actual export dimensions, frame rate, duration, joins, identity and product detail alongside sound. The historical run proves neither finished delivery nor advertising effectiveness. Use [product-image preparation](https://pr-693-preview.chatcut.dev/use-cases/product-ads) and [generation references](https://pr-693-preview.chatcut.dev/docs/generation-references) for inputs. Consult [credits](https://pr-693-preview.chatcut.dev/docs/credits-policy) before paid batches and [workflow failures](https://pr-693-preview.chatcut.dev/docs/workflow-failures) when review or project recovery blocks progress. ## Sources - [Higgsfield headphones-ad tutorial](https://www.youtube.com/watch?v=3rDs6FhFoUQ&t=708s), reviewed October 27, 2026: attributed method and wardrobe reference; no preset/project import or exact recreation claim. - **ChatCut A Ultrarealistic Headphones Ad acceptance report, October 2026**, reviewed October 28, 2026: historical local assets, tests, prompts and 25 production takes; acceptance failed with no office, packshot, assembly, mix or export. Internal report and media are not publicly hosted.