Keep real motion. Change one thing.

• Our process moved toward visual anchors and real-footage references. Historical reviews identified repeated realism failures; they were not a controlled comparison of text-to-video versus image-to-video.
• July 13: a 46-clip program led to a blind review that overturned nine earlier “clean” judgments. Matched before/after stills, restrained animation and hard cuts became the repair recipes. A later handoff records 26 assets graduated using already-paid material.
• August 26 project, captured September 3: we studied real foundation footage, generated new stills from written shot descriptions, then animated those stills. The formula explicitly kept source clips and extracted frames out of model inputs.
• Evidence of improvement: the later three-clip audit scored the crack-and-pier scene 1/5 for AI obviousness. The other two remained 3/5 because support geometry or hydraulic details looked wrong when paused. Image-to-video helped; it did not fix an incorrect starting image.
• Genjutsu is the new proposed branch. No prior Genjutsu run was found in the searched handoffs or production records. Its suitability here remains untested.

• New recommendation: preserve more of the original shot as technical complexity rises. Genjutsu preserves source motion by design; that is not a guarantee of correct tools or building physics.
• Do not turn concrete pouring into concrete leveling by changing the prompt. Keep each source within its actual trade and process.

Library register is the current index. Historical snapshot checked September 9, 2026; files were verified on Akash’s machine, not on the builder’s PC:

• There are 125 total rows, including three screen-flow entries outside these five niches. Counts include AI clips, rejected candidates and historical entries; they are not counts of fresh real footage.
• Real vs. generated: use the source and rights fields, not just a folder name. Avoid stacking another transformation on an AI clip with known defects.
• Approval and ownership are separate. The register contains 46 “no license or ownership record found” entries and 19 competitor Ad Library entries with no license record. Preserve those labels. Removing branding or using Genjutsu does not establish source rights.
• Direct-input route: choose owned or suitably licensed footage, or a source with documented clearance for this reuse/upload. Where that is unresolved, use the existing source-free approach: study the shot, write its structure, generate a new still. Do not silently upload the borrowed source.
• Pick one continuous 4–6-second action with stable framing. Exclude dissolves, speed ramps, burnt captions, blur masks, cropped hands and heavy occlusion. Start with the unretimed original when available.
• Match the intended crop before generation. A wide, low-resolution reference does not become a detailed vertical close-up merely by selecting 1080p.

• Verified September 9: both modes appear in our authenticated Higgsfield catalog. The live schema requires exactly one video and at least one reference image, offers 480p / 720p / 1080p, and defaults to 720p.
• The official launch guide says 4–30-second videos and up to 30 images; the landing FAQ says 3–30 seconds and up to 40. Use the shared safe range—4–30 seconds, at most 30 images—and follow current generator validation. Our recommended test needs just one image.
• Higgsfield says Object Swap changes a selected element; Motion Transfer rebuilds around source movement. Treat preservation claims as behavior to verify.

1. Choose the source: a cleared, continuous 4–6-second clip. Write one sentence naming the action that must survive.
2. Make one reference image: for Object Swap, clearly show only the intended replacement; for Motion Transfer, use a checked full-scene image with compatible camera height, layout and subject scale. Prefer our own material.
3. Open Video → Genjutsu: choose Object Swap for the first test. Upload the source video and the reference.
4. Describe the exact change: name its screen position or clothing and explicitly preserve everything important. Avoid cinematic/style presets for documentary jobsite footage.
5. Choose resolution: 720p for the first inspectable test; 1080p for the final if available. View the exact run cost before submitting. A fresh final render needs another full check.
6. Review beside the original: did the intended detail change, and did the repair action, object count, camera timing and shadows survive?
7. Save source, reference, prompt and output together: record the mode and exact settings, then run the checks below before admitting it to the library.

Object Swap example — wardrobe only

Change only the small background worker’s shirt to the plain charcoal work shirt in the reference image. Preserve the person’s pose, size and movement. Keep the house, soil, tools, safety equipment, hoses, repair hardware, shadows, camera path and timing unchanged. Match the existing daylight, fabric wear and phone-camera texture. No new objects, text or logos.

Motion Transfer example — simple crawl-space pan

Use the source video’s camera movement and timing for a new ordinary residential crawl-space shot matching the reference image. Keep the same low camera height, ceiling clearance and support-column layout. The liner, seams and columns are already installed and stay fixed. Only the camera moves gently. Natural work light and imperfect phone framing; no installation, material transformation, extra columns, standing worker, text or logos.

1. Study one real shot: record the setting, camera height, connected equipment and one useful action. Where source upload is not cleared, keep the footage and extracted frame analysis-only, as our August formula required.
2. Create the starting image: make a new documentary still from that description, or use an authorized real frame. Check the entire image before animation—especially how the load is supported and where hoses connect.
3. Animate only what is easy: gentle lateral camera drift, small worker movement or a finished-room pan. Our later foundation run used Seedance 2.0 with a supplied generated starting image; this is a historical recipe, not a claim that it remains the best model.
4. For before/after: create the after still using the before still as a reference. Lock camera, property, perspective and lighting. Verify both states separately. Edit with a clear hard cut for concrete curing, slab repair or hardware installation.
5. Use workers selectively: liner rollout and membrane handling improved with visible agency in July. Complex hardware installation added anatomy and continuity failures; keep the equipment installed from frame one.

Starting-image template — foundation repair

Vertical 9:16 documentary phone photograph beside an ordinary gray-stucco house. One open excavation exposes a concrete footing resting on an already-installed steel bracket and pier assembly, with continuous hoses leading out of frame. A small back-facing technician holds an unmarked straightedge near an existing wall crack. Natural shade, worn tools, believable dirt and scale. Show clearly how the footing is supported. No unsupported masonry, invented extra hardware, close face, readable text, logo or watermark. Create a new scene; do not reproduce the reference property or person.

Animation template — one modest action

Use this approved still as the first frame. One five-second documentary phone shot. The camera makes a restrained sideways movement, then holds. The technician shifts weight slightly. Keep the crack, wall, footing, bracket, pier, bolts and hose connections unchanged in shape, number and position. No house lift, crack closure, new objects, installation, morphing, dramatic camera move, readable text or logos.

• Stop rule: change the failed source, still or prompt before retrying. After two paid revisions for the same shot, park it. If over half a batch shares one defect, fix the recipe before generating more.
• Do not conceal failures with speed, blur or text. The old “feed-only” label means a known pause risk. This guide recommends keeping those out of a reusable realistic library until fixed; it does not change their existing approval records.
• Register descendants: source clip ID and hash, selected time range, source rights, reference images, model/mode, prompt, settings, job ID, credits, output path, reviewer, timestamped findings and approval. Keep originals intact.
• Generated footage remains illustrative background, never evidence of a real client job, measured lift or repair result.

★ Recommended next production step: compare three short variations against unchanged originals. Akash reviews the side-by-side results for about 10 minutes after rendering. Preparing this reference generated no video and spent no credits.

• Use unchanged real clips as the comparison baseline. Keep the test duration and viewing crop consistent.
• Only expand a route after it passes the full checks. A tiny test demonstrates feasibility; it does not establish a reliable acceptance rate.
• Before any paid generation, confirm the exact job price and the agreed allowance with Akash. Account balances and old session budgets are not reusable spending authorization.
• The official pricing snapshot lists 15 seconds at 40 / 104 / 144 credits for 480p / 720p / 1080p. Do not prorate these into a promised five-second price. Historical July rates are obsolete for planning this run. Pricing source.

Download the four prompt templates · Find source clips in the library directory

1. Save the chosen source in inputs\source.mp4, the checked image in inputs\approved-start.png, and generated output in work\generated.mp4. Keep the original file unchanged.
2. In PowerShell, create the scan folder and extract frames. Sampling helps find faults; it does not replace full playback or frame-stepping suspicious moments.

New-Item -ItemType Directory -Force "work\realism-frames"
ffmpeg -i "work\generated.mp4" -vf "fps=2,scale=720:-2" "work\realism-frames\frame-%04d.jpg"
ffprobe -v error -show_entries stream=codec_type,codec_name,width,height -show_entries format=duration -of json "work\generated.mp4"

• The old automatic text scanner uses Apple Vision and is Mac-only. On Windows, review the frames and complete playback yourself. Mark automated OCR as not run unless you actually ran another validated scanner. Never turn a failed or missing scan into zero detections.
• Record source ID, source rights, selected range, reference images, model/mode, prompt, settings, job ID, credits, tell/timestamp, reviewer and Akash’s acceptance. The supplied realism CSV carries these fields.
• Silent source clips should have no audio stream. Finished ads assembled in E must retain the locked narration. Do not apply the silent-library rule to the final voiced ad.
• For comparison, label each output “real,” “AI-generated” or “AI-edited.” A new scene modeled after PROVEN-FR32 is not that real master and does not inherit its ad results.
• Historical evidence: July 13 production laws and graduation/reshoot records; August 26 foundation shot formula and source-model blind audits; September 3 session capture. The 46-clip review and nine overturned judgments are recorded audit findings, not measured conversion lifts.
• Current product evidence: Higgsfield’s August 31 launch guide and feature page, checked September 9. Our authenticated catalog also exposed both modes. In the CLI, Object Swap is hf_mult_replace_object; Motion Transfer is hf_mult_motion_control. Both schemas required exactly one video and at least one image, with 480p, 720p or 1080p output. Discover settings again before execution.

higgsfield model get hf_mult_replace_object --json
higgsfield model get hf_mult_motion_control --json

No new Windows practice render or Genjutsu output was produced for this guide update. The builder still needs app installation, library/account access and one successful Windows practice render.