# H3 Hyperflow retention-analysis A/B

These ten pairs show no consistent visual advantage from the extra retention block or the appropriate mode markers. Explicit wardrobe, color, daylight and environment changes happen under both marker choices; selective lighting fails in both. One notable exception favors A: the environment test has one dog in A and two in B. This supports the practical importance of prose, not a claim that markers are ignored. Results are single-seed observations with this Hyperflow setup.

## Setup

20 clips / 10 paired comparisons; five shared references; 10-second viewing copies; 1344×768; full Hyperflow, trained 8-step schedule, Euler, CFG 1. Seeds and all other execution inputs are fixed within each pair. Native 243-frame originals are retained.

Study 1 (01–05): A includes the fully_preserved retention block; B omits the whole block. Study 2 (06–10): A uses the appropriate markers; B substitutes fully_preserved while keeping all explanatory prose identical.

## Findings

| Pair | Result |
| --- | --- |
| 01_camera_handoff | Similar outcome; no clear block advantage |
| 02_dog_portrait | Similar outcome; shared layout failure |
| 03_camera_orbit | Camera movement works in both |
| 04_occlusion | Similar outcome; limited occlusion stress |
| 05_dialogue | Visually similar; audio not assessed |
| 06_wardrobe_change | Requested clothing changes happen in both |
| 07_prop_recolor | Requested recoloring happens in both |
| 08_daylight_change | Relighting happens in both; shared object error |
| 09_lighting_transfer | Selective transfer fails similarly in both |
| 10_environment_inspiration | Environment changes in both; A keeps one dog, B duplicates it |

### Camera handoff: identity and object retention

Both versions preserve the recognizable people, dark clothing, café fixtures, golden dog and black camera while completing the handoff and final placement. Framing and hand positions differ, but neither shows a clear reference-retention advantage in the inspected frames. Both deviate from the requested standing adults / puppy-on-floor arrangement.

2.5–5s: handoff; 9.5s: camera resting on table. Compare both rows of the contact sheet.

[Paired frames](pairs/01_camera_handoff/contact.jpg)

### Photographing the puppy: simultaneous reference binding

Both versions show the woman crouching, raising the camera to photograph the dog, then lowering it to inspect the display. Clothing and café remain similar. Both put the dog on the tabletop instead of its requested floor mat; retaining the block does not fix that instruction failure.

5s: camera at the woman's eye; 9.5s: lowered camera. The dog is on the table throughout the sampled sequence.

[Paired frames](pairs/02_dog_portrait/contact.jpg)

### Continuous camera orbit without freezing reference poses

Both versions change viewpoint substantially while retaining the people, clothing, café and camera. The fully_preserved block does not freeze the shot. Camera handling, dog position and foreground occlusions differ, but there is no clear overall retention winner from the inspected frames.

Compare 0.5s with 9.5s: the table, people and wall are seen from different angles in both versions.

[Paired frames](pairs/03_camera_orbit/contact.jpg)

### Short occlusion and reappearance of faces and prop

Both show the man moving from the foreground toward the woman, followed by a shared look at the camera. Their appearances remain coherent in the sampled sequence. The crossing is not a clean, prolonged face occlusion, so this pair provides only weak evidence about recovery from occlusion.

The 2-fps timelines show the foreground approach and later close two-shot. Do not treat this as a strong occlusion benchmark.

[Paired frames](pairs/04_occlusion/contact.jpg)

### Dialogue, acting and multi-reference continuity

Both retain the referenced people and café through a very similar conversational composition. The woman raises the camera near the end in both. Both keep the dog on furniture instead of the floor mat. Spoken wording, voice quality and lip sync have not been evaluated; both audio tracks remain available for listening.

0.5–9.5s: broadly matched staging; 9.5s: camera raised in both.

[Paired frames](pairs/05_dialogue/contact.jpg)

### Wardrobe replacement: partially_preserved versus fully_preserved

A (partially_preserved) and B (fully_preserved) both give the woman the yellow rain jacket and the man the teal pullover, with recognizable faces and hair. Fully_preserved does not prevent the explicit wardrobe replacement in this pair. There are staging differences, but no clear benefit from the appropriate marker on the tested clothing task.

The yellow jacket and teal top are present from 0.5s through 9.5s in both versions.

[Paired frames](pairs/06_wardrobe_change/contact.jpg)

### Camera recoloring: partially_preserved versus fully_preserved

Both variants produce a red camera body with a black lens and carry that color through the handoff to the closing view. The fully_preserved marker does not block the explicitly requested recoloring. The overall camera shape is retained, but this is not a test of exact product geometry or branding.

Red camera body is visible at 0.5s, during handling, and clearly on the tabletop at 9.5s in both versions.

[Paired frames](pairs/07_prop_recolor/contact.jpg)

### Environmental relighting: partially_preserved versus fully_preserved

Both versions replace the moody café exposure with bright daylight while retaining its recognizable architecture and fixtures; the CAFE sign appears unlit in both. Both also generate two cameras despite the instruction to keep one. The appropriate retention marker does not resolve that duplication.

0.5s and 5s clearly show one camera with each person in both variants. Compare the bright left-side window and unlit lettering.

[Paired frames](pairs/08_daylight_change/contact.jpg)

### Lighting transfer: attribute_transfer versus fully_preserved

Both versions put blue light on the man, but both also leave strong blue light on the woman, contrary to the explicit request to keep her warm-neutral. Attribute_transfer does not provide a clear advantage over fully_preserved for this selective lighting assignment in the inspected frames.

0.5s and 9.5s: the woman's face and dark top remain visibly blue-lit in both rows, while the man also receives cool lighting.

[Paired frames](pairs/09_lighting_transfer/contact.jpg)

### Broad environment inspiration: weak_reference versus fully_preserved

Both A (weak_reference) and B (fully_preserved) replace the café fixtures with the explicitly requested white arches, pale stone and green plants. The mode marker does not block that substantial environment change. A has one dog, whereas B generates two dogs throughout the sampled sequence. A is preferable for the requested subject count in this pair; without repeat controls or more seeds, that does not establish a general causal benefit of weak_reference. Both put the dog(s) on furniture rather than the requested floor mat.

0.5–9.5s: both show the new lounge without the original CAFE sign, menus or green pendant lights. Every sampled B frame shows two dogs; A shows one.

[Paired frames](pairs/10_environment_inspiration/contact.jpg)

## Review limits

Unblinded visual inspection of paired full-clip contact sheets at 0.5, 2.5, 5, 7.5 and 9.5 seconds, supplemented by 2-fps timelines where needed. No audio listening/transcription or frame-perfect smoothness assessment was performed. No automatic identity or quality score was used.

One seed per scenario, five scenarios per claim; one shared five-image reference set; exploratory rather than a population-level conclusion. Marker-B deliberately conflicts with explicit requested changes. No deterministic repeat control was added. Different outputs alone do not establish a quality advantage.

Distinct outputs alone do not demonstrate usefulness; similar outcomes do not prove a token is ignored. These observations apply to this Hyperflow configuration, prompt set and reference set, not all H3 inference modes.