# V9 — complete

9 verified and explicitly reviewed planned outputs; 0 gated cells skipped; 0 terminal job exceptions. Historical setup observations and superseded attempts are preserved in SETUP_HISTORY.md and LEDGER.

## Final reviewed results

All nine planned interpolation results completed and were reviewed: fixed V1 walk, dance and realistic sword-duel sources through RIFE4.26, FILM and GIMM-VFI, final60fps at1280x720. The16fps walk/dance sources use integer4x interpolation to64fps; the24fps duel uses3x to72fps. A deterministic fps filter then selects/repeats frames to60fps and preserves duration to within one output-frame interval. Final walk/dance files have304frames/5.066667s; duel302frames/5.033333s, compared with source5.0625s/5.041667s. These are approximate original durations, not exact continuous-time equivalence. Every final is H264/yuv420p/faststart and fully decode-verified.

Each review deliberately samples20 output times away from native source-frame ticks, with three additional full-resolution inserted-frame checks and recorded frame-index provenance. The walk keeps face, jacket and silhouette coherent across all three methods; sampled differences are small and do not establish a decisive quality winner. Dance hands/forearms retain soft or translucent trails. Duel blades show soft/doubled edges and the moving fighter is blurred; a nearby native source frame confirms similar defects already exist before interpolation. None of the methods repairs source anatomy, choreography or motion blur. Stable sampled faces do not prove every intervening frame is artifact-free. No true intermediate frames or full perceptual playback benchmark are available, so avoid numerical accuracy or universally smoother-motion claims.

Measured request wall seconds (walk/dance/duel): RIFE254.70/13.45/22.45; FILM18.59/16.99/18.57; GIMM71.88/72.66/77.62. RIFE first-walk startup is a large outlier; subsequent jobs are much faster. Loading/cache state differs and these are one-pass measurements, so the first RIFE request must not be generalized as steady-state speed. GIMM was consistently slower in this tested wrapper without a clear sampled visual gain. Final60fps filtering/verification adds about2.8-4.3s per clip separately from those request times. Device peak includes resident model cache and is not isolated interpolation-model memory.

All final exports use the same lossless GPU1 NVENC finishing to avoid an extra lossy-export difference; upstream workflow encoding may already be lossy. Integer interpolation followed by64/72-to60 frame selection creates nonuniform source-time spacing, so compare the final60fps deliverables with that cadence limitation in mind. Recommended write-up: show all three methods on dance hands and duel swords, include the native source, and emphasize that interpolation cannot salvage flawed generated motion. Public comparison videos are viewing aids; originals and inserted-frame QA retain the evidence used for scoring.


Measured device VRAM includes resident models/cache. Request wall and Comfy execution times are distinct; post-resample time is additional. Visual reviews are sampled temporal contacts, not full-resolution playback or phoneme/listening analysis.

| Test | Model | Wall seconds | Execution seconds | Peak device GiB | Scores |
|---|---|---:|---:|---:|---|
| walk | rife | 254.7 | 253.881 | 43.68 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4} |
| walk | film | 18.59 | 17.864 | 30.25 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4} |
| walk | gimm | 71.88 | 70.963 | 29.7 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4} |
| dance | rife | 13.45 | 12.722 | 2.39 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 3, "style": 4} |
| dance | film | 16.99 | 16.236 | 5.61 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 3, "style": 4} |
| dance | gimm | 72.66 | 71.865 | 14.17 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 3, "style": 4} |
| duel | rife | 22.45 | 21.624 | 2.11 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 3, "style": 4} |
| duel | film | 18.57 | 17.741 | 5.21 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 3, "style": 4} |
| duel | gimm | 77.62 | 76.845 | 13.99 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 3, "style": 4} |

## Per-clip settings and review notes

- **walk / rife**: fp16 inference selected by native interpolation loader on CUDA capability>=8; checkpoint dtype may differ. source=V1_S2_M1_wan22_lightning_24601.mp4; 1280x720; input 16.0fps; integer interpolation x4 to 64.0fps; deterministic final60fps duration5.0625; seed24601. Reviewed 20 inserted-time frames and full-resolution output frames 36, 163, 268, away from native source ticks. Walking silhouette, face and jacket remain coherent; no conspicuous doubled limb or face in these samples. Mild hand-edge softness and inherited unreadable street/plate lettering persist. Output is 60fps after x4 64fps interpolation and deterministic resampling; no ground-truth intermediate frames, so these samples do not prove uniformly artifact-free motion.
- **walk / film**: fp16 inference selected by native interpolation loader on CUDA capability>=8; checkpoint dtype may differ. source=V1_S2_M1_wan22_lightning_24601.mp4; 1280x720; input 16.0fps; integer interpolation x4 to 64.0fps; deterministic final60fps duration5.0625; seed24601. Reviewed20 inserted-time frames and full-resolution output frames36,163,268. Face, hair, jacket and alternating walking silhouette remain coherent, without conspicuous doubled limbs in these samples. Mild hand-edge/background softness and inherited unreadable street/plate lettering remain. At these sampled times FILM and RIFE are visually close; no decisive quality winner established.60fps final is deterministically resampled from x4 interpolation at64fps; no intermediate ground truth is available.
- **walk / gimm**: fp32 checkpoint and explicitly selected fp32 inference. source=V1_S2_M1_wan22_lightning_24601.mp4; 1280x720; input 16.0fps; integer interpolation x4 to 64.0fps; deterministic final60fps duration5.0625; seed24601. Reviewed20 inserted-time frames and full-resolution frames36,163,268. Walking silhouette, face, jacket and background remain coherent without a conspicuous doubled limb/face in these samples. Mild soft hand edges and inherited unreadable background lettering persist. Similar sampled quality to RIFE/FILM on this relatively simple walk; no decisive visual ranking. Native16fps is interpolated x4 then deterministically resampled64-to60fps, so final cadence includes frame selection; no ground-truth intermediate images.
- **dance / rife**: fp16 inference selected by native interpolation loader on CUDA capability>=8; checkpoint dtype may differ. source=V1_M2_wan22_lightning_24601.mp4; 1280x720; input 16.0fps; integer interpolation x4 to 64.0fps; deterministic final60fps duration5.0625; seed24601. Reviewed20 inserted-time frames and full-resolution frames36,163,268. Dance poses and leg crossings remain readable with a stable face/body silhouette, but moving hands and forearms have soft translucent trails, especially near the chest. The source already contains motion blur; without ground-truth intermediates this does not isolate interpolation error. No conspicuous extra limb or doubled face in sampled output.16fps source interpolated x4 to64 then final60fps frame selection; source flaws are not repaired.
- **dance / film**: fp16 inference selected by native interpolation loader on CUDA capability>=8; checkpoint dtype may differ. source=V1_M2_wan22_lightning_24601.mp4; 1280x720; input 16.0fps; integer interpolation x4 to 64.0fps; deterministic final60fps duration5.0625; seed24601. Reviewed20 inserted-time frames and full-resolution frames36,163,268. Stable face and readable dance/leg-crossing sequence, with soft/smeared moving hands and forearms at faster gestures. Source motion blur remains a confound; no ground-truth intermediates. Sampled quality is close to RIFE, without a decisive improvement in fingers or motion edges. No conspicuous extra limb or doubled face in these samples.64fps integer interpolation is deterministically resampled to60fps.
- **dance / gimm**: fp32 checkpoint and explicitly selected fp32 inference. source=V1_M2_wan22_lightning_24601.mp4; 1280x720; input 16.0fps; integer interpolation x4 to 64.0fps; deterministic final60fps duration5.0625; seed24601. Reviewed20 inserted-time frames and full-resolution frames36,163,268. Face, torso and leg-crossing shapes remain coherent, but fast hand/forearm motion is smeared and sometimes translucent near the chest. Source blur limits attribution; no ground-truth intermediate frames. Similar overall sampled quality to RIFE/FILM, with no clear winner on this dance. No conspicuous extra limb or doubled face in samples; interpolation preserves rather than repairs source limitations.
- **duel / rife**: fp16 inference selected by native interpolation loader on CUDA capability>=8; checkpoint dtype may differ. source=V1_F1_real_hunyuan15_24601.mp4; 1280x720; input 24.0fps; integer interpolation x3 to 72.0fps; deterministic final60fps duration5.041666666666667; seed24601. Reviewed20 inserted-time frames and full-resolution frames36,161,261, plus nearby native source frame at2.666667s. Two fighters and crossing swords stay identifiable; sequence is a modest exchange followed by separation rather than complex choreography. Blades show soft/doubled edges and the advancing fighter is blurred around2.68s, also visible in the nearby source frame, so not attributable solely to interpolation. No conspicuous new extra limb or doubled face in samples.24fps source interpolated x3 to72fps then deterministically selected to60fps; sampled RIFE/FILM quality is close, without ground-truth intermediates.
- **duel / film**: fp16 inference selected by native interpolation loader on CUDA capability>=8; checkpoint dtype may differ. source=V1_F1_real_hunyuan15_24601.mp4; 1280x720; input 24.0fps; integer interpolation x3 to 72.0fps; deterministic final60fps duration5.041666666666667; seed24601. Reviewed20 inserted-time frames and full-resolution frames36,161,261, plus nearby native source frame at2.666667s. Two fighters and crossing swords stay identifiable; sequence is a modest exchange followed by separation rather than complex choreography. Blades show soft/doubled edges and the advancing fighter is blurred around2.68s, also visible in the nearby source frame, so not attributable solely to interpolation. No conspicuous new extra limb or doubled face in samples.24fps source interpolated x3 to72fps then deterministically selected to60fps; sampled RIFE/FILM quality is close, without ground-truth intermediates.
- **duel / gimm**: fp32 checkpoint and explicitly selected fp32 inference. source=V1_F1_real_hunyuan15_24601.mp4; 1280x720; input 24.0fps; integer interpolation x3 to 72.0fps; deterministic final60fps duration5.041666666666667; seed24601. Reviewed20 inserted-time frames and full-resolution frames36,161,261 with nearby source-frame comparison. Fighters, crossed blades and later separation remain coherent. Soft/doubled blade edges and blurred moving body/hand regions persist, including around2.68s where similar blur exists in the source. No conspicuous new extra limb or doubled face in samples. No decisive sampled quality advantage over RIFE/FILM;72fps integer interpolation is deterministically resampled to60fps, with no ground-truth intermediate frames.