# V3 + V4 — final results

Coverage: 22 primary videos, six planned LTX-2.5 slots skipped because access is gated, and 5 retained configuration diagnostics. All 27 actual videos passed ffprobe and full ffmpeg decode and received explicit visual review. Outputs are H.264, yuv420p, with faststart.

The four Wan FLF cases use 1280×720, 81 frames at 16 fps. Pose uses 20 steps/CFG 3.5; day-to-night, match-cut and loop use four-step Lightning/CFG 1. Reference tests use the same Mara/Kite stills, shot prompts and seeds across VACE, HuMo and SkyReels. VACE uses 20 steps/CFG 3.5, 81 frames at 16 fps; HuMo uses six steps/CFG 1, 125 frames at 25 fps, reference only and no audio; SkyReels uses 20 steps, text CFG 5/image CFG 2, 121 requested frames; the installed Phantom path removes the reference latent and returns 117 actual frames at 24 fps (4.875 seconds). No output padding was added.

VACE configuration decision: the initial four-step recipe produced weak walking, duplicated close-up figures and a static sword stance. The full-step close-up removed those severe overlays; its existing control artifact is the canonical Mara talk result. A second native adaptation of the published four-step settings produced colored noise. These four fast outputs remain diagnostics, not replacements for the six full-step reference results. This does not establish that the upstream custom-node workflow fails. See workflows/reference/vace_configuration_sources.md for the implementation differences.

HuMo runtime correction: the first reference-only attempt failed before producing video because the installed no-audio node supplied one extra temporal group. An additive adapter trims only zero-audio conditioning to the latent length before the model adds its reference padding. Text and reference conditioning remain intact; no speech audio is introduced. The failed attempt and correction are documented in workflows/reference/humo_no_audio_fix.md and state/gpu0_failures. This failed execution is separate from the actual videos.

SkyReels attention profile: {"profile": "sage2", "model": "skyreelsv3-r2v-sage2-fp8", "attention_mode": "sageattn", "sage_version": "2.2.0", "reason": "Matching720p control preserves comparable walk/face quality while sampling is1.62x faster. Output is not identical: new garbled background sign is flagged. All six primary SkyReels cases use this profile; first SDPA run retained as diagnostic.", "sdpa_total_s": 2525.147, "sage_total_s": 1645.321, "total_speed_ratio": 1.5347442839421608, "sampling_s_per_step": {"sdpa": 114.6, "sage2": 70.8}, "reference_control": "V3V4_Mara_walk_sage_control_skyreelsv3-r2v-sage2-fp8_44100", "memory_note": "Sampled total residency differs; do not attribute peak difference solely to attention."}. Any attention comparison outputs remain separately identified; do not compare cold-load timings as warm throughput. Kernel microbenchmarks alone do not establish video speed or visual equivalence. A labeled same-seed viewing aid is in comparisons/skyreels_sdpa_sage2_same_seed.mp4.

FLF findings: pose and lighting endpoints work well. The requested match-cut becomes a dissolve with ghosted sets. Identical loop input endpoints do not guarantee a seamless output: first/last RGB MAE is 5.50 versus mean adjacent-frame MAE 0.358 at 320×180. The labeled four-case comparison is in comparisons/wan_flf_four_cases.mp4.

Review method: nine evenly spaced frames including both endpoints, with notes distinguishing prompt adherence, identity and visible defects. Scores do not establish phoneme-level lip synchronization; these talk prompts are silent. Timing is ComfyUI execution time and can include cold loading; peak memory is sampled total GPU0 residency, not isolated model allocation.

| Test | Role | Settings | Seconds | Peak GiB | Scores | Review |
|---|---|---|---:|---:|---|---|
| V3V4_Kite_duel_humo17-fp8-lightning_44105 | primary | 6 steps,cfg1,1280x720,125 frames,25fps; reference-only, no audio,seed44105 | 318.57 | 44.43 | {"adherence": 5, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_identity": 5} | Reviewed nine evenly spaced frames. Single recognizable Kite takes a wide two-handed sword stance, lowers the weapon and holds it; full body and both feet remain visible. Silver hair and navy/orange outfit are stable. Blade curvature/length varies mildly during the lowering motion, and hand-grip detail is limited at this scale. A high camera angle and busy gravel texture dominate the setting. Flags: {"artifacts": "minor blade geometry variation; busy ground texture", "text_garbling": false, "watermark": false} |
| V3V4_Kite_duel_skyreelsv3-r2v-sage2-fp8_44105 | primary | 20 steps,textcfg5,imagecfg2,1280x720,121 frames,24fps; fp8 scaled fast; attention=sage2,seed44105; actual_output_frames=117 | 1441.06 | 30.19 | {"adherence": 2, "motion": 2, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_identity": 4} | Reviewed nine evenly spaced frames. Recognizable silver-haired anime Kite in navy/orange clothing holds a sword lowered beside his leg, with small head turns, blinking and slight upper-body shifts. No clear transition into a guarded duel stance or controlled footwork; lower legs and feet are cropped. Strong radial speed lines dominate an angular corridor backdrop despite restrained subject motion. Face and body remain coherent, with limited hand/blade detail and mild clothing drift. Flags: {"artifacts": "excessive speed-line background; minor clothing drift", "text_garbling": false, "watermark": false} |
| V3V4_Kite_duel_wan22-vace14-fp8_44105 | primary | 20 steps,cfg3.5,1280x720,81 frames,16fps,seed44105 | 999.33 | 72.86 | {"adherence": 4, "motion": 3, "anatomy_face": 3, "temporal": 3, "style": 4, "reference_identity": 4} | Reviewed nine evenly spaced frames. Kite turns away and swings a sword into a wide ready stance, with clear foot movement. Outfit and silver hair persist. The blade curvature/length changes, and legs become visibly soft/blurred during movement; grip details are ambiguous. The back-facing final pose limits face assessment. Flags: {"artifacts": "motion blur/soft legs and variable sword geometry", "text_garbling": false, "watermark": false} |
| V3V4_Kite_talk_humo17-fp8-lightning_44104 | primary | 6 steps,cfg1,1280x720,125 frames,25fps; reference-only, no audio,seed44104 | 320.22 | 44.43 | {"adherence": 5, "motion": 4, "anatomy_face": 5, "temporal": 5, "style": 5, "reference_identity": 5} | Reviewed nine evenly spaced frames. A single chest-up anime Kite speaks silently with clear mouth changes, a blink and subtle head/shoulder motion. Silver hair, amber eyes and navy/orange jacket remain consistent; no extra person or inset. Clean facial structure and stable linework. Background changes to a warm paneled interior. No audio, so phoneme lip-sync is not assessed. Flags: {"artifacts": false, "text_garbling": false, "watermark": false} |
| V3V4_Kite_talk_skyreelsv3-r2v-sage2-fp8_44104 | primary | 20 steps,textcfg5,imagecfg2,1280x720,121 frames,24fps; fp8 scaled fast; attention=sage2,seed44104; actual_output_frames=117 | 1446.94 | 28.52 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_identity": 4} | Reviewed nine evenly spaced frames. Coherent single anime adult with silver hair, amber eyes and navy/orange jacket. Blinks, mouth movements and a gradual camera push-in support the silent talk prompt. Framing starts medium rather than close-up; background has angular studio-like panels. Minor face/hair contour drift; no duplicate face. Lip synchronization cannot be evaluated without audio. Flags: {"artifacts": "minor contour drift", "text_garbling": false, "watermark": false} Recovery: Recovered completed ComfyUI history after collector interruption. Execution timestamps retained; VRAM is maximum of available samples before interruption, with an unobserved interval before collection. Actual hardware peak may be higher. |
| V3V4_Kite_talk_wan22-vace14-fp8_44104 | primary | 20 steps,cfg3.5,1280x720,81 frames,16fps,seed44104 | 987.84 | 72.86 | {"adherence": 2, "motion": 3, "anatomy_face": 4, "temporal": 5, "style": 5, "reference_identity": 5} | Reviewed nine evenly spaced frames. One stable anime character retains the reference outfit, hair, body proportions and studio. Small mouth/head changes suggest silent speaking, but the requested close-up is ignored: the entire body remains visible throughout. Face detail is too small to judge fine mouth anatomy; no audio or lip-sync claim. Flags: {"artifacts": false, "text_garbling": false, "watermark": false} |
| V3V4_Kite_walk_humo17-fp8-lightning_44103 | primary | 6 steps,cfg1,1280x720,125 frames,25fps; reference-only, no audio,seed44103 | 318.82 | 44.92 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_identity": 4} | Reviewed nine evenly spaced frames. A recognizable single Kite with silver hair and navy/orange jacket advances along a textured city road under a high tracking camera. Arm/body changes and moving ground imply walking, but the waist-up crop prevents assessment of leg movement or foot contacts. Face and outfit remain coherent; background road textures are unusually rough. Flags: {"artifacts": "rough road texture; foot mechanics not visible", "text_garbling": false, "watermark": false} |
| V3V4_Kite_walk_skyreelsv3-r2v-sage2-fp8_44103 | primary | 20 steps,textcfg5,imagecfg2,1280x720,121 frames,24fps; fp8 scaled fast; attention=sage2,seed44103; actual_output_frames=117 | 1446.35 | 29.00 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_identity": 4} | Recognizable silver hair, amber eyes and navy/orange clothing on a city street. Upper-body bobbing and camera movement suggest walking, but legs remain outside frame so gait/foot contact cannot be assessed. Slight face and collar drift, strongly contrasted anime shading, small garbled shop signs, unprompted mouth opening late. Single coherent character without duplicate face. Flags: {"artifacts": true, "text_garbling": true, "watermark": false} |
| V3V4_Kite_walk_wan22-vace14-fp8_44103 | primary | 20 steps,cfg3.5,1280x720,81 frames,16fps,seed44103 | 989.66 | 72.86 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_identity": 4} | Reviewed nine evenly spaced frames. Silver hair, navy/orange jacket and one coherent anime face persist; a city-street setting appears. Motion is restrained, with head turns and slight body bobbing while arms stay near the jacket; the thigh-up crop hides foot contacts, so natural walking physics cannot be confirmed. Detailed near-photographic background mixes with the anime figure. Flags: {"artifacts": "limited visible gait; feet outside frame", "text_garbling": false, "watermark": false} |
| V3V4_Mara_duel_humo17-fp8-lightning_44102 | primary | 6 steps,cfg1,1280x720,125 frames,25fps; reference-only, no audio,seed44102 | 318.74 | 44.43 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 5, "reference_identity": 4} | Reviewed nine evenly spaced frames. Coherent single Mara in a wide sword-ready stance, with clear arm swings and torso turns; teal jacket, trousers and boots remain. Sword stays recognizable across the sequence, with blade blur and occasional uncertain grip during swings. Footwork is limited and the lower feet are partly cropped. Background becomes an indoor practice space. No duplicated person. Flags: {"artifacts": "blade motion blur and minor grip ambiguity", "text_garbling": false, "watermark": false} |
| V3V4_Mara_duel_skyreelsv3-r2v-sage2-fp8_44102 | primary | 20 steps,textcfg5,imagecfg2,1280x720,121 frames,24fps; fp8 scaled fast; attention=sage2,seed44102; actual_output_frames=117 | 1447.76 | 27.64 | {"adherence": 2, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 3, "reference_identity": 4} | Walks and turns away across a studio with sword lowered, missing the requested ready stance; feet intermittently cropped. Coherent body and face with mild identity drift, harsh contrast and a changing/darkening sword blade. Studio equipment remains visible. Flags: {"artifacts": true, "text_garbling": false, "watermark": false} |
| V3V4_Mara_duel_wan22-vace14-fp8_44102 | primary | 20 steps,cfg3.5,1280x720,81 frames,16fps,seed44102 | 990.43 | 72.92 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 2, "style": 4, "reference_identity": 4} | Reviewed nine evenly spaced frames. Unlike the nearly static fast baseline, the full-step output raises a sword and steps into a low ready stance. Body, outfit and single face remain coherent. The long blade visibly shortens to a dagger around the middle and remains short; weapon continuity is poor. Reference studio persists. Flags: {"artifacts": "weapon length and shape change markedly", "text_garbling": false, "watermark": false} |
| V3V4_Mara_duel_wan22-vace14-lightning-fp8_44102 | diagnostic | 4 steps,cfg1,1280x720,81 frames,16fps,seed44102 | 118.45 | 73.83 | {"adherence": 1, "motion": 1, "anatomy_face": 4, "temporal": 5, "style": 4, "reference_identity": 5} | Nine samples: reference identity and clothes stay stable, but requested sword-ready stance, weapon and controlled footwork are absent. Almost static studio pose. Poor action adherence under the four-step configuration; retain rather than counting pipeline execution as creative success. Flags: {"artifacts": false, "text_garbling": false, "watermark": false} |
| V3V4_Mara_talk_full_control_wan22-vace14-fp8_44101 | primary | 20steps,cfg3.5,1280x720,81frames,16fps; same seed/prompt/reference as failed four-step close-up; no Lightning LoRA,seed=44101 | 975.28 | 87.03 | {"adherence": 4, "motion": 4, "anatomy_face": 5, "temporal": 5, "style": 5, "reference_identity": 3} | Nine samples: coherent single-person talking close-up, natural mouth changes, stable framing and plain studio background. The duplicated full-body overlays from the original four-step configuration disappear. Hair changes to a shorter bob with fringe; facial details/identity less exact than the reference. Same prompt/reference/seed;20steps,cfg3.5,noLoRA versus4steps,cfg1,LoRA. Execution975.28s. Silent reference test, so no phoneme lip-sync score. Flags: {"artifacts": false, "text_garbling": false, "watermark": false} |
| V3V4_Mara_talk_humo17-fp8-lightning_44101 | primary | 6 steps,cfg1,1280x720,125 frames,25fps; reference-only, no audio,seed44101 | 317.49 | 44.43 | {"adherence": 1, "motion": 3, "anatomy_face": 2, "temporal": 4, "style": 2, "reference_identity": 4} | Reviewed nine evenly spaced frames. The large face has coherent speech-like mouth changes and recognizable Mara features, but an unwanted second full-body copy of the reference person remains at the right throughout. This is a major duplicated-character/composite failure, not an acceptable single close-up. Individual faces are coherent; the composition is wrong. No audio or phoneme lip-sync claim. Retained as a negative benchmark result under the same configuration as the successful walk. Flags: {"artifacts": "FAIL: duplicated full-body character/inset beside close-up", "text_garbling": false, "watermark": false} |
| V3V4_Mara_talk_official_fast_control_wan22-vace14-lightning-official-fp8_44101 | diagnostic | 4steps,cfg1,shift12,uni_pc,high3+low1 (boundary0.875),1280x720,81frames,16fps; published VideoX-Fun sampler settings adapted to native nodes; same prompt/reference/seed,seed=44101 | 120.02 | 62.87 | {"adherence": 1, "motion": 1, "anatomy_face": 1, "temporal": 1, "style": 1, "reference_identity": 1} | Nine samples contain colored noise with no usable subject. The published VideoX-Fun sampler settings did not transfer successfully to this native-node split-sampler adaptation. This is a local adaptation failure, not evidence that the original custom workflow fails. No further fast-reference fixes after the two controls; adopt the validated20-step baseline. Flags: {"artifacts": true, "text_garbling": false, "watermark": false} |
| V3V4_Mara_talk_skyreelsv3-r2v-sage2-fp8_44101 | primary | 20 steps,textcfg5,imagecfg2,1280x720,121 frames,24fps; fp8 scaled fast; attention=sage2,seed44101; actual_output_frames=117 | 1444.69 | 27.91 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 2, "reference_identity": 3} | Reviewed nine evenly spaced frames. A single close-up has clear silent speaking mouth shapes, blinks and head tilts; no inset or duplicate body. The realistic reference becomes strongly stylized/plastic-looking, with heavy red lipstick, saturated smooth skin and a straighter darker bob. Teal jacket and broad facial traits remain, but realism and identity fidelity weaken. Gaze often falls slightly off camera. No audio; phoneme lip-sync not assessed. Flags: {"artifacts": "strong stylization and reference identity/hair drift", "text_garbling": false, "watermark": false} |
| V3V4_Mara_talk_wan22-vace14-lightning-fp8_44101 | diagnostic | 4 steps,cfg1,1280x720,81 frames,16fps,seed44101 | 125.43 | 73.83 | {"adherence": 1, "motion": 2, "anatomy_face": 1, "temporal": 2, "style": 2, "reference_identity": 4} | Nine samples reveal a severe composition failure: a talking close-up is overlaid with multiple full-body copies of the reference. Duplicated subjects/face composition is a fail despite recognizable Mara traits. This four-step T2V-LoRA/VACE configuration needs a full-step control on the same prompt before attributing the defect to base VACE. Flags: {"artifacts": true, "text_garbling": false, "watermark": false} |
| V3V4_Mara_walk_humo17-fp8-lightning_44100 | primary | 6 steps,cfg1,1280x720,125 frames,25fps; reference-only, no audio,seed44100 | 389.82 | 44.40 | {"adherence": 5, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 5, "reference_identity": 4} | Reviewed nine evenly spaced frames. Clear forward walk along an outdoor brick-walled street, with alternating arm/leg positions and a coherent single face. Teal jacket, charcoal trousers, brown boots and auburn bob remain. Slight facial/hair drift; feet move outside the frame as she approaches, limiting late foot-contact assessment. Corrected zero-audio conditioning works; silent output. Flags: {"artifacts": "minor face/hair drift", "text_garbling": false, "watermark": false} |
| V3V4_Mara_walk_sage_control_skyreelsv3-r2v-sage2-fp8_44100 | primary | 20steps,textcfg5,imagecfg2,1280x720,121frames,24fps; fp8 scaled fast; SageAttention2 auto dispatcher; matching SDPA baseline seed44100; actual_output_frames=117 | 1645.32 | 28.29 | {"adherence": 5, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 3, "reference_identity": 4} | Reviewed nine evenly spaced frames against matching SDPA baseline. Coherent street walk, single face, bob and outfit identity are comparable; similarly harsh contrast and blue clothing drift remain. Scene geometry differs and a new background sign has garbled letters. This is not bit-identical or artifact-free. Total1645.321s versus2525.147s SDPA, approximately70.8 versus114.6s per sampling step. Adopted for the six primary SkyReels cases with the text artifact explicitly retained. Flags: {"artifacts": "crushed shadows/clipped highlights and clothing-color drift", "text_garbling": true, "watermark": false} |
| V3V4_Mara_walk_skyreelsv3-r2v-fp8_44100 | diagnostic | 20 steps,textcfg5,imagecfg2,1280x720,121 frames,24fps; fp8 scaled fast,seed44100; actual_output_frames=117 | 2525.15 | 38.62 | {"adherence": 5, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 3, "reference_identity": 4} | Reviewed nine evenly spaced frames. Coherent single Mara walks toward camera along a real-looking city sidewalk with alternating steps and arm swing. Face, auburn bob and broad outfit identity remain, but jacket/trousers shift toward blue and the image has strongly crushed shadows and clipped highlights. Foot contacts are readable in most samples, with some crop near the end. SDPA baseline;2525.147s including cold loading, roughly114.6s per sampling step. Flags: {"artifacts": "harsh contrast/clipped tonal detail and clothing-color drift", "text_garbling": false, "watermark": false} |
| V3V4_Mara_walk_wan22-vace14-fp8_44100 | primary | 20 steps,cfg3.5,1280x720,81 frames,16fps,seed44100 | 976.64 | 77.37 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_identity": 4} | Reviewed nine evenly spaced frames. Coherent forward walk with natural arm swing and a single face; teal jacket, auburn bob and boots remain. The requested city street is replaced by the reference studio. Face changes mildly near the end; tighter crop follows the approach. Flags: {"artifacts": "minor face drift", "text_garbling": false, "watermark": false} |
| V3V4_Mara_walk_wan22-vace14-lightning-fp8_44100 | diagnostic | 4 steps,cfg1,1280x720,81 frames,16fps,seed44100 | 586.54 | 74.12 | {"adherence": 2, "motion": 2, "anatomy_face": 4, "temporal": 3, "style": 4, "reference_identity": 5} | Nine samples: Mara identity/outfit strong, but only a few small steps and studio backdrop persists instead of city street. Studio geometry shifts behind her. Reference conditioning overpowers scene/action instructions. Cold load dominates586.54s execution; retain this low-adherence result. Flags: {"artifacts": true, "text_garbling": false, "watermark": false} |
| V3V4_daynight_wan22-flf14-lightning-fp8_44002 | primary | 4 steps,cfg1,1280x720,81 frames,16fps,seed44002 | 96.60 | 82.72 | {"adherence": 5, "motion": 4, "anatomy_face": 4, "temporal": 5, "style": 5} | Nine samples including endpoints show a clear daylight-to-night transition with practical lights coming on. Mara stays stable and nearly still, as expected for the timelapse framing; background traffic changes. No doubled features or obvious text/watermark. Strong endpoint control. Flags: {"artifacts": false, "text_garbling": false, "watermark": false} |
| V3V4_loop_wan22-flf14-lightning-fp8_44004 | primary | 4 steps,cfg1,1280x720,81 frames,16fps,seed44004 | 89.35 | 50.57 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 3, "style": 5} | Nine samples show stable Mara/garden with very subtle movement. Identical supplied endpoints did not produce a pixel-seamless loop: decoded first/last RGB MAE5.50 versus average neighboring-frame MAE0.358 at320x180. Visible pose is close, but residual background/face drift makes the seam stronger than ordinary frame changes. No postprocessing used to conceal the result. Flags: {"artifacts": true, "text_garbling": false, "watermark": false} |
| V3V4_matchcut_wan22-flf14-lightning-fp8_44003 | primary | 4 steps,cfg1,1280x720,81 frames,16fps,seed44003 | 89.51 | 50.57 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} | Nine samples including endpoints: ceramic cup becomes a glowing orb, then diner dissolves into spaceship bridge. Endpoint identities and composition mostly retained. This is a morph/cross-dissolve, not a crisp editorial match cut; overlapping sets ghost during the transition. Little body movement. Garbled diner sign is inherited from the start keyframe. Flags: {"artifacts": true, "text_garbling": true, "watermark": false} |
| V3V4_pose_wan22-flf14-fp8_44001 | primary | 20 steps,cfg3.5,1280x720,81 frames,16fps,seed44001 | 825.11 | 85.96 | {"adherence": 5, "motion": 4, "anatomy_face": 4, "temporal": 5, "style": 5} | Reviewed nine evenly spaced frames including both endpoints: Mara raises her right arm naturally and holds the greeting pose. Identity, outfit, studio and framing stay stable; endpoint guidance is strong. Slight finger softness at intermediate poses. Full-step execution 825.1 s; sampling review rather than frame-by-frame hand validation. Flags: {"artifacts": true, "text_garbling": false, "watermark": false} |

Best-scoring primary clips by mean of adherence, motion, anatomy/face, temporal consistency and style (identity is listed separately in the table):

- [V3V4_Kite_talk_humo17-fp8-lightning_44104](V3V4_Kite_talk_humo17-fp8-lightning_44104.mp4): 4.80/5. Reviewed nine evenly spaced frames. A single chest-up anime Kite speaks silently with clear mouth changes, a blink and subtle head/shoulder motion. Silver hair, amber eyes and navy/orange jacket remain consistent; no extra person or inset. Clean facial structure and stable linework. Background changes to a warm paneled interior. No audio, so phoneme lip-sync is not assessed.
- [V3V4_pose_wan22-flf14-fp8_44001](V3V4_pose_wan22-flf14-fp8_44001.mp4): 4.60/5. Reviewed nine evenly spaced frames including both endpoints: Mara raises her right arm naturally and holds the greeting pose. Identity, outfit, studio and framing stay stable; endpoint guidance is strong. Slight finger softness at intermediate poses. Full-step execution 825.1 s; sampling review rather than frame-by-frame hand validation.
- [V3V4_daynight_wan22-flf14-lightning-fp8_44002](V3V4_daynight_wan22-flf14-lightning-fp8_44002.mp4): 4.60/5. Nine samples including endpoints show a clear daylight-to-night transition with practical lights coming on. Mara stays stable and nearly still, as expected for the timelapse framing; background traffic changes. No doubled features or obvious text/watermark. Strong endpoint control.
- [V3V4_Mara_talk_full_control_wan22-vace14-fp8_44101](V3V4_Mara_talk_full_control_wan22-vace14-fp8_44101.mp4): 4.60/5. Nine samples: coherent single-person talking close-up, natural mouth changes, stable framing and plain studio background. The duplicated full-body overlays from the original four-step configuration disappear. Hair changes to a shorter bob with fringe; facial details/identity less exact than the reference. Same prompt/reference/seed;20steps,cfg3.5,noLoRA versus4steps,cfg1,LoRA. Execution975.28s. Silent reference test, so no phoneme lip-sync score.
- [V3V4_Mara_walk_humo17-fp8-lightning_44100](V3V4_Mara_walk_humo17-fp8-lightning_44100.mp4): 4.40/5. Reviewed nine evenly spaced frames. Clear forward walk along an outdoor brick-walled street, with alternating arm/leg positions and a coherent single face. Teal jacket, charcoal trousers, brown boots and auburn bob remain. Slight facial/hair drift; feet move outside the frame as she approaches, limiting late foot-contact assessment. Corrected zero-audio conditioning works; silent output.

Mara comparison (reviewed before completing Kite): HuMo best follows the outdoor walking setting, but its close-up introduces a severe second full-body figure. VACE full20 produces a clean single close-up with changed fringe/facial identity; its walk retains the studio background and its sword blade shrinks substantially. SkyReels Sage2 follows the walking setting with harsh contrast and garbled signage, renders the close-up with a strongly stylized/plastic face, and substitutes a studio walk with lowered sword for the ready stance. All three preserve recognizable clothing cues more reliably than exact face identity. These are three shots from one reference/seed per shot, not broad model rankings.

## Kite reference findings

HuMo produced the strongest silent close-up in this small comparison: clear mouth changes, stable anime linework and recognizable identity. Its sword stance also best matched the requested action and retained the full body, though blade shape varied. Its busy ground textures and high camera angles were less convincing.

VACE kept the reference appearance especially faithfully in the talk shot, but ignored the requested close-up and retained a full-body studio composition. Its sword motion was more active than SkyReels but introduced blurred legs and changing blade geometry. SkyReels preserved recognizable hair and outfit and produced a coherent talk shot, yet its sword case stayed nearly static with cropped feet and excessive background speed lines. All three walking cases cropped the legs enough to limit gait assessment.

The labeled Kite comparison is in comparisons/kite_reference_models.mp4. The contact review and the composite agree on these framing differences. These are single-seed, reference-specific observations, not a universal model ranking.