# V2 — complete

16 verified and explicitly reviewed planned outputs; 8 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 Historical setup observations and superseded attempts are preserved in SETUP_HISTORY.md and LEDGER.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 eight canonical K style references were animated once per available model at seed24601: Wan2.2 I2V fp8 full20steps CFG3.5 shift5,81frames16fps; Hunyuan1.5 I2V fp16 CFG-distilled20steps CFG1 shift7,121frames24fps with tiled VAE512/64 temporal64/8. S4-S6 retain960x720 4:3; others1280x720. LTX2.5 stock access is gated, so its eight cells are skipped. Every available result was decoded, checked against its reference/start/middle/end panel and sampled through time; targeted full-resolution weapon frames confirmed S1 continuity problems.

Both models are more dependable on simple walking and social dance than on precise action choreography. S2 realistic walk is a useful identity-preservation example for both, though Hunyuan barely performs the requested turn. S3 preserves monochrome/period setting but weakly expresses newsreel grain/gate weave. S4 keeps the kitchen and identity; Hunyuan develops horizontal banding. Wan S6 retains the VHS treatment better: Hunyuan quickly removes the reference tracking lines despite the preservation request. The S5 canonical reference itself is anime Kite on a bridge, so neither result establishes a live-action1970s television look; Wan barely moves, while Hunyuan advances into a close-up that crops the eyes and misses stop-and-turn.

Neither model completes the S7 windmill/headspin/freeze sequence. Wan substitutes standing motion and a cartwheel-like inversion; Hunyuan uses upright kicks/turns with visible ghosted limbs. S1 single-sword continuity fails in both: Wan has implausible scabbard/blade geometry, Hunyuan changes into dual swords. Both S8 results introduce an on-screen opponent despite the off-screen instruction; Hunyuan also has severe body trails and unstable blade/contact geometry. These quality failures are retained as benchmark evidence, not rerolled away. Suggested write-up examples: S2 for reference consistency, S6 for style-preservation differences, S1/S7/S8 for the gap between recognizable characters and controlled action.

Measured request wall times: Wan468.24-848.57s, median763.44s; Hunyuan270.56-496.94s, median439.07s. These are single-seed mixed-resolution measurements with differing native frame counts and cache/loading behavior, not a controlled universal speed ratio. Additional postcollection/QA time is excluded from inference timing. No model-ranking claim extends beyond this small matrix.


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 |
|---|---|---:|---:|---:|---|
| S1 | wan22_i2v_fp8 | 848.57 | 847.554 | 89.28 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 3, "style": 4, "reference_identity": 4} |
| S1 | hunyuan15_i2v_fp16 | 496.94 | 496.065 | 81.66 | {"adherence": 2, "motion": 2, "anatomy_face": 4, "temporal": 2, "style": 4, "reference_consistency": 3} |
| S2 | wan22_i2v_fp8 | 767.13 | 766.299 | 72.78 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_identity": 4} |
| S2 | hunyuan15_i2v_fp16 | 439.59 | 438.518 | 71.75 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_consistency": 4} |
| S3 | wan22_i2v_fp8 | 764.63 | 763.948 | 62.33 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 3, "reference_identity": 4} |
| S3 | hunyuan15_i2v_fp16 | 441.16 | 440.427 | 72.64 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 3, "reference_consistency": 4} |
| S4 | wan22_i2v_fp8 | 471.1 | 470.215 | 57.69 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 3, "reference_identity": 4} |
| S4 | hunyuan15_i2v_fp16 | 273.31 | 272.456 | 56.43 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 3, "style": 3, "reference_consistency": 4} |
| S5 | wan22_i2v_fp8 | 468.24 | 467.281 | 56.61 | {"adherence": 2, "motion": 2, "anatomy_face": 4, "temporal": 3, "style": 3, "reference_identity": 3} |
| S5 | hunyuan15_i2v_fp16 | 275.24 | 274.259 | 55.28 | {"adherence": 2, "motion": 3, "anatomy_face": 4, "temporal": 3, "style": 3, "reference_consistency": 3} |
| S6 | wan22_i2v_fp8 | 469.1 | 468.321 | 57.47 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 3, "style": 4, "reference_identity": 4} |
| S6 | hunyuan15_i2v_fp16 | 270.56 | 269.751 | 56.46 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 3, "style": 2, "reference_consistency": 3} |
| S7 | wan22_i2v_fp8 | 769.09 | 768.368 | 62.3 | {"adherence": 2, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_identity": 4} |
| S7 | hunyuan15_i2v_fp16 | 438.55 | 437.72 | 56.43 | {"adherence": 1, "motion": 3, "anatomy_face": 3, "temporal": 2, "style": 4, "reference_consistency": 4} |
| S8 | wan22_i2v_fp8 | 762.25 | 761.486 | 62.18 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 3, "style": 4, "reference_identity": 4} |
| S8 | hunyuan15_i2v_fp16 | 440.35 | 439.319 | 56.48 | {"adherence": 2, "motion": 2, "anatomy_face": 3, "temporal": 2, "style": 4, "reference_consistency": 3} |

## Per-clip settings and review notes

- **S1 / wan22_i2v_fp8**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. seed24601,1280x720,steps20; cfg3.5,shift5,81frames16fps;fp8 hi/lo. Reviewed20 temporal samples, reference/start/middle/end panel and full-resolution2.75s/3.25s frames. Kite retains silver hair, coherent face and navy/orange costume in anime style. Mostly planted stance with arm/sword raising and return to front guard; little clear forward stepping or decisive swing. Black sheath-like material remains attached around the outer blade and changes shape during the motion, producing an implausible partially sheathed weapon. Hands and sword grip are approximate; no doubled facial features. Background remains stable; feet partly crop as framing tightens.
- **S1 / hunyuan15_i2v_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. seed24601,1280x720,steps20; cfg1,shift7,121frames24fps;cfg-distilled fp16; tiledVAE512 overlap64 temporal64 overlap8. Reviewed 20-frame temporal contact, reference/start/middle/end panel and full-resolution3.5s frame. Face, silver hair and navy/orange outfit remain recognizable, with cel-shaded styling. The single sword/scabbard configuration morphs into two unsheathed swords; extra curved blade shapes appear during the draw/swing. Feet remain broadly planted while camera pushes in and crops the legs, so forward step and steady full-body framing are weak. Weapon continuity is a clear failure despite stable facial anatomy.
- **S2 / wan22_i2v_fp8**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. seed24601,1280x720,steps20; cfg3.5,shift5,81frames16fps;fp8 hi/lo. Reviewed20 temporal samples and reference/start/middle/end panel. Mara walks forward, makes a small hand gesture and turns her head/body while retaining the short auburn bob, teal jacket, dark trousers and recognizable face. Realistic city, natural light and reference composition are maintained before the camera closes in; later foot contact cannot be assessed because legs crop. Background cars move consistently at sampled times. No evident doubled face or extra limbs; fine hand/traffic detail is limited by sampled review.
- **S2 / hunyuan15_i2v_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. seed24601,1280x720,steps20; cfg1,shift7,121frames24fps;cfg-distilled fp16; tiledVAE512 overlap64 temporal64 overlap8. Reviewed 20-frame contact and reference/start/middle/end panel. Natural forward walk with small arm swing; face, short dark hair, teal jacket and olive top remain consistent with the reference. Realistic city lighting and background layout remain coherent. She faces the camera throughout with little distinct requested turn; advancing framing crops the feet and eventually lower body. No conspicuous doubled facial features or extra limbs in sampled frames.
- **S3 / wan22_i2v_fp8**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. seed24601,1280x720,steps20; cfg3.5,shift5,81frames16fps;fp8 hi/lo. Reviewed20 temporal samples and reference/start/middle/end panel. Initial held stance transitions into a coherent forward walk with small head turns. Reference face, short hair, jacket/trousers and period street/cars remain recognizable. Monochrome is preserved, but the image stays relatively clean and stable with little evident film grain or gate weave. Boots/feet crop later as the camera closes. No conspicuous doubled features or new large scene artifacts at sampled times.
- **S3 / hunyuan15_i2v_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. seed24601,1280x720,steps20; cfg1,shift7,121frames24fps;cfg-distilled fp16; tiledVAE512 overlap64 temporal64 overlap8. Reviewed 20-frame contact and reference/start/middle/end panel. Coherent forward sidewalk walk, stable face, short hair and jacket, with period cars and monochrome appearance preserved. The image remains relatively clean and stable rather than visibly grainy with newsreel gate weave; those requested style elements are weak. Framing gradually crops the legs as she approaches. No conspicuous doubled features or extra limbs in sampled frames.
- **S4 / wan22_i2v_fp8**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. seed24601,960x720,steps20; cfg3.5,shift5,81frames16fps;fp8 hi/lo. Reviewed20 temporal samples and reference/start/middle/end panel. Single recognizable Mara performs playful small dance steps, knee bends, raised-arm gestures and a turn, mostly full-body within the4:3 kitchen composition. Face, teal jacket and brown boots remain coherent. Cabinets, appliances and floor hold stable; no extra person or doubled face. Warm clean cinematic lighting remains stronger than the requested flat studio/video-bloom aesthetic, although period domestic styling is recognizable.
- **S4 / hunyuan15_i2v_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. seed24601,960x720,steps20; cfg1,shift7,121frames24fps;cfg-distilled fp16; tiledVAE512 overlap64 temporal64 overlap8. Reviewed 20-frame contact and reference/start/middle/end panel. Full-body playful side-steps and raised-arm dance preserve face, teal jacket, boots and kitchen composition. Warm lighting remains more cinematic than flat multicamera sitcom lighting. Fine horizontal scanline/banding texture becomes more apparent over the later frames, including the cabinets and figure; not a clear clean slight-bloom treatment. No conspicuous doubled face or extra limbs in the sampled dance.
- **S5 / wan22_i2v_fp8**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. seed24601,960x720,steps20; cfg3.5,shift5,81frames16fps;fp8 hi/lo. Reviewed20 temporal samples and reference/start/middle/end panel. Canonical S5 reference is an anime Kite on a saturated space bridge; this anime appearance is inherited from the supplied keyframe, not newly introduced by the video model. The clip gives a small orientation change and apparent forward drift with largely planted legs, not two clearly readable steps followed by a turn. Face and silver hair remain coherent, but orange collar/waist trim and jacket opening simplify or shift. Starfield/planet composition remains stable. Clean anime graphics lack the requested1970s television softness; distinguish reference styling limitation from motion failure.
- **S5 / hunyuan15_i2v_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. seed24601,960x720,steps20; cfg1,shift7,121frames24fps;cfg-distilled fp16; tiledVAE512 overlap64 temporal64 overlap8. Reviewed temporal contact and reference/start/middle/end panel. Canonical K reference already depicts anime Kite on a bridge; this is not a newly introduced conversion from live action. The figure advances toward the camera with limited early leg motion, then grows into a poorly framed close-up that crops the top of his head and eyes. No clear two-step stop-and-turn sequence. Silver hair and navy/orange outfit remain recognizable, but jacket/collar proportions and folds shift. Saturated space background persists; requested1970s soft-focus television look is weak.
- **S6 / wan22_i2v_fp8**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. seed24601,960x720,steps20; cfg3.5,shift5,81frames16fps;fp8 hi/lo. Reviewed20 temporal samples and reference/start/middle/end panel. Mara progresses from a short held stance into hops, side steps, knee bends and broad arm gestures with full body largely visible. Teal jacket, boots, short hair and broad facial identity persist. The shirt hem rises/shortens to expose the midriff as motion develops, an outfit-consistency miss. Inherited horizontal colored tracking lines vary in strength and motion blur grows; these are partly requested VHS effects, not all new defects. No clear duplicated face or extra limb at sampled times; hand/facial fine detail is blurred during faster movement.
- **S6 / hunyuan15_i2v_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. seed24601,960x720,steps20; cfg1,shift7,121frames24fps;cfg-distilled fp16; tiledVAE512 overlap64 temporal64 overlap8. Reviewed temporal contact and reference/start/middle/end panel. Varied hip-hop-like steps, knee lifts and upper-body gestures keep the broad face/hair and teal-jacket identity. Strong colored tracking lines in the canonical reference and first frame largely disappear almost immediately, leaving a clean studio image instead of the requested persistent VHS/chroma-bleed/480i appearance. Some transient face/hand smearing during faster movement and slight garment-proportion drift; no clear extra limb in sampled frames.
- **S7 / wan22_i2v_fp8**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. seed24601,1280x720,steps20; cfg3.5,shift5,81frames16fps;fp8 hi/lo. Reviewed20 temporal samples and reference/start/middle/end panel. Kite retains silver hair, navy/orange outfit and cel-shaded face. Standing arm poses and a turn lead into a one-hand cartwheel/brief inverted split, then a crouch and upright reset. Requested windmills, headspin and final freeze are not established. Full body and clean background remain visible; transitions and support mechanics are approximate, but no conspicuous added limbs or duplicate face are seen at sample scale. Face is hidden during the inverted/back-facing action.
- **S7 / hunyuan15_i2v_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. seed24601,1280x720,steps20; cfg1,shift7,121frames24fps;cfg-distilled fp16; tiledVAE512 overlap64 temporal64 overlap8. Reviewed temporal contact and reference/start/middle/end panel. Upright kicks, turns and arm gestures end in a wide standing pose. No requested windmill, headspin or floor freeze appears in sampled sequence. Several fast turns/kicks show conspicuous translucent duplicate limb silhouettes and smeared body edges. Face, silver hair, navy/orange outfit and cel-shaded look remain broadly recognizable, but body scale/framing shifts during motion.
- **S8 / wan22_i2v_fp8**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. seed24601,1280x720,steps20; cfg3.5,shift5,81frames16fps;fp8 hi/lo. Reviewed20 temporal samples and reference/start/middle/end panel. Mara raises the sword, advances toward camera, makes an overhead strike/parry-like motion and returns to guard. Face, hair and clothing remain recognizable where visible. An opponent enters the foreground despite the off-screen instruction, obscuring action; the camera also crops Mara head/legs during the exchange. Blade shape changes from straight to visibly curved and contact mechanics are unclear. The second person is a newly visible opponent rather than a duplicated Mara face. Studio backdrop and photographic look are retained; no clear overlay text/watermark.
- **S8 / hunyuan15_i2v_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. seed24601,1280x720,steps20; cfg1,shift7,121frames24fps;cfg-distilled fp16; tiledVAE512 overlap64 temporal64 overlap8. Reviewed temporal contact and reference/start/middle/end panel. Mara raises/swings the sword, turns and finishes in a wide guard, with broad face/outfit identity and photographic studio appearance preserved. A second blonde opponent enters prominently despite the off-screen-opponent request. Fast turns and the encounter show severe transparent body/arm trails; blade length/curvature and contact geometry are unstable. Framing crops the head during the middle action. This is not a clean controlled parry sequence.