# V1 — complete

34 verified and explicitly reviewed planned outputs; 17 gated cells skipped; 0 terminal job exceptions. Historical setup observations and superseded attempts are preserved in SETUP_HISTORY.md and LEDGER.

## Final reviewed results

The available grid compares Wan14B Lightning at four steps, CFG1, shift5 and81 frames/16fps against HunyuanVideo1.5 at20 steps, CFG6, shift9 and121 frames/24fps. Both use seed24601 and720p, with960x720 for the4:3 styles. LTX2.5 accounts for17 gated skips. The plan lists eight example style prompts but requires nine; a realistic two-person social dance supplies the ninth, alongside the eight motion-ladder prompts. This is a practical configured-workflow comparison, not equal compute or equal native frame rate.

Hunyuan generally preserves the explicitly requested anime look, including the neutral motion-ladder fights that Wan rendered photographically. Both models produce stable-looking adults in several simple scenes, but semantic action completion remains weaker than visual polish. Sword clips usually give a brief exchange followed by a long guard hold. The martial-arts clips do not establish a convincing controlled throw. Hunyuan's neutral anime martial-arts result includes kicks and a leg catch followed by a crouching drop, with a late background patch; the realistic version gives kicks and a clinch/reset. Correct rendering style does not imply correct combat mechanics.

Framing is another recurring limitation. Hunyuan's neutral walk remains a head-and-shoulders shot, and its neutral dance repeatedly crops the head or feet. The realistic sword motion-ladder clip keeps both fighters fully visible and is a better candidate for studying weapon motion. Wan's neutral dance has useful full-body framing and readable steps. Hunyuan's Mara walk preserves broad identity cues but changes the requested field jacket into a blazer and ends in a close-up; that makes it useful for face/hair detail evaluation but weaker for canonical outfit adherence.

The analog styles expose different weaknesses. Hunyuan's newsreel is near-monochrome with a period setting, while Wan's remains visibly desaturated color. Neither VHS result convincingly reproduces the requested analog texture; Hunyuan adds malformed timestamp-like glyphs. The sitcom Hunyuan clip adds a third adult to a two-person prompt. Retro bridge scenes broadly retain the requested set and framing but often substitute a small pivot or pose for walking across the bridge.

Hunyuan grid decoding uses512-pixel tiles,64-pixel overlap, temporal64 and overlap8. The first grid result received an explicit seam/continuity review before the rest were released; the original untiled V0 smoke remains separate. No controlled decoder-speedup claim is made. Warm Hunyuan grid jobs take roughly14 minutes at16:9 and8.4 minutes at4:3 in this configuration; Wan Lightning is much faster, with cold shared-storage loading causing substantial outliers. Final per-clip timings and whole-device VRAM are tabulated below. Retained models/cache mean VRAM is not the minimum standalone requirement.

The final gymnastics and breakdance cases also miss their requested motion sequences: Hunyuan gymnastics ends in a long handstand, while breakdance substitutes standing kicks and an aerial flip for windmills/headspin/freeze. These are retained quality failures. Hero selection is handled by GPU0 with all four mandatory fight prompts and six highest-ranked remaining prompts, using three distinct seeds; its matched Kandinsky control may affect the second-model choice.


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_F1 | wan22_lightning | 196.28 | 195.578 | 50.27 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| S1_F1 | hunyuan15 | 928.17 | 927.469 | 38.82 | {"adherence": 3, "motion": 2, "anatomy_face": 4, "temporal": 4, "style": 4} |
| S1_M6 | wan22_lightning | 261.21 | 260.011 | 49.87 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| S1_M6 | hunyuan15 | 837.56 | 836.714 | 38.78 | {"adherence": 2, "motion": 2, "anatomy_face": 3, "temporal": 3, "style": 4} |
| S2_F1 | wan22_lightning | 415.52 | 414.215 | 91.67 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| S2_F1 | hunyuan15 | 836.63 | 836.039 | 38.83 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| S2_M1 | wan22_lightning | 139.89 | 138.794 | 89.98 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4} |
| S2_M1 | hunyuan15 | 838.59 | 837.933 | 38.78 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "character_description": 3} |
| S3_M1 | wan22_lightning | 85.59 | 84.18 | 86.85 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 3} |
| S3_M1 | hunyuan15 | 839.45 | 838.835 | 38.83 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4} |
| S4_M2 | wan22_lightning | 56.39 | 55.276 | 82.04 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 3} |
| S4_M2 | hunyuan15 | 506.37 | 505.236 | 38.78 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| S5_M1 | wan22_lightning | 54.65 | 53.553 | 82.04 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4} |
| S5_M1 | hunyuan15 | 506.12 | 505.483 | 38.75 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| S6_M2 | wan22_lightning | 55.73 | 54.575 | 82.04 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 2} |
| S6_M2 | hunyuan15 | 505.65 | 505.026 | 38.75 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 2} |
| S2_M2 | wan22_lightning | 84.99 | 83.966 | 84.96 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| S2_M2 | hunyuan15 | 836.98 | 836.414 | 38.78 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4} |
| M1 | wan22_lightning | 88.43 | 87.344 | 84.96 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4} |
| M1 | hunyuan15 | 835.23 | 834.547 | 38.78 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| M2 | wan22_lightning | 85.29 | 84.198 | 84.96 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4} |
| M2 | hunyuan15 | 840.2 | 839.56 | 38.75 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4} |
| F1_anime | wan22_lightning | 86.11 | 85.053 | 84.96 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 1} |
| F1_anime | hunyuan15 | 835.99 | 835.45 | 38.78 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| F1_real | wan22_lightning | 85.61 | 84.055 | 85.47 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| F1_real | hunyuan15 | 834.49 | 833.838 | 38.83 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| F2_anime | wan22_lightning | 85.68 | 84.226 | 84.96 | {"adherence": 2, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 1} |
| F2_anime | hunyuan15 | 836.63 | 835.962 | 38.78 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 3, "style": 4} |
| F2_real | wan22_lightning | 85.49 | 84.159 | 84.96 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| F2_real | hunyuan15 | 838.18 | 837.625 | 38.83 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| M5 | wan22_lightning | 85.53 | 84.158 | 84.96 | {"adherence": 2, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| M5 | hunyuan15 | 838.6 | 837.955 | 38.78 | {"adherence": 2, "motion": 3, "anatomy_face": 3, "temporal": 3, "style": 4} |
| M6 | wan22_lightning | 85.88 | 84.694 | 84.96 | {"adherence": 4, "motion": 3, "anatomy_face": 3, "temporal": 3, "style": 4} |
| M6 | hunyuan15 | 834.75 | 834.083 | 38.78 | {"adherence": 2, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |

## Per-clip settings and review notes

- **S1_F1 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. Reviewed 20 temporal samples across five seconds. Two cel-shaded adult fighters perform a visible blade exchange and separate; courtyard and identities remain stable. Footwork is limited and sword-contact mechanics are simplified. No doubled facial features, added text, or watermark observed in samples. Mild sword/hand ambiguity; promising pipeline result, not a perfect combat simulation.
- **S1_F1 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed20 temporal samples plus full-resolution0.5s/2s/4s frames for tiled-decoder gate. No visible spatial tile boundaries or abrupt window break in these samples; horizontal bright edge streaks are consistent with prompted speed-line styling, not an evident tile grid. Coherent cel-shaded adult-intended fighters and courtyard; a brief high parry occurs around1-2s, then long planted poses, with weak footwork and limited exchange. Faces remain singular/coherent. Decoder512px overlap64 temporal64 overlap8 is disclosed;928.17s wall927.469exec38.815GiBpeak. This is not a controlled untiled-vs-tiled speed test.
- **S1_M6 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. 20-frame temporal contact review: corrected silver-haired/navy-orange adult Kite identity retained. First half is mostly a planted pose and arm flourish, followed by cartwheel/handstand and one-hand freeze. No complete windmill or headspin demonstrated; background arena lettering is garbled. Mild hand-ground/contact ambiguity; fixed arena and cel style stable.
- **S1_M6 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed20 temporal samples plus full-resolution2.25sframe. Silver hair and navy jacket with orange trim broadly match canonical Kite, but model invents an angular kite-like handheld prop, performs standing kicks/airborne rotation, and ends holding it in a planted pose. No windmill or headspin; some early framing crops the head and support/flight transitions are strained. Brief motion trails and body/prop transitions reduce temporal clarity. Cel-shaded style retained; no clear doubled facial features seen. The prop may relate to the ambiguous name Kite, but causation is untested. This quality miss is retained without reroll.
- **S2_F1 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. 20 sampled frames: two distinct adult fighters perform a readable overhead parry and return to guard. Stable courtyard and clothing; limited shuffle-like footwork and stiff body/contact mechanics. Blades/hands briefly ambiguous at crossing. No doubled face or visible watermark/text at contact scale. Photorealistic style retained.
- **S2_F1 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed20 temporal samples plus full-resolution1.75sframe. Two stable realistic adult fighters perform an overhead strike, crossed-blade parry, lower swords and reset; full-body courtyard framing and faces stay coherent. Footwork is stiff and much of the final portion is static. Full-resolution inspection shows one broad bright blade rather than a clear extra weapon, but blade-edge/guard detail is artificial/ambiguous across motion; fine contact physics remain uncertain. No doubled facial features seen. Retain this measured single-seed result.
- **S2_M1 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. 20 sampled frames: believable forward walk, stable adult face/teal jacket and sunlit street. Starts full-body but camera does not maintain full-body tracking as she approaches; no distinct turn from profile toward camera. Natural-light realistic style strong, readable early footfalls; tiny face/finger detail limited by contact scale. No obvious doubled features, watermark or conspicuous artifact.
- **S2_M1 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Twenty temporal samples show a coherent rear-view walk, turn toward camera and stable face through a closing camera move. Olive-toned skin, short dark bob, prominent cheek mark, teal outerwear, gray trousers and brown boots broadly follow the description, but the field jacket becomes a tailored blazer and full-body tracking becomes a close-up for most of the latter clip. Eye/hair nuances and exact canonical identity are not guaranteed by T2V description alone. Natural sunlight and city remain coherent; no doubled facial features or clear tile boundary seen. Potential detail-upscale source because the face/hair are large and stable, subject to final grid selection.
- **S3_M1 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. 20 sampled frames: coherent adult woman walking on a crowded street with period-looking clothing and cars. Full body visible initially; foreground passerby crosses naturally. Image is desaturated rather than strictly black-and-white (colored lights/buildings/skin remain), with little visible film grain or gate weave; looks like a clean modern period reconstruction rather than newsreel. Main face/body stable at contact scale.
- **S3_M1 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Twenty temporal samples show a coherent waist-up tracking walk with a stable adult face, period hair/clothing and vintage traffic. Near-monochrome cool-gray rendering is a stronger period/newsreel cue than Wan S3, but film grain and gate weave are weak and the requested full-body view is absent throughout, so leg/foot mechanics cannot be assessed. Foreground/background motion appears coherent in samples; no doubled facial features or obvious tile boundary. Clean classic-film look rather than a complete distressed-newsreel treatment.
- **S4_M2 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,960x720,81frames,16fps; fp8 models fp16 umT5. 20 sampled frames: two stable adult performers in full-body4:3 kitchen staging, mostly planted feet with arm waving and body bobbing; limited synchronized social-dance footwork. Bright studio staging reads as sitcom, but modern-looking appliances and crisp HD texture weaken the1980s analog-bloom target. Faces/outfits/background stable; no conspicuous visual corruption at contact scale.
- **S4_M2 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,960x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Twenty temporal samples preserve full-body4:3 kitchen-set framing, stable separate adult faces and broad classic-sitcom staging. The prompt asks for two synchronized dancers, but there are three adults: the left man and center woman perform largely independent steps/gestures while the right man mostly stands with hands at his hips. Motion is coherent but choreography weak. Warm wood cabinets/period-ish staging support the style; analog bloom remains subtle. Extra third person is a prompt-count miss, not duplicated facial features. No obvious technical artifact seen in samples.
- **S5_M1 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,960x720,81frames,16fps; fp8 models fp16 umT5. 20 sampled frames: adult officer walks across raised bridge area, turns toward camera and approaches. Stable ochre uniform/face; another crew member crosses behind her (occlusion, not duplicated facial features).4:3 retro staging and saturated colors read well, although space panorama is crisper/more digital than a painted backdrop. No conspicuous corruption at contact scale.
- **S5_M1 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,960x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Twenty temporal samples show a coherent adult officer in a white retro uniform, plastic-looking bridge console and saturated red/blue space backdrop in4:3. She mainly pivots away and back with a few small steps, rather than clearly walking across the bridge; latter half is a posed hold with hands clasped. Face/body remain coherent through camera movement. Broad retro-space-television styling succeeds, but the colorful nebula is digitally crisp rather than clearly painted, and soft-focus film texture is weak. No obvious doubled features or tile boundary seen.
- **S6_M2 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,960x720,81frames,16fps; fp8 models fp16 umT5. 20 sampled frames: coherent full-body shuffle/hip-hop-like steps and stable adult performer/studio.4:3 output correct, but image remains crisp modern HD with little tracking noise or chroma bleed and no readable timestamp; a few corner marks do not establish VHS look. Requested analog degradation largely absent.
- **S6_M2 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,960x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed 20 temporal samples plus enlarged top-edge crop at 2.5 s. Stable adult dancer with coherent shuffle, hops and knee lifts in a 4:3 studio frame. Image remains clean and digitally sharp, with little convincing VHS chroma bleed, tracking damage or interlaced texture. Small timestamp-like white glyphs along the upper edge are malformed and unreadable; no identifiable watermark. Face/body remain coherent at sampled times.
- **S2_M2 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. 20 sampled frames: two adults in a naturally lit studio retain face/outfit/full-body framing. Mostly mirrored extended-arm gestures with limited planted-foot weight shifts and small hops; synchronized but shallow social-dance choreography. Stable room/reflections and realistic lighting, no conspicuous corruption at contact scale.
- **S2_M2 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed 20 temporal samples: two distinct adults perform small synchronized social-dance steps, weight shifts and arm gestures in a bright spacious studio. Full bodies remain readable, with stable faces, clothes and separation. Movement is modest but more clearly stepped than the Wan style-grid partner dance. A later gray-shirted figure at left is visible within the wall mirror, consistent with a reflection rather than an extra dancer. Fine reflection geometry and hand detail are not fully resolved by the contact sheet; no obvious doubled facial features or tile boundaries at sampled times.
- **M1 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. 20 sampled frames: stable fictional adult woman walks naturally toward camera on a coherent street. Early full-body footfalls readable; framing tightens and feet approach lower edge near end instead of sustained full-body tracking. Face/outfit stable, neutral realistic styling. No conspicuous visual corruption at contact scale.
- **M1 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed 20 temporal samples. Stable adult woman in a gray jacket, with coherent face, hair and city setting during a gentle tracking/orbiting camera move. Framing stays at head and shoulders rather than the requested full-body view; legs and foot contacts are never visible, so walking gait cannot be established. Motion is mostly subtle upper-body/head movement and camera travel. Neutral cinematic rendering is coherent; no obvious doubled facial features, severe flicker or tile boundary at sampled times.
- **M2 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. 20 sampled frames: clear full-body hip-hop-like stepping, knee lifts and varied arm gestures on a spacious studio floor. More movement variety than the two-person social-dance cases; face, gray shirt and room stable. Glossy floor reflections visible but fine reflection/hand physics not fully resolved at contact scale. No conspicuous doubled features or watermark.
- **M2 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed 20 temporal samples. A single adult dancer performs coherent knee lifts, forward kicks, small hops and lateral steps with stable clothing and limbs. The framing repeatedly cuts off the head in the first part and feet later, missing the requested full-body wide view. Face is coherent where visible, but cannot be assessed during head crops. Studio background and neutral cinematic rendering remain stable; no evident doubled features or tile seams. Wan M2 is a stronger full-body source candidate despite this result having readable dance motion.
- **F1_anime / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. Twenty sampled frames show two stable adult fighters performing a readable high parry and reset, with limited footwork and ambiguous blade contact. The requested anime rendering is absent: the result is photographic/live-action throughout. No doubled facial features seen in these samples.
- **F1_anime / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed 20 temporal samples plus full-resolution 1.75-second frame. Clearly anime-rendered fighters retain distinct faces, hair and outfits, unlike the photographic Wan result for this same prompt. A short blade exchange/parry is followed by separation and a long guard hold. Motion and footwork are limited, and lower legs/feet are cropped rather than full-body wide framing. Two swords remain identifiable; one free hand touches a blade during the exchange, so combat mechanics are approximate. No evident doubled facial features or major weapon duplication at sampled times.
- **F1_real / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. Twenty sampled frames show stable realistic adult fighters and courtyard, a high parry, follow-through and reset. Sword geometry/contact is occasionally ambiguous and footwork remains stiff; no doubled facial features seen in the samples.
- **F1_real / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed 20 temporal samples. Two distinct adult fighters remain in full-body view through a brief controlled sword exchange, parry and separation. Faces, clothing, two weapons and stone courtyard remain coherent at sampled times. Footwork is small and stiff, with a long static guard hold in the latter half, limiting motion score. Camera travels gently backward/sideways without an evident tile seam or doubled facial features. Better full-body framing than the matching Hunyuan anime motion-ladder duel.
- **F2_anime / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. Twenty sampled frames show coherent alternating front kicks and guards with stable adult identities, but no controlled throw. Requested anime rendering is absent: photographic martial-arts training throughout. Hands and kick contact remain approximate at contact-sheet resolution.
- **F2_anime / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed 20 whole-clip samples, an additional 10fps contact sheet over seconds3-5, and full-resolution4.75s. Proper anime style and two distinct faces/outfits persist. Both fighters kick and make blocking/contact gestures; the ending shows a leg catch followed by the black-clad fighter dropping into a crouch, not a clearly readable controlled throw. Lower legs and landing area are cropped. Contact mechanics and support are approximate, with held poses. A pale human-like patch emerges in the background panel late in the clip, reducing temporal consistency; no doubled foreground facial features.
- **F2_real / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. Twenty sampled frames show a readable front kick toward pads, return to guard, and small stance shifts, with stable realistic adult identities. The requested controlled throw never appears; this reads as pad practice rather than the full kick-block-throw sequence. Fine foot/contact physics remain approximate.
- **F2_real / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed 20 temporal samples plus full-resolution2.5s and3s frames. Two distinct adult men retain coherent faces and clothing through repeated high kicks, blocking/guard gestures and a brief clinch, then reset. No controlled throw is visible. Much of the clip crops lower legs and feet, limiting assessment of support and missing full-body wide framing. Kick/contact mechanics remain approximate. Red/yellow lettering-like clothing graphics are malformed or unreadable, especially the back patch; no identifiable overlay watermark. Background and identities are otherwise stable at sampled times.
- **M5 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. Twenty sampled frames show a coherent sideways cartwheel into a standing balance, rather than the requested backward handspring and full aerial twist. Body remains mostly coherent through inversion, with ambiguous hand support and floor reflections. Wide padded-floor setting and cinematic style preserved; no doubled facial features seen. Sampled review does not certify precise gymnastics mechanics.
- **M5 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed 20 whole-clip samples,36 samples at12fps over the first3seconds, and full-resolution3.5s handstand. The gymnast performs repeated backward aerial rotations and settles into a prolonged handstand, ending with legs opening rather than a clear standing landing. The requested back-handspring plus full-twist sequence is not established. Takeoff/rotation transitions are floaty and support mechanics approximate; hands/wrists merge or become difficult to distinguish in the handstand, and facial detail is often hidden or blurred. Gym apparatus/background detail changes during camera travel. Neutral cinematic gym styling remains recognizable. This is a retained quality failure, not a reroll candidate.
- **M6 / wan22_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. steps=4,cfg=1,shift=5,seed=24601,1280x720,81frames,16fps; fp8 models fp16 umT5. Twenty sampled frames show substantial floor rotation, hand-supported inversion, and a head-supported final pose, closer to the requested breakdance sequence than canonical Kite S1_M6. Leg/body configurations and support transitions become strained; exact windmill/headspin mechanics are not certified from samples. Clothing rides up consistently, while audience details drift slightly. No clearly doubled facial features, but inverted face is poorly resolved.
- **M6 / hunyuan15**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16; tiledVAE512 overlap64 temporal64 overlap8 (untiled smoke retained separately). Reviewed 20 temporal samples. Adult dancer performs standing kicks, turns and a tucked aerial flip, then resets to upright arm gestures and a crouch. Requested windmills, headspin and floor freeze are not established. Full body is mostly visible; face and clothing remain coherent, with approximate aerial physics and motion blur. Stable bright studio background; no evident doubled face or watermark.

## Follow-up selection

Ten hero prompts from the fixed selection rule: F1_anime, F1_real, F2_anime, F2_real, M2, S2_M1, S2_M2, M1, S2_F1, S5_M1. The final hero model choice and execution are owned by GPU0.

V9: Wan Mara walk, Wan neutral dance, Hunyuan realistic sword duel. V10: Hunyuan Mara face close-up and Wan neutral dance. V6: Wan Mara walk for rain/night, Wan neutral dance for anime restyle. Exact source paths and reasons are in state/gpu1_V9_sources.json, gpu1_V10_sources.json, and gpu1_V6_sources.json. The source choice does not erase the original framing, identity, or physics limitations.