# V7 — complete

6 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 Historical setup observations and superseded attempts are preserved in SETUP_HISTORY.md and LEDGER.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

InfiniteTalk is the strongest completed dialogue option in this set: both its single-speaker and spatially routed two-speaker clips retain stable identities and broadly plausible mouth activity. The two-person result preserves Mara on the left and Rowan on the right, although its lighting is more cinematic than flat sitcom lighting. These are coarse temporal visual judgments, not phoneme-level synchronization measurements.

HuMo replaced gated LTX in the three-model comparison after its smoke retained a more coherent face than the initial S2V smoke. HuMo preserves both reference identities, but mouth motion extends beyond the short speech intervals, including the quiet tail. Its two-person workflow uses mixed audio without explicit left/right routing. Partly unmatched image-embedding LoRA keys were reported, while 813 transformer patches applied; results describe that actual configuration.

Kandinsky Pro single-speaker is a retained quality failure: a second copy of Mara enters from the left, despite the requested single speaker. Correctly recognized speech does not rescue the duplicated-person failure. Native audio begins later than the generic-TTS inputs, so comparisons use actual output timing. The two-person Kandinsky clip preserves both distinct identities and left/right positions without an extra person, but Rowan keeps an open-mouth expression across both listener and quiet intervals. Coarse mouth timing is weaker than InfiniteTalk. Cinematic lighting also misses the requested flat sitcom look.

Local CPU ASR broadly recovered the intended short lines, but the final Kandinsky pair transcript reads “Coffee? Ah, yes, peas.” This could be recognition error; exact audio wording remains uncertain. This checks transcript intelligibility only. Audio listening quality and phoneme alignment remain unscored. No voice cloning was used: audio-driven inputs use generic Piper voices. Timings include model loading and, for Kandinsky single, a long untiled VAE decode; they are measured workflow costs rather than controlled kernel benchmarks. The auxiliary pair-reference PNG is excluded from the six-video count.


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 |
|---|---|---:|---:|---:|---|
| single | infinitetalk_single_fp8 | 366.79 | 365.966 | 63.26 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "lip_sync": 4} |
| single | kandinsky6_pro | 1296.83 | 1295.285 | 94.96 | {"adherence": 2, "motion": 3, "anatomy_face": 2, "temporal": 2, "style": 4, "lip_sync": 3, "reference_consistency": 4, "audio_quality": null} |
| single | humo17_fp8_lightning | 360.55 | 359.188 | 92.45 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_consistency": 5, "lip_sync": 3} |
| two_person | infinitetalk_multi_fp8 | 620.71 | 619.646 | 89.34 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 3, "lip_sync": 4} |
| two_person | kandinsky6_pro | 653.85 | 653.071 | 94.96 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 3, "lip_sync": 2, "audio_quality": null} |
| two_person | humo17_fp8_lightning | 305.01 | 303.674 | 52.78 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 3, "lip_sync": 3, "reference_consistency": 4, "audio_quality": null} |

## Per-clip settings and review notes

- **single / infinitetalk_single_fp8**: fp8 base checkpoint plus fp16 speech adapter; Comfy mixed inference. Wan2.1 I2V14B fp8 + InfiniteTalk single fp16 patch + Lightx2v;steps6,cfg1,seed24601,1280x720,125frames25fps;Piper generic audio5s. 20 temporal samples reviewed against the known input speech window (starts0.25s, about1.7s of speech). Mouth articulates during the first portion then settles during padded silence; identity, face and jacket remain stable. Lip-sync score is coarse speech-window alignment only: phoneme-level timing and perceived voice quality were not assessed. Long quiet tail is intentional input padding. No doubled features or added watermark seen.
- **single / kandinsky6_pro**: BF16/F32 mixed checkpoint tensors; default BF16 diffusion inference (runtime log); fp16 Qwen encoder where used. 1280x720,121frames24fps,steps10,cfg1,seed24601; I2V K Mara reference;native audio speech markers. Twenty temporal samples plus full-resolution2.5sframe confirm an unwanted second copy of Mara entering from left in a single-speaker scene. Individual faces retain reference appearance, but duplicated character is a clear composition/anatomy-integrity failure; artifact flagged, no reroll. Primary mouth moves through speech and rests late, but begins before stronger audio activity (threshold-based silence check approx2.14-4.33s, not reliable phoneme timing); coarse lip-sync3 provisional. Local ASR recovers Good morning. Nice to see you. Nativeaudio listening quality unassessed.1296.83s wall includes ~8.4min untiled video decode;1280x720121f24,5.042s video/5.062sAAC.
- **single / humo17_fp8_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. 6steps,cfg1,shift8,seed24601,1280x720,125frames25fps; generic Piper; mixed audio without spatial speaker routing for two-person case. 20 temporal samples: canonical face/jacket/plain background preserved, coherent head motion and speaking-like mouth changes during supplied first~2s. Mouth opens again late in intentionally silent tail, lowering coarse sync confidence. Lip_sync3 is speech-window visual assessment only, not phoneme matching. No doubled features or obvious visual artifacts. Runtime warns that some Lightx2v image-embedding keys are unmatched; transformer patches applied (813), matching supplied HuMo template behavior.
- **two_person / infinitetalk_multi_fp8**: fp8 base checkpoint plus fp16 speech adapter; Comfy mixed inference. Wan2.1 I2V14B fp8 + InfiniteTalk multi fp16 + Lightx2v;steps6,cfg1,seed24601,960x720,125frames25fps;two generic voices,left/right masks;5s. 20 full-scene and20 paired face crops sampled at4fps: stable Mara-left/Rowan-right identities, coherent restrained gestures, no doubled features. Woman mouth changes around supplied first speech window0.5s and man around second window2.3s; listener comparatively still. Coarse turn-taking/lip timing4 only, not phoneme-level validation or listening quality. Warm shadowed kitchen looks more cinematic than flat80s sitcom. Audio5.056s/video5s differ by AAC padding.
- **two_person / kandinsky6_pro**: BF16/F32 mixed checkpoint tensors; default BF16 diffusion inference (runtime log); fp16 Qwen encoder where used. 4:3 960x720,121frames24fps,steps10,cfg1,seed24601; I2V canonical pair reference;native audio alternating dialogue. Reviewed 20 whole-scene and upper-body temporal samples. Mara remains on the left and Rowan on the right with two coherent faces and no extra person; kitchen and outfits stay stable. Restrained hand gestures accompany dialogue, but Rowan retains a conspicuous open-mouth expression even before audio starts and during listener/quiet periods. Audio above -30dB occurs approximately1.07-1.56s and2.26-3.79s with short gaps; visible mouth timing is weak at this coarse scale. Local tiny.en ASR returned Coffee? Ah, yes, peas.; this may be recognition error and does not prove exact wording or audio quality. Cinematic shadows differ from requested flat sitcom lighting. Lip-sync score is a coarse sampled timing judgment, not phoneme alignment or listening review.
- **two_person / humo17_fp8_lightning**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. 6steps,cfg1,shift8,seed24601,960x720,125frames25fps; generic Piper; mixed audio without spatial speaker routing for two-person case. Reviewed20 full-scene samples plus20 upper-body crops. Both canonical adult identities/outfits and kitchen remain stable, with coherent gestures and no doubled facial features seen. Woman initially gestures/speaks and man responds, but broad smiles and mouth movement extend outside the short known speech windows, including the silent tail; mixed audio offers no explicit spatial speaker routing. Coarse lip-sync3 only, not phoneme-level assessment. ASR recovered Coffee? Yes, please. Lighting remains cinematic rather than flat80s sitcom; audio listening quality unassessed. 5.000s video/5.056sAAC padding.