# V0B — complete

25 verified and explicitly reviewed planned outputs; 2 gated cells skipped; 0 terminal job exceptions. One additional optional Avatar model failed before rendering. Historical setup observations and superseded attempts are preserved in SETUP_HISTORY.md and LEDGER.

## Final reviewed results

This suite establishes which installed workflows execute, rather than ranking all models on a common-quality benchmark. Smoke lengths, sources and dimensions differ; cached and cold model-loading costs also differ. Use the later V1/V2 and matched V9/V10 comparisons for model selection.

Useful public examples are Kandinsky Pro's street walk/turn and canonical Mara I2V, Hunyuan's adult walk and anime I2V, the corrected native SR clips, and the controlled Fun Control parry. Kandinsky Pro preserved a coherent face and street through the turn; Lite's foreground remained coherent while background detail drifted. FramePack preserved the anime scene in its short window but invented blue energy effects. Its 37-frame output is 1.233 seconds, not evidence of long-take performance. Fun Control followed the initial source parry in the two-second Canny smoke, but blade contact and footwork remain approximate.

The diagnostic chart makes invented detail visible. ESRGAN is a useful deterministic learned baseline; SeedVR2 3B introduced conspicuous text-like marks, while 7B-sharp was cleaner in this sample but still modified tiny chart detail. FlashVSR also altered fine chart patterns. Kandinsky VSR sharpened eye/hair/seam edges while largely preserving the street source; Hunyuan SR redrew facial and costume details more noticeably on its different anime source. These are source-dependent observations, not a controlled claim that one SR model wins. The later V10 suite uses matched sources and exact 4K finishing.

Two output problems required one correction each: Hunyuan SR initially bypassed its latent upscaler and stayed at 720p; MMAudio initially generated only 3.25 seconds because its wrapper interprets feature frames at 25fps. The corrected SR is 1920x1072, and corrected MMAudio covers the source's 5.06 seconds after resampling feature input to 25fps. Originals and logs remain archived. Kandinsky's 2.25x VSR rounds 864x480 to 1952x1088. Report actual ffprobe sizes instead of nominal 1080p labels.

Generic Piper speech is intelligible to the local ASR sanity check. Audio generation and durations are verified, but listening quality, Foley semantics and phoneme-level alignment are unscored. Interpolation chart smokes verify execution; they do not establish performance on faces, occlusions or sword motion, which belong in V9. Early ESRGAN peak-VRAM measurements used five-second sampling and may miss short peaks; later jobs sample at half-second intervals. All peak figures are whole-device occupancy, including retained models and cache, not minimum standalone model requirements.

LTX-2.5 distilled and dev were skipped because gated. HunyuanVideo-Avatar could not be made compatible after two constrained dependency-install attempts; no shared package downgrades or access workarounds were used. SkyReels A2V passed after one local compatibility correction: the wrapper wav2vec override needed the output-capture decorator used by the installed Transformers parent. A small CPU regression verified hidden states before restarting GPU1 only. Its 1.96-second result preserved Mara and showed plausible coarse speech-mouth timing; it is not a full five-second dialogue comparison. LongCat passed its final corrective attempt after fixing the LoRA input and selecting explicit Comfy RoPE. Its 2.2-second close-up preserves a coherent face but is nearly stationary; no long-take or walking-performance claim follows.


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 |
|---|---|---:|---:|---:|---|
| esrgan_x2 | RealESRGAN_x2plus | 5.21 | 3.354 | 0.81 | {"adherence": 4, "motion": 4, "anatomy_face": null, "temporal": 4, "style": 4} |
| esrgan_x4 | RealESRGAN_x4plus | 5.24 | 3.336 | 0.87 | {"adherence": 4, "motion": 4, "anatomy_face": null, "temporal": 4, "style": 4} |
| ultrasharp | 4x-UltraSharp | 5.25 | 2.58 | 0.93 | {"adherence": 4, "motion": 4, "anatomy_face": null, "temporal": 4, "style": 4} |
| interp_rife | rife_vfi | 2.13 | 1.597 | 1.0 | {"adherence": 4, "motion": 4, "anatomy_face": null, "temporal": 4, "style": 4} |
| interp_rife_v4.25 | rife_v4.25 | 1.91 | 1.305 | 1.25 | {"adherence": 4, "motion": 4, "anatomy_face": null, "temporal": 4, "style": 4} |
| interp_rife_v4.25_heavy | rife_v4.25_heavy | 2.02 | 1.509 | 1.32 | {"adherence": 4, "motion": 4, "anatomy_face": null, "temporal": 4, "style": 4} |
| interp_rife_v4.26_heavy | rife_v4.26_heavy | 1.89 | 1.286 | 1.38 | {"adherence": 4, "motion": 4, "anatomy_face": null, "temporal": 4, "style": 4} |
| interp_film | film_vfi | 2.48 | 2.047 | 1.34 | {"adherence": 4, "motion": 4, "anatomy_face": null, "temporal": 4, "style": 4} |
| interp_gimm | gimm_vfi | 4.69 | 4.169 | 1.97 | {"adherence": 4, "motion": 4, "anatomy_face": null, "temporal": 4, "style": 4} |
| tts | piper_lessac | 7.153442448936403 |  | CPU | {"adherence": 4, "motion": null, "anatomy_face": null, "temporal": null, "style": null, "audio_quality": null} |
| t2v | hunyuan15_fp16 | 882.7 | 882.056 | 74.62 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4} |
| i2v | hunyuan15_i2v_fp16 | 506.93 | 505.361 | 78.14 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| sr | hunyuan15_sr_fp16 | 1111.56 | 1109.812 | 94.97 | {"adherence": 3, "motion": 3, "anatomy_face": 3, "temporal": 4, "style": 4, "sharpness": 4, "detail_preservation": 3} |
| kandinsky_t2v | kandinsky6_pro | 266.33 | 265.369 | 94.96 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "audio_quality": null} |
| i2v | kandinsky6_pro | 177.12 | 175.779 | 94.81 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "reference_consistency": 5, "audio_quality": null} |
| vsr | kandinsky6_vsr | 271.61 | 269.506 | 51.74 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "sharpness": 4, "detail_preservation": 4, "audio_quality": null} |
| seedvr_3b | seedvr2_3b_fp16 | 16.33 | 15.65 | 59.61 | {"adherence": 3, "motion": 4, "anatomy_face": null, "temporal": 3, "style": 3} |
| seedvr_7b | seedvr2_7b_sharp_fp16 | 84.18 | 82.894 | 63.69 | {"adherence": 4, "motion": 4, "anatomy_face": null, "temporal": 4, "style": 4, "sharpness": 4, "detail_preservation": 4} |
| flashvsr | flashvsr_v11 | 43.85 | 42.729 | 63.09 | {"adherence": 4, "motion": 4, "anatomy_face": null, "temporal": 4, "style": 4} |
| framepack_i2v | framepack | 238.65 | 237.141 | 88.68 | {"adherence": 3, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4} |
| longcat_t2v | longcat | 32.96 | 32.476 | 57.84 | {"adherence": 3, "motion": 2, "anatomy_face": 4, "temporal": 4, "style": 4} |
| mmaudio_foley | mmaudio | 37.62 | 36.576 | 74.93 | {"adherence": null, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4, "audio_quality": null, "audio_sync": null} |
| kandinsky6_lite | kandinsky_lite | 95.84 | 94.723 | 94.97 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 3, "style": 4, "audio_quality": null} |
| skyreels_a2v | skyreels_a2v | 200.29 | 199.681 | 23.41 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "lip_sync": 4, "reference_consistency": 4, "audio_quality": null} |
| wan_fun_control | wan_fun_control | 166.15 | 164.62 | 43.88 | {"adherence": 4, "motion": 3, "anatomy_face": 4, "temporal": 4, "style": 4, "control_alignment": 4} |

## Per-clip settings and review notes

- **esrgan_x2 / RealESRGAN_x2plus**: fp32 model/inference; native upscale node passes float32 tiles. deterministic chart 320x192,25frames,24fps; scale=2; no generative steps/cfg; seed=24601 label only. Technical motion-chart smoke only; four sampled frames inspected. Preserves gross color bars and motion; altered fine checkerboard texture and tiny numerals. Anatomy/lip sync not applicable; no added watermark observed. Coarse 5s VRAM sampling may miss short allocation peaks. Comfy execution 3.354s; ledger wall includes polling.
- **esrgan_x4 / RealESRGAN_x4plus**: fp32 model/inference; native upscale node passes float32 tiles. deterministic chart 320x192,25frames,24fps; scale=4; no generative steps/cfg; seed=24601 label only. Technical motion-chart smoke only; four sampled frames inspected. Preserves gross color bars and motion; altered fine checkerboard texture and tiny numerals. Anatomy/lip sync not applicable; no added watermark observed. Coarse 5s VRAM sampling may miss short allocation peaks. Comfy execution 3.336s; ledger wall includes polling.
- **ultrasharp / 4x-UltraSharp**: fp32 model/inference; native upscale node passes float32 tiles. deterministic chart 320x192,25frames,24fps; scale=4; no generative steps/cfg; seed=24601 label only. Technical motion-chart smoke only; four sampled frames inspected. Preserves gross color bars and motion; altered fine checkerboard texture and tiny numerals. Anatomy/lip sync not applicable; no added watermark observed. Coarse 5s VRAM sampling may miss short allocation peaks. Comfy execution 2.58s; ledger wall includes polling.
- **interp_rife / rife_vfi**: fp16 inference selected by native interpolation loader on CUDA capability>=8; checkpoint dtype may differ. source=320x192,25frames,24fps; multiplier=2; output49frames48fps; seed=24601. Technical chart smoke; four temporal samples visually checked. Gross geometry and progression retained; small moving blocks/numerals occasionally blend. Anatomy and lip sync not applicable; no added watermark observed. Real-source motion comparison deferred to V9.
- **interp_rife_v4.25 / rife_v4.25**: fp16 inference selected by native interpolation loader on CUDA capability>=8; checkpoint dtype may differ. source=320x192,25frames,24fps; multiplier=2; output49frames48fps; seed=24601. Technical chart smoke; four temporal samples visually checked. Gross geometry and progression retained; small moving blocks/numerals occasionally blend. Anatomy and lip sync not applicable; no added watermark observed. Real-source motion comparison deferred to V9.
- **interp_rife_v4.25_heavy / rife_v4.25_heavy**: fp16 inference selected by native interpolation loader on CUDA capability>=8; checkpoint dtype may differ. source=320x192,25frames,24fps; multiplier=2; output49frames48fps; seed=24601. Technical chart smoke; four temporal samples visually checked. Gross geometry and progression retained; small moving blocks/numerals occasionally blend. Anatomy and lip sync not applicable; no added watermark observed. Real-source motion comparison deferred to V9.
- **interp_rife_v4.26_heavy / rife_v4.26_heavy**: fp16 inference selected by native interpolation loader on CUDA capability>=8; checkpoint dtype may differ. source=320x192,25frames,24fps; multiplier=2; output49frames48fps; seed=24601. Technical chart smoke; four temporal samples visually checked. Gross geometry and progression retained; small moving blocks/numerals occasionally blend. Anatomy and lip sync not applicable; no added watermark observed. Real-source motion comparison deferred to V9.
- **interp_film / film_vfi**: fp16 inference selected by native interpolation loader on CUDA capability>=8; checkpoint dtype may differ. source=320x192,25frames,24fps; multiplier=2; output49frames48fps; seed=24601. Technical chart smoke; four temporal samples visually checked. Gross geometry and progression retained; small moving blocks/numerals occasionally blend. Anatomy and lip sync not applicable; no added watermark observed. Real-source motion comparison deferred to V9.
- **interp_gimm / gimm_vfi**: fp32 checkpoint and explicitly selected fp32 inference. source=320x192,25frames,24fps; multiplier=2; output49frames48fps; seed=24601. Technical chart smoke; four temporal samples visually checked. Gross geometry and progression retained; small moving blocks/numerals occasionally blend. Anatomy and lip sync not applicable; no added watermark observed. Real-source motion comparison deferred to V9.
- **tts / piper_lessac**: stock ONNX CPU model; internal operator precision not instrumented. CPU ONNX Lessac medium; generic synthetic voice; 22050Hz mono; 1.7066666666666668s; nominal filename seed24601 (no stochastic seed exposed by Piper CLI). Decoded mono WAV. Local official Whisper tiny.en CPU ASR recovered exact intended words Good morning, nice to see you. Adherence4 reflects recognized text only; perceived voice/audio quality unassessed. Visual scoring dimensions and visual flags inapplicable.
- **t2v / hunyuan15_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. steps=20,cfg=6,shift=9,seed=24601,1280x720,121frames,24fps; Qwen fp16. Reviewed 20 temporal samples across five seconds. Realistic adult woman moves along a sunlit street, turns toward the camera and smiles; natural hair/head motion and stable facial identity. Framing is medium/close, so foot-contact mechanics cannot be assessed. No doubled facial features, conspicuous added text, or watermark observed at contact-sheet scale. Small background details remain uncertain.
- **i2v / hunyuan15_i2v_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. cfg-distilled; steps20,cfg1,shift7,seed24601,1280x720,121frames,24fps; original fictional anime source from V1. 20-frame temporal contact review: two adult anime duelists retain identity and courtyard; readable single sword exchange, limited footwork and intermittent blade/contact ambiguity. Speed-line decoration appears; no visible watermark or text. Not full-frame playback assessment.
- **sr / hunyuan15_sr_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. encoded V1 source; nativeSR1920x1080,81frames,16fps; distilled8-step scheduler split4+4,noise0.7,cfg1,shift2,seed889 (template). Corrected nativeSR output1920x1072,81f16fps verified.20 temporal samples plus matched2.5s detail crop versus source/Lanczos display: motion/layout and cel style retained, edges crisper, but face/hair redrawn, jacket folds simplified, sword hilt/grip and wall texture changed. This invents/repaints detail rather than faithfully recovering source detail. Untiled high-resolution VAE decoding dominates long1111.56s wall; no second fix attempted or interruption used.
- **kandinsky_t2v / kandinsky6_pro**: BF16/F32 mixed checkpoint tensors; default BF16 diffusion inference (runtime log); fp16 Qwen encoder where used. native smoke 864x480,121frames,24fps,steps10,cfg1,seed24601; no beautifier or MagCache. 20-frame temporal review: realistic adult woman turns from rear view toward camera and smiles; face/hair and street remain coherent. Medium framing hides footwork so walking physics not fully assessed. Native44.1kHz audio stream5.062s matches5.042s video within AAC padding; sound quality/ambience not listened to or scored. No obvious visual artifact at contact-sheet scale.
- **i2v / kandinsky6_pro**: BF16/F32 mixed checkpoint tensors; default BF16 diffusion inference (runtime log); fp16 Qwen encoder where used. 864x480,121frames24fps,steps10,cfg1,seed24601; reference from K; native audio ambience. 20 temporal samples: canonical Mara face, auburn bob, teal jacket and neutral backdrop strongly retained. Natural blink, small outward head turn, return to camera and smile. Slight extra speech-like mouth activity toward end despite ambience prompt. No doubled face or obvious visual artifact at contact scale. Native audio duration verified; listening/semantic audio quality not assessed.
- **vsr / kandinsky6_vsr**: BF16/F32 mixed checkpoint tensors; default BF16 diffusion inference (runtime log); fp16 Qwen encoder where used. native 2.25x VSR;121frames24fps,distilled2-step checkpoint,5 scheduler points,cfg1,seed24601. Twenty temporal samples preserve the source walk, head turn, smile, face identity and street layout. Matched 2.5s crop shows clearer eye/nose edges, hair strands, jacket seam and background foliage than source Lanczos display, with subtle reconstruction/smoothing of skin and hair rather than exact detail recovery. No obvious doubled features or gross temporal collapse in samples. Actual output1952x1088 (rounded 2.25x),121f24fps; native audio retained, listening quality unassessed. This smoke is not the later controlled two-source 4K comparison.
- **seedvr_3b / seedvr2_3b_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. 320x192 to640x384,25frames,24fps,steps1,cfg1,seed24601. Four temporal samples of the diagnostic chart reviewed. Gross layout and motion survive, but tiny counters are rewritten and checkerboard texture becomes invented glyph-like patterns. This is a successful executable pipeline with a detail-faithfulness failure. Anatomy/lip sync not applicable. No added watermark observed.
- **seedvr_7b / seedvr2_7b_sharp_fp16**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. 320x192 to640x384,25frames,24fps,steps1,cfg1,seed24601. Four temporal samples of1.04s diagnostic chart: layout, color fields and moving shapes retained, less obvious counter/pattern corruption than the3B sample. Fine checkerboard still modified and at least one tiny counter glyph ambiguous/changed. No natural-scene/face conclusion; empty contact cells are unused.84.18s cold wall, not directly comparable to3B warm loading.
- **flashvsr / flashvsr_v11**: bf16 diffusion and low-quality projection inference explicitly set by installed FlashVSR native node. 320x192 to1280x768,25frames,24fps,scale4,seed24601,bf16; fixed official prompt embedding. Four sampled frames of short diagnostic chart: color fields and moving geometry retained, fine checkerboard softened and tiny counter edges imperfect. No natural-scene or face conclusion from chart. Blank contact cells are unused, not black video frames.
- **framepack_i2v / framepack**: bf16 explicitly selected base precision; SDPA. bf16;640x384,1s target30fps,steps30,cfg1,guidance10,seed24601,SDPA,TeaCache off. Five temporal samples across the actual 37 frames / 30fps (1.233s; requested 1s rounds to latent window) preserve both anime adult fighters and courtyard. Small parry motion with largely planted feet; strong blue energy/glow is invented around sword contact, absent from requested controlled ordinary parries. Black unused contact-sheet cells are padding, not black video frames. Short smoke only, no long-take claim.
- **longcat_t2v / longcat**: bf16 explicitly selected base precision; SDPA. bf16+distillLoRA unmerged;832x480,33frames15fps,steps12,cfg1,shift12,seed24601,SDPA; fix2 explicit Comfy RoPE (head-wise norm); fix1 corrected LoRA input. Reviewed nine temporal samples across the 2.2-second smoke. A coherent adult woman smiles in a close-up on a sunlit street. Small hair/head changes occur but the framing and body are almost stationary; this is not evidence of a readable walking gait or long-take ability. Face, plaid shirt and background remain stable at sampled times. Final corrective attempt executes after the LoRA-input and explicit Comfy RoPE fixes; only the short tested configuration is established.
- **mmaudio_foley / mmaudio**: fp16 explicitly selected model/feature precision. fp16;V1 anime duel original16fps5.0625s; feature input resampled25fps127frames;44kHz,steps25,cfg4.5,seed24601;fix1 prevents wrapper truncation. Corrected feature-frame rate: audio5.085s versus video5.0625s (AAC padding), non-silent RMS-28.50dBFS and sample peak1.124. Original visual source unchanged and previously inspected: stable two adult anime duelists with limited contact physics. Sound semantics, perceived quality and synchronization unassessed; adherence/audio scores deliberately null. Fix1 resolves premature audio truncation.
- **kandinsky6_lite / kandinsky_lite**: BF16/F32 mixed checkpoint tensors; default BF16 diffusion inference (runtime log); fp16 Qwen encoder where used. 864x480,121frames24fps,steps10,cfg1,seed24601. 20-frame contact review: plausible adult woman walking with camera tracking, turns toward camera and smiles; foreground face/hair coherent. Street/cars and distant people change less cleanly during the pan, signage unclear. Feet out of frame. Native44.1kHz audio duration aligns with video, but listening quality unassessed. This small-model smoke is visually viable but background stability weaker than Pro in these different samples.
- **skyreels_a2v / skyreels_a2v**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. fp8 scaled mixed;832x480,49frames25fps,4steps,cfg1,shift5,seed24601; SDPA; generic Piper; Transformers5 audio-capture compatibility fix1. Eight temporal samples across1.96s show one stable canonical Mara, natural blink/head motion, and mouth activity during the supplied generic Piper line with quieter start/end poses. Coarse lip-sync4 only; no phoneme/listening-quality assessment and no five-second or multi-person claim. No doubled features or duplicate character seen.49f25fps,832x480;AAC1.984s padding. First attempt failed in Transformers5 hidden-state capture; restored inherited capture_outputs decorator on custom wav2vec forward, CPU regression passed, GPU1 restarted idle, first corrected attempt successful. No dependency versions changed.
- **wan_fun_control / wan_fun_control**: fp8 checkpoint with Comfy mixed inference; adapter precision follows saved workflow/checkpoint. fp8 hi/lo+I2V Lightx2v;832x480,33frames16fps,4steps,cfg1,shift8,seed24601; Canny control from V1. Eight temporal samples from2.0625s/33f16 preserve cel-shaded two-adult duel and courtyard, with high parry/arm trajectory broadly matching the initial Canny-controlled source segment. Feet remain fairly planted and blade/contact detail is approximate. No doubled features seen at contact resolution; not a pixel/pose-error measurement. Blank contact cells are padding. Four-step I2VLightx2v used with FunControl as an explicit fast smoke, not a full-step quality claim.
LTX2.5 distilled/dev skipped because gated. HunyuanVideo-Avatar failed after two constrained dependency-install attempts; see SETUP_HISTORY and LEDGER.