# V10 — complete

12 verified and explicitly reviewed planned outputs; 2 gated cells skipped; 0 terminal job exceptions. Historical setup observations and superseded attempts are preserved in SETUP_HISTORY.md and LEDGER.

## Final reviewed results

All12 available-method/source cells completed and were reviewed at final3840x2160: two fixed V1 sources (Hunyuan walk/turn with face detail, Wan Lightning studio dance) through Hunyuan1.5 SR, Kandinsky VSR, SeedVR2 7B-sharp, FlashVSR1.1, RealESRGANx4plus and deterministic Lanczos. Two LTX upscaler cells are gated/skipped. Reviews include20-frame temporal contacts and matched2.5s source/result crops. No true high-resolution ground truth exists, so sharper generated texture must not be described as recovered fact.

For the face example, Hunyuan SR and Kandinsky VSR provide stronger apparent eye/hair/clothing detail; SeedVR2 gives a more restrained improvement close to source structure. On the dance, Kandinsky and SeedVR2 keep the original composition and pose timing with useful definition; Hunyuan adds more conspicuous face/beard and room-detail reconstruction. FlashVSR is sharp but changes some fine features and uses a different input regularization, reducing comparability. RealESRGAN offers modest definition with waxy/painted flattening, particularly in the face clip background. Lanczos preserves source content best but adds no genuine detail. Moving hand blur persists across methods; upscale is not a cure for source motion/anatomy defects. These are two-source observations, not universal model rankings.

The chains are materially different. Hunyuan uses native nominal1080p SR (actual1920x1072) then Lanczos4K; Kandinsky uses its native VSR target1080p then Lanczos4K; SeedVR2 restores directly at3x4K; FlashVSR first regularizes1280x720 to960x544, performs4x, then finishes exact4K, with a slight crop/geometry change; ESRGAN uses framewise4x followed by reduction to4K. The deterministic control uses the unmodified source. Do not label these as equivalent native4K generation or an identical-input-resolution benchmark.

Measured request wall seconds, face/dance: Hunyuan873.12/406.41; Kandinsky492.80/297.34; SeedVR2 727.65/458.24; FlashVSR2798.56/1882.26; ESRGAN197.71 for face (dance exact timing in table below). Lanczos baseline processing was about5s per clip; final export/verification is additional for every path. FlashVSR was the slowest tested workflow at this size: streaming inference plus a lengthy CPU wavelet color-correction stage, not an execution failure. The small smoke chart did not predict this full-video cost. Device peaks include resident caches, and loading/cache behavior differs, so times and memory are observations for these graphs, not isolated theoretical model requirements.

All final exports use common lossless GPU1 NVENC H264/yuv420p/faststart finishing. An81-frame control confirmed decoded-pixel equality against the same Lanczos resize, avoiding an extra lossy-export confound. Earlier CRF18 exports were archived and remastered from original native/source frames; no inference was repeated for that repair. Intermittent CPU x264 initialization stalls motivated the export workaround; their root cause remains unresolved. Model request wall time, Comfy execution time, collection and final resize/verification time are recorded separately. The workflow outputs themselves may already be compressed; lossless final finishing does not undo upstream compression. Native stage files and pre-final metadata remain available for audit.

Suggested write-up comparison: Lanczos versus SeedVR2 versus Kandinsky on the matched face and dance crops, with Hunyuan and FlashVSR as examples of stronger detail reinterpretation and the sharpness/time tradeoff. Inspect the originals for temporal artifacts; contact sheets are sampled evidence, not a guarantee of frame-by-frame stability.


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 |
|---|---|---:|---:|---:|---|
| face | hunyuan15_sr | 873.12 | 872.3 | 75.52 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "sharpness": 4, "source_preservation": 4} |
| face | kandinsky_vsr | 492.8 | 491.047 | 94.97 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "sharpness": 4, "source_preservation": 4} |
| face | seedvr2_7b | 727.65 | 725.99 | 83.76 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "sharpness": 4, "source_preservation": 4} |
| face | flashvsr | 2798.56 | 2796.92 | 84.51 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "sharpness": 4, "source_preservation": 3} |
| face | esrgan | 197.71 | 196.91 | 7.41 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 3, "sharpness": 3, "source_preservation": 3} |
| face | lanczos | 4.997 |  | CPU | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "sharpness": 3, "source_preservation": 5} |
| dance | hunyuan15_sr | 406.41 | 405.538 | 79.61 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 3, "style": 4, "sharpness": 4, "source_preservation": 3} |
| dance | kandinsky_vsr | 297.34 | 295.822 | 67.18 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "sharpness": 4, "source_preservation": 4} |
| dance | seedvr2_7b | 458.24 | 457.111 | 85.42 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "sharpness": 4, "source_preservation": 4} |
| dance | flashvsr | 1882.26 | 1881.256 | 66.7 | {"adherence": 3, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "sharpness": 4, "source_preservation": 3} |
| dance | esrgan | 129.12 | 128.386 | 6.27 | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 3, "sharpness": 3, "source_preservation": 4} |
| dance | lanczos | 5.061 |  | CPU | {"adherence": 4, "motion": 4, "anatomy_face": 4, "temporal": 4, "style": 4, "sharpness": 3, "source_preservation": 5} |

## Per-clip settings and review notes

- **face / hunyuan15_sr**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. source=V1_S2_M1_hunyuan15_24601.mp4; native1920x1072 (nominal1080p) SR8steps split4+4,noise0.7,cfg1,shift2,seed24601;tiledVAE512 overlap64 temporal64 overlap8; thenLanczos3840x2160. Reviewed20 temporal samples and matched2.5s detail crop against source displayed with Lanczos. Native1920x1072 SR followed by deterministic3840x2160 finishing sharpens hair strands, eyes, skin and jacket seams substantially. It also redraws small facial marks, shirt folds/neckline and background car/building detail; added detail is generated rather than recovered ground truth. Identity, camera trajectory and scene geometry remain broadly consistent across samples, with no obvious tile seam or doubled facial features. No major temporal instability visible at contact scale. Native1072 height introduces a slight aspect adjustment during exact4K finishing. Export revision: regenerated from original native/source frames using Lanczos and validated GPU1 NVENC losslessH264; inspected fresh2.5s matched crop. Earlier temporal assessment applies to unchanged model/source sequence. No model inference repeated; prior CRF18 master archived.
- **face / kandinsky_vsr**: BF16/F32 mixed checkpoint tensors; default BF16 diffusion inference (runtime log); fp16 Qwen encoder where used. source=V1_S2_M1_hunyuan15_24601.mp4; nativeVSR5steps,seed24601,target1080p; thenLanczos3840x2160. Reviewed20-frame contact and matched source/result2.5s face crop at the final3840x2160 export. Clearer eyes, lip edges, hair strands and jacket seams than the source enlarged for display; facial identity, turn and street composition remain coherent. Fine hair, skin and blouse/button details are reconstructed and slightly changed, so this is plausible generated detail rather than recovered ground truth. Some smoothing/painterly background detail remains; no conspicuous doubled facial features in samples. Native Kandinsky VSR stage is followed by deterministic exact4K resize, not native4K generation.
- **face / seedvr2_7b**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. source=V1_S2_M1_hunyuan15_24601.mp4; 7B-sharp fp16,tiledVAE,1step,cfg1,seed24601,3x to3840x2160. Reviewed20-frame contact and matched2.5s face crop. Facial outline, gaze, mole, hair shape and clothing remain close to source; eye/lip/hair edges gain definition while skin and background remain softly smoothed. Improvement is restrained in this crop, with less conspicuous reconstruction than the stronger native-SR examples. Fine texture is still generated and not ground-truth recovery; no conspicuous doubled face or temporal structural break in samples. Direct3x SeedVR2 restoration yields final3840x2160.
- **face / flashvsr**: bf16 diffusion and low-quality projection inference explicitly set by installed FlashVSR native node. source=V1_S2_M1_hunyuan15_24601.mp4; v1.1 bf16,seed24601,regularized960x544 input,x4 thenLanczos3840x2160; source downsample/crop disclosed. Reviewed20-frame contact and matched2.5s crop. Face, turn and outfit remain recognizable with clearly sharper eye/lip/hair and jacket edges. Fine hair texture has a faint grid-like/artificial pattern, and blouse/button detail is noticeably reinterpreted. The pipeline first regularizes the source to960x544 before4x upscaling and exact3840x2160 finishing, so framing/geometry differs slightly from the other paths; not an identical-input-resolution comparison. No conspicuous doubled facial features in samples. Long measured request includes native streaming inference and CPU wavelet color correction; this tested path is not fast at this size.
- **face / esrgan**: fp32 model/inference; native upscale node passes float32 tiles. source=V1_S2_M1_hunyuan15_24601.mp4; RealESRGANx4plus native4x thenLanczos3840x2160. Reviewed20-frame contact and matched2.5s face crop. Identity, turn and broad clothing layout remain intact, but the face gains relatively little useful fine detail. Background building/window and car shapes become flattened, waxy and outlined compared with the source, creating a painted appearance. Some edge sharpening is visible without the convincing facial detail of the stronger SR examples. Native framewise4x output is resized down to exact3840x2160; no conspicuous doubled face in samples.
- **face / lanczos**: deterministic image processing; no learned-model precision. CPU ffmpeg Lanczos3840x2160 from V1_S2_M1_hunyuan15_24601.mp4;no model inference;seed in filename is source bank only. Reviewed20 temporal samples and matched2.5s source/output detail crops. Exact3840x2160 deterministic Lanczos control preserves source identity, geometry, motion and style with no evident added structures or temporal instability. Fine detail remains source-limited; this resize does not recover genuine new texture. Original framing/action limitations remain. Source crop is displayed with Lanczos at common coordinates, so close visual agreement is expected, not independent evidence of super-resolution quality. Export revision: regenerated from original native/source frames using Lanczos and validated GPU1 NVENC losslessH264; inspected fresh2.5s matched crop. Earlier temporal assessment applies to unchanged model/source sequence. No model inference repeated; prior CRF18 master archived.
- **dance / hunyuan15_sr**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. source=V1_M2_wan22_lightning_24601.mp4; native1920x1072 (nominal1080p) SR8steps split4+4,noise0.7,cfg1,shift2,seed24601;tiledVAE512 overlap64 temporal64 overlap8; thenLanczos3840x2160. Reviewed20 temporal samples and matched2.5s source/output detail crop. Motion sequence and full-body composition broadly follow the source. SR sharpens face, shirt, bricks and window edges but noticeably redraws the face/beard/hair, clothing folds and wall geometry; white pillars and individual bricks change thickness/shape. Face/mouth details vary during action, reducing faithful preservation despite clean anatomy. No extra person or obvious tile boundary. Native1920x1072 is finished at3840x2160 with Lanczos and GPU1 NVENC losslessH264; exact81-frame decoded-pixel equality to the resize was validated, so the detail changes originate upstream in SR rather than the final encoder.
- **dance / kandinsky_vsr**: BF16/F32 mixed checkpoint tensors; default BF16 diffusion inference (runtime log); fp16 Qwen encoder where used. source=V1_M2_wan22_lightning_24601.mp4; nativeVSR5steps,seed24601,target1080p; thenLanczos3840x2160. Reviewed20-frame contact and matched source/result2.5s crop. Dance timing, pose, clothing and room geometry remain broadly preserved. Facial/hair edges and brick/pillar texture are crisper, with mild regenerated texture and smoothing; the moving hand remains blurry rather than gaining reliable finger detail. No conspicuous extra limb or doubled face in sampled frames. Native VSR followed by deterministic4K resize adds apparent detail, not evidence of true recovered scene information.
- **dance / seedvr2_7b**: fp16 checkpoint/inference; see saved workflow for encoder and VAE variants. source=V1_M2_wan22_lightning_24601.mp4; 7B-sharp fp16,tiledVAE,1step,cfg1,seed24601,3x to3840x2160. Reviewed20-frame contact and matched2.5s source/result crop. Pose timing, face, shirt folds and room layout remain close to source. Beard/hair, trouser edges and masonry gain fine definition; the blurred moving hand is still soft, and some brick/skin texture is reconstructed. No conspicuous extra limb or doubled face in samples. Useful restrained4K restoration in this example, without evidence that generated texture represents true missing detail.
- **dance / flashvsr**: bf16 diffusion and low-quality projection inference explicitly set by installed FlashVSR native node. source=V1_M2_wan22_lightning_24601.mp4; v1.1 bf16,seed24601,regularized960x544 input,x4 thenLanczos3840x2160; source downsample/crop disclosed. Reviewed20-frame contact and matched2.5s source/result crop. Dance and room composition remain recognizable; face/beard, fabric and brick edges become sharper, with stronger invented masonry texture. Moving hand remains soft. No conspicuous doubled face or extra limb in sampled frames. Input was first regularized to960x544, so slight crop/geometry changes complicate exact source preservation; final output is3840x2160 after4x and finishing resize. Long request includes CPU wavelet color correction and is much slower than other tested paths on this source.
- **dance / esrgan**: fp32 model/inference; native upscale node passes float32 tiles. source=V1_M2_wan22_lightning_24601.mp4; RealESRGANx4plus native4x thenLanczos3840x2160. Reviewed20-frame contact and matched2.5s source/result crop. Dance pose and room layout remain close to source. Face, shirt and pillar edges gain some definition, while fine brick/skin texture is smoothed into broad waxy surfaces rather than richly recovered detail. Moving hand stays blurred. No conspicuous extra limb or doubled face in samples; framewise native4x result is reduced to exact4K. Less detailed than the stronger restoration examples, but broad source structure remains intact.
- **dance / lanczos**: deterministic image processing; no learned-model precision. CPU ffmpeg Lanczos3840x2160 from V1_M2_wan22_lightning_24601.mp4;no model inference;seed in filename is source bank only. Reviewed20 temporal samples and matched2.5s source/output detail crops. Exact3840x2160 deterministic Lanczos control preserves source identity, geometry, motion and style with no evident added structures or temporal instability. Fine detail remains source-limited; this resize does not recover genuine new texture. Original framing/action limitations remain. Source crop is displayed with Lanczos at common coordinates, so close visual agreement is expected, not independent evidence of super-resolution quality. Export revision: regenerated from original native/source frames using Lanczos and validated GPU1 NVENC losslessH264; inspected fresh2.5s matched crop. Earlier temporal assessment applies to unchanged model/source sequence. No model inference repeated; prior CRF18 master archived.