# Godfortune Seedance Quality Loop Seedance 2.0 is normalized to 100 as a reference-only video-quality target. Studio-Core does not call the Seedance API. The reference defines the required motion stability, physical plausibility, instruction following, micro-expression quality, multi-subject consistency, camera control, and temporal visual finish. ## Gate Order 1. Approve a production keyframe for each qualification shot. 2. Generate four I2V candidates from the same keyframe and prompt. 3. Reject hard failures before weighted scoring. 4. Shortlist two candidates with technical measurements. 5. Record human Seedance-relative scores for all eight categories. 6. Promote one winner per shot only when the total is at least 90 and no category is below 80. 7. Use the four approved winners in a ten-second integration test. ## First Local Candidate The first measured engine is Wan 2.2 TI2V 5B FP16 through ComfyUI native offloading. The official ComfyUI guide states that the 5B model should fit well on 8GB VRAM, while the upstream model supports 720p I2V at 24fps and uses the Apache 2.0 license. The first quality recipe is 704x1280, 121 frames, 24fps, 20 steps, CFG 5, `uni_pc`, `simple`, and model shift 8. No four-step speed LoRA or reduced-step distillation is allowed in the baseline. ## First Measured Calibration The first local shot completed in 1002.982 seconds. It is H.264 at 704x1280, 24fps, 121 frames, and 5.041667 seconds. Peak GPU allocation was 11911MiB and minimum available system memory was 0.855GiB. Music and video generation must therefore use one exclusive GPU queue; ACE-Step was stopped during Wan inference and restored after the ComfyUI models were unloaded. The internal visual calibration is 72.7/100 over the observable single-character categories. This is not a qualification score. The input was a padded research state rather than an approved full-scene keyframe, singing synchronization was not present, and multi-subject interaction could not be observed. The result proves that the engine is runnable and preserves the simple identity, but it does not clear the Seedance-relative production gate. ## Approved-Keyframe Run The user approved the 93.0-point HiDream neutral-front state on 2026-07-24. Studio composited that state with the approved 94.05-point HiDream stage at 704x1280, recorded the layer hashes and ground-contact placement, and reserved the lower caption-safe area. The first production-keyframe motion test uses one bounded action: one gentle sway, one presenting-hand raise, a smile, and a final hold with a locked camera. This isolates temporal identity, body mechanics, instruction completion, hand integrity, and background stability before song timing or camera motion is introduced. The previous 72.7 calibration remains failure-history evidence and is not reused as the production input. The first approved-keyframe candidate scored a provisional 81.3/100 over the observable categories. The character, stage, and locked composition remained recognizable, but the left arm rose despite a right-hand-only instruction and finger topology changed during the gesture. It is evidence of an 8.6-point improvement over the research calibration, not an approved motion asset. The second candidate keeps the same approved keyframe, model, resolution, sampler, steps, and camera. Its defect-control prompt locks the left arm, requests only a right-forearm presenting gesture, keeps both eyes open, and explicitly forbids the two-arm pose and finger failures observed in the first candidate. This run determines whether prompt control is sufficient or whether the pipeline must introduce an approved end-pose or layered animation path. The second candidate scored a provisional 84.2. It kept the left arm down and reduced the hand failure, but ignored the open-eye instruction and did not hold the final gesture. Prompt-only full-scene I2V therefore remains below the production target. ## Separated Character Motion The third candidate allocates 900 vertical pixels to the approved character on a flat blue field, generates only the character motion, then composites it over the exact approved background. FFmpeg combines a blue-screen alpha mask with a spatially limited source-shadow mask, scales the generated canvas by 0.8, and places the feet on the approved 825-pixel anchor. The source background never enters the final shot. The generated arm rise completed too quickly, so the first 0.75 seconds were retimed to 1.25 seconds and 0.5 seconds of redundant hold were removed. The final video remains 121 frames, 24fps, and 5.041667 seconds. No generative repaint, frame interpolation, or character identity edit occurs during this timing pass. This separated composite initially produced a 94.1 stability diagnostic over the observable single-character categories. Human review correctly identified that the score was inflated by low task complexity: the shot has one neutral expression and one simple arm gesture. It does not contain a speaking or singing face, a comedy reaction, or multiple independent motion beats. The 94.1 value is therefore retained only as a separated-composite stability measurement. Its qualification score is null and its complexity gate fails. Future candidates must satisfy the shot's expression-count, motion-beat, vocal face, and start/end eye-topology requirements before weighted Seedance-relative scoring begins. ## Complexity And Scale Correction Human review found that the first approved-keyframe candidate contained more varied and natural motion than the separated stability candidate, but its 720-pixel character height was too large for the approved stage and its opening and closing eye images did not match. These are independent quality dimensions: motion variety cannot excuse composition scale or identity drift, and a simple stable gesture cannot outrank a harder expressive shot without a complexity gate. Studio now renders a 540, 600, 660, and 720-pixel scale comparison at the same 825-pixel feet anchor. The 720-pixel placement is rejected and 600 pixels is the working target for the next full-stage candidate. The next complexity-matched shot must visibly complete all of these beats before weighted scoring: 1. neutral start with both round eyes open; 2. a right-palm presentation while the left arm remains down; 3. a surprised mouth-and-brow expression with a small torso recoil; 4. a return to the original neutral eye topology and face. Missing expression states, missing motion beats, or different opening and closing eyes cause a hard failure and leave the qualification score null. The first 600-pixel complexity-matched run completed the expression and motion count, and the reduced scale fit the stage materially better. It still failed before scoring: dusty-pink hair became orange, the right hand became an orange object-like form, the opening round eyes became larger outlined eyes, and the opening smile returned as a different worried mouth. The eye-region endpoint MAE was 0.201739, retained only as a diagnostic; human topology review owns the failure decision. This establishes the next architecture boundary. Wan 2.2 5B may supply torso timing and non-critical transitional deformation, but it cannot own exact identity restoration in an expressive shot. The face, eye, mouth, and hand states must come from approved layers or approved endpoint states. The approved background remains static and the 600-pixel full-stage scale remains the working target. The first layer-locked candidate reused the stable separated right-arm motion, played the raise forward and backward, inserted a fixed O-mouth and eyebrow layer, applied a whole-character recoil, and ended on the opening frame. It passed the visual expression-count and endpoint identity checks, but human review rejected it. The eyebrow layer introduced anatomy that the source hairstyle hides, and the retimed, reversed, and held motion was not natural. Temporal diagnostics confirmed the review. The layer-locked candidate had a 0.683333 frozen frame-pair ratio and concentrated its discontinuities around the expression overlay. The research calibration had no frozen pairs and the first approved-keyframe candidate had a 0.016667 frozen-pair ratio. Low pixel motion is no longer treated as a quality advantage. ## Direct Temporal Generation And Local Identity Repair The next candidate rebuilt the approved stage at a 600-pixel character height and restored direct Wan generation for all 121 frames. It used the same 20-step, 24fps, `uni_pc`, no-speed-LoRA recipe as the smoother historical candidates. No generated segment was stretched, reversed, frozen, or assembled. Direct generation restored continuous motion and retained the round eye topology, but Wan drew eyebrow arcs during the surprise expression despite an explicit hidden-eyebrow instruction. Prompting alone therefore does not enforce this character invariant. The repair candidate preserves the direct Wan frame order and timestamps and suppresses only the two localized eyebrow regions after the arcs are detected. It does not modify the eyes, mouth, hands, or body motion. Its frozen frame-pair ratio is 0.083333, compared with 0.683333 for the rejected layer-locked candidate. The frame-delta diagnostic finds no late discontinuities after frame 14. This establishes the current production direction: 1. generate the complete shot motion directly in Wan; 2. keep the approved 600-pixel scene composition; 3. repair only localized identity violations; 4. never build primary body timing from reverse playback, hard holds, or a suddenly enabled expression layer; 5. keep the score null until human review confirms that the repair is invisible at playback speed. The current candidate still holds its surprised mouth and raised hand instead of returning to neutral, and the requested torso recoil is subtle. Those are instruction-adherence issues for the next shot design, not reasons to discard the recovered temporal architecture. ## Boundaries - A good motion model cannot repair a weak input keyframe. - The same approved input, prompt, resolution, duration, and output probe are used for every engine comparison. - Runtime is measured but is not part of the quality score. - Automatic image similarity and frame-difference metrics are diagnostics only. - Human review owns expression, comedy timing, action completion, and the final Seedance-relative score. - Seedance reference assets cannot be represented as locally generated output. Machine-readable policy: `config/godfortune_seedance_benchmark_v1.json`. Primary sources: - - - -