Seedance2.5/MiniMaxH3測評影片摘要
認真研究了幾天, agent也配合24hr工作了幾天, 整理一些容易踩到的坑
也公布這次最後使用的提示詞, agent最終測試了16版才定稿
這篇文章的概念是一切調整先在0.2mp完成
最後才正式在0.7mp生成結果
不管是抽卡流 細節流 還是工程流, 如果能在0.2mp都先搞定一切
就能省下大量的時間
而目標是: 0.7mp要能重現0.2mp的構圖邏輯, 只有細節不同
如果只是把0.2mp改成0.7mp這麼簡單
也就不用特地寫出來, 實務面上至少要做以下幾件事
1.所有參數要一樣,包括seed和prompt
2.但ref_video的width是例外
從上一篇可知ref_video的解析度差異可讓運算速度差3倍
所以調整一個合理的解析度就非常重要
操作時會遇到兩種情境:
A. ref_video解析度非常高,假設是4K影片,那麼在沒處理的情況下會耗用大量記憶體實際上H3節點最高只能看到1mp左右, 也就是1344x768 這是它的adapt_canvas上限所以在產0.2mp時可以把custom_width設為720, 而0.7mp時設為1024
B. ref_video解析度很低,也就是我這次的情況,來自年代久遠的mv
在產0.2mp時由於解析度略高於0.2mp畫布,會被自動縮為0.2mp (這沒問題)
而當產0.7mp時由於解析度遠低於0.7mp畫布,
加上H3節點的關鍵原則是「小圖不放大」,所以低解析進0.7mp畫布會有填不滿的問題,導致構圖邏輯改變,我這次就是踩到這個坑 (很有問題)
當我把custom_width手動設為1024之後, 構圖邏輯就完全和0.2mp相同了
解析度與mp對應表格: (比值 1.0 附近未測;表外的組合請自行驗證。)
| ref_video 原始 |adapt_canvas| custom_width |參考畫布px |可安全支撐到|----------------------|------------|----------------|-----------|--------| 640×480(VGA) | 1024×768 | 1024 | 786,432 | 0.7 mp
| 704×480(DVD 寬螢幕) | 1120×768 | 1120 | 860,160 | 0.8 mp
| 720×480(NTSC DV) | 1152×768 | 1152 | 884,736 | 0.8 mp
| 1280×720 | 1344×768 | 1344 | 1,032,192 | 0.9 mp
| 1920×1080 | 1344×768 | 1344(省記憶體) | 1,032,192 | 0.9 mp
| 3840×2160(4K) | 1344×768 | 1344(省記憶體) | 1,032,192 | 0.9 mp
無論來源多大,參考畫布的天花板都是 1,032,192 px,所以 1.0mp 以上的輸出
在 ref2va 下無解, 構圖邏輯必然會改變。
想要更高畫質請用 upscaler,不要提高生成解析度。
關於模型:
unet: 優先使用pruned版本, 這是reddit討論群大家都認可的
clip: 跑得動int8版本的話優先使用, 會比nvfp4版本聰明
關於prompt:
這是我踩到的另一個坑, 或應該說是agent踩的..
上篇N大在推文中建議可以讓agent去閱讀官方的prompt guide
這樣生成的prompt會更精確
這是個實用技巧,我的agent是從零自己摸索,然後得到兩個結論:
1.否定句對H3無效
2.寫在prompt的外觀描述會出現在生成中,所以不要去描述你要換的對象外觀
最終在detailed_description只使用了39個英文字就生成了上一篇的結果
接著我把H3_Prompt_Guide餵給agent看, 要他根據內容來調整
他看完說Guide提到detailed_description最好要有350-500個字
於是開始大幅改寫, 根據Guide的規則多寫了很多東西進去
最終detailed_description介於367-616個字之間
總長度786-1266個字, 共6個版本
但產出結果卻異常的慘 全軍覆沒
於是我們回到精簡版: detailed_description 89個字, 總長度363
生成結果又回到原本水準上了
這次:
雖然我自己看也看得出來bug點很多, 但工程流純驗證就不花時間修
不抽卡seed都一樣, 0.2mp跑通後就上到0.7mp一次生成
再補一個用 2x跑出來的結果: (時間:1484.7s)
覺得不太值得, 似乎沒有多什麼細節, 反而動態變差
用過Scail2之後我才知道H3真的很聰明
H3分得出來銀幕上的人物要怎麼處理
舞台上舞台下多人時 景深和層次 怎麼安排
對嘴能力和臉部表情更是一流
Scail2則是只要遮罩框到一律換成真人
於是兩個大銀幕就給我跳出兩個巨人
最終版的Prompt:
subject_definitions:
[Shot 1] Replace the woman singing into a handheld microphone in <Video 1> with <Subject 1> (S1): a young woman with long straight silver-white hair fading to coral pink at the tips, straight blunt bangs, pale violet eyes, and fair smooth skin, as defined by <Picture 1> and <Picture 2>. She keeps the source performer's pose, motion, timing, and position in frame from <Video 1>. Everything else in <Video 1> is unchanged. (S1) is the only person who sings in the target video.
summary:
[reference generation + video editing] The target video is an edited version of <Video 1>. In [Shot 1], <Subject 1> replaces only the source performer described as singing woman from <Video 1>. The declared reference roles use <Subject 1>, <Picture 1>, <Video 1>.
retention_analysis:
<Subject 1> (appears in [Shot 1]): fully_preserved - the defined identity and appearance are preserved.<Video 1> ([Shot 1]): partially_preserved - the source timeline remains identifiable; only the declared character replacement is applied. Its source performance and timing plus scene, other people, lighting, camera, framing, and cuts are preserved.
detailed_description:
Cinematic live-action matching the source video.
[Shot 1] Using <Video 1> as the source timeline, replace only the source performer described as the woman singing into a handheld microphone with <Subject 1>. Apply <Subject 1>'s complete referenced identity and appearance, including wardrobe. Preserve that source performer's pose, motion, timing, gaze, expression, and physical interactions from <Video 1>. Keep every other person, the environment, props, lighting, camera movement, framing, and cuts from <Video 1> unchanged. Maintain seamless replacement identity continuity through occlusion, profile views, motion blur, and every source cut.
overall_soundscape:
The lead vocal heard throughout this shot is sung by (S1) alone. Her lips, jaw, and breathing move in exact sync with that lead vocal continuously, from the first frame to the last. Every other person on stage keeps the mouth position and facial behavior they have in <Video 1>. Synchronize physical sounds with the original soundtrack of <Video 1> and maintain exact audio-visual alignment with the source performance.
non_diegetic_music:
Reuse the complete original vocal song and background music from <Video 1> as the primary soundtrack without adding new music or extra vocal layers.
--