bitGen StudioTwo sets of real-card numbers, and what bitGen Studio changed because of them.
== RTX 3090: a tester's blind benchmark ==
MiniMax H3 in ComfyUI (int8 main model, ComfyUI 0.42). The same picture and the same seed, today's settings against the speed-up nodes; two rounds of 32 pairs, judged blind by two AI judges that had not seen the code.
Sampling time, today's settings -> with the speed-ups:
- Turbo v4, full size: 796 s -> 355 s (Veda)
- PDD4, full size: 398 s -> 177 s
- 12.25 s long shot: 628 s -> 222 s
- 20-step two-person lip-sync: 1389 s -> 828 s
- Short fight scene: 108 s -> 84 s
- Lip-sync with Spectrum: 1403 s -> 717 s
- Long shot, KJ head-chunk attention + Sage: 628 s -> 521 s
His verdicts:
- Veda sparse attention on full-size, long and dialogue shots: passed - 41-65% faster and the judges did not prefer the original.
- Veda on fights and transformations: not worth it - only 22% faster and the picture got softer.
- Spectrum on 20-step lip-sync: twice as fast, but faces blur.
- KJ head-chunk attention + Sage on a long shot: only 17% faster.
- bitGen's Qwen-Image still ran 25 steps where his Turbo runs 8.
What bitGen Studio 2.717.0 changed because of it:
- Veda skips a shot whose prompt describes a fight, a transformation, an explosion or a chase (tick "Veda on action shots too" in the video style to keep it).
- With Long video, Veda now takes the place of the low-VRAM attention instead of being left out.
- Spectrum keeps the first 2 and the last 3 steps real on runs of 12 steps or more, so faces stay sharp.
- A new Qwen Image 2.1 edit Turbo8 style: 8 steps instead of 25. Qwen Image 2.1 and the Turbo8 LoRA are non-commercial (Qwen Research License). The scene cards' First frame defaults to Qwen Image Edit 2511 in 4 steps (Apache 2.0 - commercial use is fine) when it is installed.
== RTX 5090 (32 GB): measured by the developer ==
- Qwen Image 2.1 edit, 25 steps: 16 s a picture
- Qwen Image 2.1 edit Turbo8, 8 steps: 11 s a picture
- MiniMax H3 Fast Omni, 540p: a 15 s clip in 1 min 48 s, a 10 s clip in 1 min 12 s
Reading them:
- Turbo8 cuts the steps 3x but the time only 1.5x. About 8-9 s of every picture is fixed work - loading the model, the 8B text encoder, decoding, the upload - and sampling is about 0.3 s a step. Part of the fixed time was reloading the model before every render; since bitGen 2.722.0 the model stays on the card while the renders in a row use it (Settings > ComfyUI > Free it: when the model changes - the default).
- Fast Omni works out to 7.2 s of rendering per second of video at both lengths - it scales in a straight line, no memory swapping.
- These are not like-for-like with the 3090 chart: other models (Fast Omni against the standard H3 graphs), 540p against full size, and the chart gives no clip lengths. For a true card-to-card comparison, render the same full-size Turbo v4 clip with Veda on and off.
== Update: a larger picture, the model kept on the card (bitGen 2.722.0) ==
The same RTX 5090, a larger picture, the first run -> the second run in a row:
- Qwen Image 2.1 edit, 25 steps: 39 s -> 33 s
- Qwen Image 2.1 edit Turbo8, 8 steps: 17 s -> 15 s
Reading them:
- At this size a step takes about 1 s (about 0.3 s on the smaller picture). Step time grows faster than the picture, so Turbo8's saving grows with it: 2.2x faster here (33 s -> 15 s) against 1.5x on the smaller picture.
- The fixed part (loading, the text encoder, decoding, the upload) stays at about 6-7 s whatever the size.
- The second run is quicker because 2.722.0 keeps the model on the card between renders that use it: 6 s saved on Qwen edit, 2 s on Turbo8. The first run includes loading the model.
Measured your own card? Reply with the card, the model, the size, the clip length and the time.
== 繁體中文 ==
兩組實機數據，以及 bitGen Studio 因此做的修改。
== RTX 3090：一位測試者的盲評 ==
ComfyUI 上的 MiniMax H3（int8 主幹、ComfyUI 0.42）。同圖同 seed，對照現行設定與加速節點；兩輪 32 對，由兩位沒看過程式的 AI 盲評。
取樣秒數，現行 -> 加速節點：
- Turbo v4 全尺寸：796 s -> 355 s（Veda）
- PDD4 全尺寸：398 s -> 177 s
- 12.25 s 長場：628 s -> 222 s
- 20 步雙人對嘴：1389 s -> 828 s
- 打鬥短場：108 s -> 84 s
- 對嘴 Spectrum：1403 s -> 717 s
- 長場 KJ 切頭注意力 + Sage：628 s -> 521 s
他的結論：
- Veda 稀疏注意力用在全尺寸、長場、對話：通過 - 快 41-65%，盲評沒輸。
- Veda 用在打鬥、變形這種大動作：不值得 - 只快 22%，畫面也變軟。
- Spectrum 用在 20 步對嘴：快一倍，但臉會糊。
- KJ 切頭注意力 + Sage 用在長場：只快 17%。
- bitGen 的 Qwen-Image 還是 25 步，他的 Turbo 只要 8 步。
bitGen Studio 2.717.0 因此做的修改：
- 提示詞描述打鬥、變形、爆炸或追逐的鏡頭，Veda 會跳過（在影片風格勾選「Veda on action shots too」可保留）。
- Long video 時，Veda 現在取代省 VRAM 的注意力，而不是被略過。
- 12 步以上的執行，Spectrum 保留前 2 步與最後 3 步實際計算，臉保持清晰。
- 新的 Qwen Image 2.1 edit Turbo8 風格：8 步取代 25 步。Qwen Image 2.1 與 Turbo8 LoRA 僅限非商業用途（Qwen Research License）。安裝了 Qwen Image Edit 2511 時，場景卡片的 First frame 預設用它以 4 步生成（Apache 2.0，可商用）。
== RTX 5090（32 GB）：開發者實測 ==
- Qwen Image 2.1 edit，25 步：每張 16 秒
- Qwen Image 2.1 edit Turbo8，8 步：每張 11 秒
- MiniMax H3 Fast Omni，540p：15 秒片段 1 分 48 秒，10 秒片段 1 分 12 秒
怎麼看：
- Turbo8 步數少了 3 倍，時間只少 1.5 倍。每張約有 8-9 秒是固定工作（載入模型、8B 文字編碼器、解碼、上傳），取樣每步約 0.3 秒。固定時間的一部分是每次算圖前重新載入模型；從 bitGen 2.722.0 起，連續算圖使用同一模型時，模型會留在顯示卡上（Settings > ComfyUI > Free it：when the model changes，預設）。
- Fast Omni 兩種長度都是每 1 秒影片約 7.2 秒算圖 - 線性成長，沒有記憶體交換。
- 這些數字和 3090 的圖表不能直接比：模型不同（Fast Omni 對標準 H3 流程）、540p 對全尺寸，而且圖表沒有片段長度。要真正比較兩張卡，請用同一段全尺寸 Turbo v4 片段，分別開關 Veda 來算。
== 更新：較大的圖片，模型留在顯示卡上（bitGen 2.722.0）==
同一張 RTX 5090，較大的圖片，第一次 -> 連續的第二次：
- Qwen Image 2.1 edit，25 步：39 秒 -> 33 秒
- Qwen Image 2.1 edit Turbo8，8 步：17 秒 -> 15 秒
怎麼看：
- 這個尺寸每步約 1 秒（較小的圖約 0.3 秒）。每步時間增加得比圖片大小更快，所以 Turbo8 省下的時間也跟著變多：這裡快 2.2 倍（33 秒 -> 15 秒），較小的圖只快 1.5 倍。
- 固定部分（載入、文字編碼器、解碼、上傳）不論尺寸都約 6-7 秒。
- 第二次比較快，是因為 2.722.0 在連續算圖使用同一模型時把模型留在顯示卡上：Qwen edit 省 6 秒，Turbo8 省 2 秒。第一次包含載入模型的時間。
也測了自己的卡嗎？請回覆顯示卡、模型、尺寸、片段長度和時間。