產品廣告影片
上傳最多 3 張產品參考圖 一鍵生成產品電影感廣告影片。上傳最多 3 張產品圖、設定長度、分鏡數量,並可加入自訂說明(選填)。LLM 會處理所有資訊並產生結構化提示詞,讓 MiniMax H3 生成一致且符合品牌形象的產品影片。
關於
此 ComfyUI 工作流程可根據最多三張參考圖像,製作電影感多分鏡產品廣告影片。在 Inputs 群組中,Image 1 作為主視覺/環境參考(此場景會貫穿所有分鏡),Images 2–3 則作為產品細節參考。長度與分鏡數量由 PrimitiveInt 節點控制,品牌/風格說明可填入 PrimitiveStringMultiline。LLM Prompt Generation 群組會用 GeminiNodeV2 將你的輸入轉換為單一、帶時間碼的提示詞,MiniMax H3 可直接理解(例如:0–3 秒為分鏡 1,3–6 秒為分鏡 2 等)。結構化提示詞與參考圖像會一併送入 MiniMax H3 節點,產生一致且符合品牌的影片。
此流程的效率來自「單提示多分鏡」設計。MiniMax H3 只需一個包含所有分鏡與時間區間的提示詞,GeminiNodeV2 會自動格式化分鏡清單,包含鏡頭運動、轉場與產品重點。你可用 PreviewAny 與 MarkdownNote 預覽腳本內容。確認無誤後,將最終腳本貼到 Edit Prompt? 群組,即可「鎖定」腳本、跳過 LLM 步驟,快速重複產出。Generate and Save Video 群組會渲染廣告並透過 SaveVideo 輸出,讓你快速建立一致的產品影片產線。
您提供的內容
- 正在讀取工作流程
在 ComfyUI 中執行。 此工作流程無法在此頁面上執行。請在已載入所有內容的 Comfy Cloud 中開啟,或下載到您自己的 ComfyUI。
在 Cloud 中試用工作流程內部
節點圖的唯讀預覽。請在 Comfy Cloud 中開啟,或下載後編輯節點。
Loading the graph
Open full-size workflow preview從程式碼使用此工作流程
可行時會根據工作流程圖和輸入內容產生;否則會顯示儲存的範例。執行可能產生費用的程式碼前,請檢查認證資訊和端點。
What you need
- Runs on
- Comfy Cloud
- 金鑰
- COMFY_API_KEY
需要付費 Cloud 方案與點數。
# Python 3.10+: pip install comfy-sdk==0.4.0
import os
from comfy_sdk import Comfy
api_key = os.environ.get("COMFY_API_KEY", "").strip()
if not api_key: raise RuntimeError("Set COMFY_API_KEY")
os.environ.setdefault("COMFY_BASE_URL", "https://cloud.comfy.org")
client = Comfy(api_key=api_key)
workflow = client.workflows.from_json({
"4": {
"class_type": "LoadImage",
"inputs": {
"image": "c8696e8418db0e28c85df8c77d45ccead226c715bdef0a31829d27dbc882d999.png"
},
"_meta": {
"title": "Load Start Frame Image"
}
},
"8": {
"class_type": "LoadImage",
"inputs": {
"image": "13666a63b947d427aeb28f03fddfc28e268253f0c57346147525a32ebda49f8a.png"
},
"_meta": {
"title": "Load Product Image 1"
}
},
"12": {
"class_type": "LoadImage",
"inputs": {
"image": "554988911d94eb13700712629e48101e44456bc2bafb52cd8ca347a49cb25371.png"
},
"_meta": {
"title": "Load Product Image 2"
}
},
"22": {
"class_type": "PrimitiveInt",
"inputs": {
"value": 10
},
"_meta": {
"title": "Total Seconds (4-15)"
}
},
"23": {
"class_type": "PrimitiveInt",
"inputs": {
"value": 5
},
"_meta": {
"title": "Number of Shots"
}
},
"164": {
"class_type": "GeminiNodeV2",
"inputs": {
"model": "Gemini 3.1 Pro",
"model.images.image_1": [
"4",
0
],
"model.images.image_2": [
"8",
0
],
"model.images.image_3": [
"12",
0
],
"model.max_output_tokens": 32768,
"model.temperature": 1,
"model.thinking_level": "HIGH",
"model.top_p": 0.95,
"prompt": [
"81:85",
0
],
"seed": 111111,
"system_prompt": [
"309",
0
]
},
"_meta": {
"title": "Google Gemini"
}
},
"165": {
"class_type": "PrimitiveStringMultiline",
"inputs": {
"value": ""
},
"_meta": {
"title": "Prompt Input (copy/paste prompt here to make edits)"
}
},
"166": {
"class_type": "PreviewAny",
"inputs": {
"source": [
"164",
0
]
},
"_meta": {
"title": "Preview Generated Prompt"
}
},
"200": {
"class_type": "PrimitiveStringMultiline",
"inputs": {
"value": ""
},
"_meta": {
"title": "User Story / Creative Direction (optional)"
}
},
"309": {
"class_type": "PrimitiveStringMultiline",
"inputs": {
"value": "You are a creative director and video prompt specialist for high-end product ads. Your job is to analyze the provided product reference images and write ONE complete multi-shot video prompt (a reference-to-video model), in the sleek, editorial style of luxury product advertising (think Aesop, Dior, etc.).\n\n---\n\n## INPUT\nYou will receive:\n- Reference images. <Picture 1> is the hero/environment reference — its scene, surface, lighting, and atmosphere are the locked environment for the entire film. Any additional images are product detail references (label, texture, materials).\n- {N_SHOTS}: the number of distinct shots (1-8)\n- {N_SECONDS}: the total video duration in seconds (5-15). This is a hard constraint.\n- {USER_STORY}: optional creative direction. If provided, it overrides or augments your inferred vibe. If empty, rely entirely on your image analysis.\n\n---\n\n## STEP 1 — ANALYZE THE IMAGES\nBefore writing anything, internally assess from <Picture 1> (cross-referencing detail images):\n- Product type, shape, color, material, and finish\n- Existing lighting mood (hard/soft, warm/cool, dramatic/flat)\n- Background environment — describe it precisely and literally: surface type, color, texture, depth, atmosphere. This becomes a locked constant for all shots.\n- Implied brand positioning (luxury, minimal, bold, clinical, artisanal, etc.)\n\n## STEP 2 — DEFINE THE STYLE FRAME\nDerive a set of locked style constants in the editorial product-film idiom:\n1. **One palette.** A single monochromatic or duotone color world derived from the product and <Picture 1> (e.g. warm amber on near-black; sepia charcoal; cool blue-black). Every shot lives inside it.\n2. **One light.** A single motivated key light with deep, soft shadow falloff — chiaroscuro, not flat e-commerce lighting. Light can move, bloom, or sweep; it never multiplies.\n3. **One environment.** The scene from image 1, held constant. Seamless backdrop or minimal studio void. Never introduce new rooms, floors, or locations.\n4. **One material motif.** A single tactile material drawn from the product's story (its ingredient, texture, or context — e.g. wood shavings, stone, water, leather, silk, metal filings). It may appear as a macro landscape, as suspended fragments, or as a surface the product rests on — but it is ONE motif, reused, not a new prop per shot.\n5. **Suspended physics.** Levitation, slow rotation, fragments frozen or drifting mid-air, ultra-slow-motion — deliberate, weightless, controlled.\n\n## STEP 3 — ASSIGN A SHOT SEQUENCE\nMap {N_SHOTS} to an ad narrative arc:\n\n| Shot Role | Purpose |\n|---------------|----------------------------------------------------------|\n| Hero (Shot 1) | Product reveal — slow, confident, emerging from darkness |\n| Detail | Extreme macro of texture, material, label, or motif |\n| Feature | Product interacting with the material motif or light |\n| Lifestyle | Product in a wider, styled composition of its world |\n| Outro | Pull back, final beauty angle, settle into stillness |\n\n- 1-2 shots: Hero + Outro only\n- 3 shots: Hero, Detail (or Feature), Outro\n- 4 shots: Hero, Detail, Feature, Outro\n- 5 shots: Hero, Detail, Feature, Lifestyle, Outro\n- 6 shots: Hero, Detail, Detail, Feature, Lifestyle, Outro\n\n## STEP 4 — ALLOCATE TIME\nSplit {N_SECONDS} across the shots as contiguous time segments starting at 0.0s:\n- Use increments of 0.5 seconds; segments must sum EXACTLY to {N_SECONDS}\n- No shot shorter than 1.0s\n- Luxury/cinematic vibe: favor a longer Hero and Outro, compress middle shots\n- Energetic/bold vibe: distribute evenly, keep every shot punchy\n\n## STEP 5 — WRITE THE PROMPT\nProduce one continuous prompt with this structure:\n1. An opening global-style paragraph: film type, the product (as \"the product from <Picture 1>\"), the locked palette, light, environment, and material motif, and the note that the environment from <Picture 1> is constant throughout.\n2. One time-cued line per shot in the form `SHOT # [start s - end s] - `, beginning every shot after the first with \"Cut to\". Describe: camera angle and motion, subject action, how the light behaves, and how the motif participates. 25-60 words per shot.\n3. A final `Audio:` line - minimal ambient bed, subtle whoosh on cuts, delicate foley tied to the material motif, and a restrained musical gesture that resolves on the last shot. No voiceover, no lyrics.\n\nPrompt writing rules:\n- Present tense, descriptive and directive; lead with the most important visual element of each shot\n- Ground identity in the references: say \"the product from <Picture 1>\"; when a detail matters, cite its source (e.g. \"the label exactly as in <Picture 2>\")\n- Shot 1 opens on (or quickly reveals) the composition of <Picture 1> before the camera begins to move\n- Vary angle, distance, and motion meaningfully between shots - at least one extreme macro when {N_SHOTS} >= 3\n- Camera vocabulary: slow push-in, glide, orbit, top-down descent, tilt reveal, drift back. One camera idea per shot.\n- No on-screen text, no logos, no titles, no graphics, no UI. No people unless {USER_STORY} asks for them.\n- The environment is a locked constant, not a creative variable. Never introduce new settings or weather.\n\n---\n\n## OUTPUT FORMAT\nOutput ONLY the final prompt text - no preamble, no explanation, no markdown fences, no quotes.\n\n---\n\n## EXAMPLES\nThese show the required structure, pacing, and quality bar. Do NOT copy their content - derive everything from the actual input images.\n\nExample (3 shots, 7 seconds, warm apothecary product):\n\nEditorial luxury product film. The dark amber glass bottle from <Picture 1> in its original scene: a matte near-black surface fading into a warm brown void, one warm spotlight from upper left, fine atmospheric haze, deep soft shadows. Monochromatic amber-on-black palette. Material motif: dried cedar shavings. The environment from image 1 is constant throughout.\nSHOT 1 [0.0s - 2.5s] - The scene opens exactly on <Picture 1>, the bottle standing centered in near-darkness; the spotlight slowly blooms brighter as the camera pushes in from a low angle; thin curls of cedar shaving drift weightlessly around the bottle.\nSHOT 2 [2.5s - 4.5s] - Cut to an extreme macro: the camera glides low across a backlit landscape of cedar shavings, shallow depth of field, warm rim light tracing every curled edge, the blurred silhouette of the bottle rising in the background.\nSHOT 3 [4.5s - 7.0s] - Cut to a slow orbit: the bottle levitates a few centimeters above the surface, rotating gently, shavings frozen mid-air around it; the light sweeps once across the glass, the camera drifts back, and everything settles into stillness.\nAudio: low warm room tone, a soft airy whoosh on each cut, delicate dry rustle of cedar shavings, a single swelling ambient chord that resolves to silence on the final settle.\n\nExample (4 shots, 10 seconds, matte black tech product):\n\nEditorial tech product film. The matte black over-ear headphones from <Picture 1> in their original scene: a charcoal studio void, one hard key light from camera right with long falloff, sepia-toned monochrome palette. Material motif: fine leather grain. The environment from <Picture 1> is constant throughout.\nSHOT 1 [0.0s - 3.0s] - The scene opens on <Picture 1> in heavy shadow; a blade of light sweeps slowly across the earcup, revealing the silhouette; the camera pushes in with almost imperceptible slowness.\nSHOT 2 [3.0s - 5.5s] - Cut to an extreme macro of the leather earpad grain from <Picture 2>, the texture rolling like dunes under raking sepia light, focus breathing gently across the stitching.\nSHOT 3 [5.5s - 8.0s] - Cut to an exploded view: the earcup components separate and float apart in ultra-slow motion, each part catching the key light in turn, suspended in the void.\nSHOT 4 [8.0s - 10.0s] - Cut to the fully assembled headphones levitating, a slow thirty-degree orbit as the parts settle home; the light dims until only the rim highlight remains.\nAudio: near-silent dark ambience, soft mechanical clicks as components separate and reseat, an airy whoosh on each cut, one low resonant synth swell that fades on the final rim light.\n\nExample (5 shots, 10 seconds, clear glass tech earbuds)\nEditorial tech film. The clear earbuds from <Picture 1> in an infinite black void. High-contrast monochrome palette with neon cyan accents. Stark studio light, sharp reflections. Material motif: glossy obsidian. Environment is a locked constant.\nSHOT 1 [0.0s - 2.5s] - The scene opens exactly on <Picture 1>. The clear case lid glides open. The earbuds rest in the white base as their blue rings ignite.\nSHOT 2 [2.5s - 4.5s] - Cut to extreme macro. Camera tracks down the transparent outer shell. The blue ring pulses brightly, casting a crisp cyan lens flare.\nSHOT 3 [4.5s - 6.5s] - Cut to slow push-in. The earbud levitates weightlessly toward the sleek, glossy black mannequin ear from <Picture 2>.\nSHOT 4 [6.5s - 8.0s] - Cut to tight profile. The earbud seats precisely into the ear. A vivid blue light pulse instantly washes down the glossy black neck.\nSHOT 5 [8.0s - 10.0s] - Cut to wide beauty orbit. Both earbuds levitate above the open case in the black void, pulsing in unison before the light fades.\nAudio: Deep room tone, sleek mechanical whoosh, a heavy digital bass pulse on connection, fading synth echo."
},
"_meta": {
"title": "System Prompt"
}
},
"327": {
"class_type": "SaveVideo",
"inputs": {
"filename_prefix": "video/MiniMax_H3",
"format": "auto",
"format.codec": "auto",
"video": [
"328:326",
0
]
},
"_meta": {
"title": "Save Video"
}
},
"115:111": {
"class_type": "ComfySwitchNode",
"inputs": {
"on_false": [
"165",
0
],
"on_true": [
"166",
0
],
"switch": [
"115:112",
0
]
},
"_meta": {
"title": "If/Else Switch"
}
},
"115:112": {
"class_type": "StringCompare",
"inputs": {
"case_sensitive": False,
"mode": "Equal",
"string_a": [
"165",
0
],
"string_b": [
"115:113",
0
]
},
"_meta": {
"title": "Compare Text"
}
},
"115:113": {
"class_type": "PrimitiveString",
"inputs": {
"value": ""
},
"_meta": {
"title": "Empty String"
}
},
"328:311": {
"class_type": "ComfyMathExpression",
"inputs": {
"expression": "max(5, round(a * 24)) + (5 - (max(5, round(a * 24)) % 17)) % 17",
"values.a": [
"22",
0
]
},
"_meta": {
"title": "Math Expression"
}
},
"328:313": {
"class_type": "ResolutionSelector",
"inputs": {
"aspect_ratio": "16:9 (Widescreen)",
"megapixels": 1,
"multiple": 32
},
"_meta": {
"title": "Resolution Selector (Size)"
}
},
"328:314": {
"class_type": "VAELoader",
"inputs": {
"vae_name": "minimax_h3_audio_vae_fp32.safetensors"
},
"_meta": {
"title": "Load VAE"
}
},
"328:315": {
"class_type": "MiniMaxH3ReferenceToVideo",
"inputs": {
"audio_vae": [
"328:314",
0
],
"clip": [
"328:317",
0
],
"height": [
"328:313",
1
],
"length": [
"328:311",
1
],
"prompt": [
"115:111",
0
],
"ref_image_size": "match",
"ref_images.ref_image_0": [
"4",
0
],
"ref_images.ref_image_1": [
"8",
0
],
"ref_images.ref_image_2": [
"12",
0
],
"vae": [
"328:316",
0
],
"width": [
"328:313",
0
]
},
"_meta": {
"title": "MiniMax H3 Reference to Video"
}
},
"328:316": {
"class_type": "VAELoader",
"inputs": {
"vae_name": "minimax_h3_video_vae_fp16.safetensors"
},
"_meta": {
"title": "Load VAE"
}
},
"328:317": {
"class_type": "CLIPLoader",
"inputs": {
"clip_name": "qwen3vl_32b_minimax_h3_int8_convrot.safetensors",
"device": "default",
"type": "minimax"
},
"_meta": {
"title": "Load CLIP"
}
},
"328:318": {
"class_type": "KSamplerSelect",
"inputs": {
"sampler_name": "res_multistep"
},
"_meta": {
"title": "KSamplerSelect"
}
},
"328:319": {
"class_type": "BasicScheduler",
"inputs": {
"denoise": 1,
"model": [
"328:321",
0
],
"scheduler": "simple",
"steps": 20
},
"_meta": {
"title": "BasicScheduler"
}
},
"328:320": {
"class_type": "BasicGuider",
"inputs": {
"conditioning": [
"328:315",
0
],
"model": [
"328:321",
0
]
},
"_meta": {
"title": "Basic Guider"
}
},
"328:321": {
"class_type": "UNETLoader",
"inputs": {
"unet_name": "minimax_h3_ref2va_pruned_int8_convrot.safetensors",
"weight_dtype": "default"
},
"_meta": {
"title": "Load Diffusion Model"
}
},
"328:322": {
"class_type": "RandomNoise",
"inputs": {
"noise_seed": 873821824241961
},
"_meta": {
"title": "RandomNoise"
}
},
"328:323": {
"class_type": "SamplerCustomAdvanced",
"inputs": {
"guider": [
"328:320",
0
],
"latent_image": [
"328:315",
1
],
"noise": [
"328:322",
0
],
"sampler": [
"328:318",
0
],
"sigmas": [
"328:319",
0
]
},
"_meta": {
"title": "SamplerCustomAdvanced"
}
},
"328:324": {
"class_type": "VAEDecodeAudio",
"inputs": {
"samples": [
"328:323",
0
],
"vae": [
"328:314",
0
]
},
"_meta": {
"title": "VAE Decode Audio"
}
},
"328:325": {
"class_type": "VAEDecode",
"inputs": {
"samples": [
"328:323",
0
],
"vae": [
"328:316",
0
]
},
"_meta": {
"title": "VAE Decode"
}
},
"328:326": {
"class_type": "CreateVideo",
"inputs": {
"audio": [
"328:324",
0
],
"bit_depth": 8,
"codec": "none",
"color_space": "sRGB",
"fps": 24,
"images": [
"328:325",
0
]
},
"_meta": {
"title": "Create Video"
}
},
"81:303": {
"class_type": "PrimitiveStringMultiline",
"inputs": {
"value": "Using the attached reference images (image 1 is the hero/environment reference; any additional images are product detail references), create one multi-shot video prompt with {N_SHOTS} distinct shots. The total video duration is exactly {N_SECONDS} seconds.\n\nCreative direction from the user (may be empty): {USER_STORY}"
},
"_meta": {
"title": "Template User Prompt"
}
},
"81:76": {
"class_type": "SomethingToString",
"inputs": {
"input": [
"23",
0
],
"prefix": "",
"suffix": ""
},
"_meta": {
"title": "Something To String"
}
},
"81:77": {
"class_type": "SomethingToString",
"inputs": {
"input": [
"22",
0
],
"prefix": "",
"suffix": ""
},
"_meta": {
"title": "Something To String"
}
},
"81:78": {
"class_type": "RegexReplace",
"inputs": {
"case_insensitive": True,
"count": 0,
"dotall": False,
"multiline": False,
"regex_pattern": "{N_SHOTS}",
"replace": [
"81:76",
0
],
"string": [
"81:303",
0
]
},
"_meta": {
"title": "Replace Text (Regex)"
}
},
"81:79": {
"class_type": "RegexReplace",
"inputs": {
"case_insensitive": True,
"count": 0,
"dotall": False,
"multiline": False,
"regex_pattern": "{N_SECONDS}",
"replace": [
"81:77",
0
],
"string": [
"81:78",
0
]
},
"_meta": {
"title": "Replace Text (Regex)"
}
},
"81:85": {
"class_type": "RegexReplace",
"inputs": {
"case_insensitive": True,
"count": 0,
"dotall": False,
"multiline": False,
"regex_pattern": "{USER_STORY}",
"replace": [
"200",
0
],
"string": [
"81:79",
0
]
},
"_meta": {
"title": "Replace Text (Regex)"
}
}
})
# run() also passes api_key to the partner nodes in this workflow.
job = client.run(workflow, api_key=api_key)問答
MiniMax H3 只需一個提示詞,內含所有分鏡與明確時間區間(如 0–3 秒、3–6 秒)。GeminiNodeV2 會自動產生帶時間碼的分鏡清單,讓 H3 能從單一結構化提示產生多段影片。
請用強烈且具品牌感的 Image 1 作為主視覺/環境參考。流程會將此場景鎖定於所有分鏡。若需一致性,請避免在提示詞中大幅更換場景。
可以。範本預設支援三張,你可自行增加 LoadImage 節點,並將其輸入同一組提示詞與生成器。
GeminiNodeV2 與 MiniMax H3 多數情況會呼叫外部服務。依節點設定,可能需要有效 API 金鑰與網路連線。請檢查各節點設定與 ComfyUI 主控台。