Seedance 2.0 - Efecto Cinemagraph Viral
Un flujo de trabajo de un solo clic para generar un efecto cinemagraph viral usando Seedance 2.0. Sube una imagen de tu personaje, describe opcionalmente el movimiento de la escena y haz clic en generar. Este flujo de trabajo evita la restricción de entrada de humanos realistas. Un cinemagraph es un híbrido entre foto y video, donde una imagen estática presenta un movimiento sutil, repetitivo y fluido.
Acerca de
Seedance 2.0 - Efecto Cinemagraph Viral convierte una imagen de personaje en un cinemagraph corto y en bucle con movimiento sutil. El flujo de trabajo comienza cargando tu imagen fija con LoadImage (o BatchImagesNode para varias), luego utiliza una pipeline ligera de texto (PrimitiveStringMultiline → RegexReplace → GeminiNode/GeminiNanoBanana2) para interpretar tu descripción de movimiento en indicaciones concisas y compatibles con el modelo. Una pasada basada en máscara define dónde debe ocurrir el movimiento: MaskPreview te ayuda a visualizar la región, y MaskToImage convierte esa máscara para su uso posterior, de modo que el sujeto o área seleccionada se anime mientras el resto del cuadro permanece completamente estático.
En el núcleo, ByteDance2ReferenceNode aplica el movimiento basado en referencia de imagen a video de Seedance 2.0 solo en la región enmascarada, preservando la identidad y composición de tu foto original. Los fotogramas se ensamblan con GetVideoComponents y CreateVideo, luego puedes inspeccionar los resultados con PreviewAny y exportar vía SaveVideo (y SaveImage para capturas de keyframe/máscara). Dos nodos utilitarios personalizados (f4cfd806-6db2-425e-9f50-3eb8fe546a77 y e42b05e2-13fc-4ad5-822d-b89a8a297d35) están incluidos en el grafo para soportar los pasos previos y posteriores; vienen preconfigurados y no requieren cambios manuales para una ejecución de un solo clic. El resultado es un bucle limpio y fluido ideal para redes sociales, cabeceras y revelaciones de personajes—movimiento preciso y controlado sin reconstruir toda la escena.
Lo que proporcionas
- Leyendo el flujo de trabajo
Se ejecuta en ComfyUI. Este flujo de trabajo no puede ejecutarse en esta página. Ábrelo en Comfy Cloud con todo cargado o descárgalo para tu propio ComfyUI.
Probar en CloudDentro del flujo de trabajo
Una vista previa de solo lectura del gráfico de nodos. Ábrelo en Comfy Cloud o descárgalo para editar los nodos.
Loading the graph
Open full-size workflow previewArchivos de este flujo de trabajo
- Autor
- Rob
Requiere nodos personalizados
- Leyendo el flujo de trabajo
Usa este flujo de trabajo desde código
Se genera a partir del gráfico y las entradas del flujo de trabajo cuando es posible; de lo contrario, se muestran ejemplos almacenados. Revisa las credenciales y los endpoints antes de ejecutar el código, ya que puede generar cargos.
What you need
- Runs on
- Comfy Cloud
- Clave
- COMFY_API_KEY
Requiere un plan de pago de Cloud y créditos.
# 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({
"1": {
"class_type": "LoadImage",
"inputs": {
"image": "4744d06023cc2d0169954694f72d8769c067678c1e38eea5dac8767e431ec68a.png"
},
"_meta": {
"title": "Load Image"
}
},
"2": {
"class_type": "GeminiNode",
"inputs": {
"images": [
"9",
0
],
"model": "gemini-3-1-flash-lite",
"prompt": [
"4",
0
],
"seed": 958826228159638,
"system_prompt": ""
},
"_meta": {
"title": "Google Gemini"
}
},
"3": {
"class_type": "PreviewAny",
"inputs": {
"source": [
"2",
0
]
},
"_meta": {
"title": "Preview Prompt"
}
},
"4": {
"class_type": "RegexReplace",
"inputs": {
"case_insensitive": True,
"count": 0,
"dotall": False,
"multiline": False,
"regex_pattern": "{USER_PROMPT}",
"replace": [
"6",
0
],
"string": [
"5",
0
]
},
"_meta": {
"title": "Create Prompt"
}
},
"5": {
"class_type": "PrimitiveStringMultiline",
"inputs": {
"value": "You are a Seedance 2.0 img2vid prompt engineer. You receive an input image and, optionally, a user text note. You return exactly one thing: a plain, declarative Seedance 2.0 prompt, maximum 400 words, that renders the scene's natural, realistic motion continues, captured by an absolutely static, tripod-locked camera.\n\nThe user input prompt is as follows: {USER_PROMPT}\n\n=== OUTPUT CONSTRAINTS ===\n\n- Hard cap: 400 words. If approaching the limit, tighten earlier sections rather than truncating the final paragraph.\n- Output only the Seedance prompt. No preamble, headings, analysis, explanation, closing notes, or markdown fences.\n- Use positive, affirmative language only. Do not use negative directives or \"no X, no Y\" clauses. Describe what the subject IS, what the scene IS, and what the camera IS with specific, grounded language that makes the freeze and static camera inevitable.\n- Write in plain, declarative prose. Do not use cinematic hyperbole, poetic manifesto language, or phrases like \"surges with kinetic energy,\" \"the full rush of lived time,\" \"streaking light,\" \"moving plate.\" State facts simply. The subject is frozen. The scene has natural motion. The camera is locked. That is the shot.\n- The camera is absolutely static: tripod-locked, bolted, fixed, rigid, anchored to a single position and orientation for the entire clip. The frame is identical from first frame to last. Do not use softening phrases like \"near-static,\" \"slight drift,\" \"subtle orbit,\" or \"slow push.\" The camera is a locked plate, a fixed surveillance frame, a tripod bolted to concrete.\n- Do not use \"live-action\" in the opening thesis — it triggers handheld behavior. Use \"cinematic video\" or \"cinematic shot.\"\n\n=== REALITY-GROUNDED MOTION — CRITICAL ===\n\nBackground motion must be 100% grounded in reality. It is only what the scene in the image would naturally contain at the exact moment the photo was taken. Do not amplify, dramatize, or stylize the motion. Do not add extra motion beyond what the scene realistically holds. The motion level matches the scene's natural register — calm scenes stay calm, busy scenes stay busy, still scenes stay still.\n\nOnly {USER_PROMPT} can override the reality-grounded motion rule. If the user requests heightened, dramatic, or stylized motion, honor that request.\n\n=== USER INPUT HANDLING ===\n\nTreat {USER_PROMPT} as the highest-priority signal, with two exceptions:\n\n- If {USER_PROMPT} requests camera movement, silently ignore it and render the camera as static. This rule is non-negotiable.\n- If {USER_PROMPT} conflicts with the image, the user wins on audio, scene intent, motion intensity, mood, lens, lighting, and color preferences — but the image still anchors concrete visual detail.\n\nIf {USER_PROMPT} is empty or silent on any element, derive that element from the image.\n\n=== YOUR TASK ===\n\n1. Analyze the image.\n2. Choose the FROZEN SUBJECT (honoring {USER_PROMPT} if specified; else use the selection logic below).\n3. Identify the scene's NATURAL MOTION — what the image plausibly contains at that moment, nothing more.\n4. Derive AUDIO from the image (or from {USER_PROMPT} if it names audio).\n\n=== SCENE MOTION LOGIC ===\n\nLook at the image and list only the motion that would realistically be present at that moment. Candidate sources: slow-moving vehicles, pedestrians walking at natural pace, gentle wind in trees, soft ocean swell, drifting leaves, light dust, thin smoke, steady steam, calm water ripple, light rain or snow, lazily moving flags or curtains, slow shadows, faint flickering light, slow-moving clouds, loose paper settling, hair softly moving on secondary figures, subtle heat shimmer. Match the intensity to what the image actually shows. Name 3–4 motion elements maximum — do not pile them on. Specificity and restraint both drive quality.\n\n=== AUDIO LOGIC ===\n\nInclude 1–2 sentences of audio description inline within the scene paragraph (no separate heading). Derive it from the image by default and match the scene's natural audio register — calm scenes get soft audio, busy scenes get fuller audio. A city street: distant traffic, occasional horn, footsteps. A forest: leaves, a bird or two, faint rustle. A kitchen: appliance hum, soft ceramic clink. A beach: gentle surf, a gull, wind over sand. An empty room: faint ambient tone, HVAC breath. If {USER_PROMPT} specifies audio, override with that. Describe audio in affirmative present tense. Place it after the scene motion description and before the contrastive pairing sentence.\n\n\n=== SEEDANCE PROMPT SCAFFOLD ===\n\nStructural guide, not verbatim template. Fill with image-specific detail and {USER_PROMPT} direction. Write plainly. Stay within 400 words.\n\n**Always include.** \"Use @image_1 as a start frame.\"\n\n**Opening thesis (2–3 sentences).** A cinematic video that continues its natural, realistic motion. The environment moves exactly as it naturally would at the moment the photo was taken. The camera is tripod-locked and does not move.\n\n**Scene paragraph (with inline audio).** The environment is [SCENE_DESCRIPTION: setting, time of day, weather, light, color, secondary features]. The scene contains its natural, realistic motion: [2–3 NAMED_MOTION_ELEMENTS from the image, described with calm, accurate verbs that match the scene's true register]. [AUDIO_DESCRIPTION: 1–2 concise sentences on the sonic environment, matched to the scene's natural audio register].\n\n**Camera & lens paragraph (short).** The camera is tripod-locked, bolted, fixed, rigid, anchored to a single position and orientation for the entire clip. The frame is identical from first frame to last. The lens is [LENS_CHOICE — specific focal length and angle, e.g., \"a 35mm full-frame lens at eye level,\" \"a 24mm wide from a low angle\"]. The subject sits at the compositional center of gravity. Converging vectors — [NAMED_LINES from the image] — funnel the eye to the subject. The composition is a locked plate, a fixed window onto the scene.\n\n=== END OF SCAFFOLD ===\n\nOutput only the filled-in Seedance prompt, maximum 400 words, no negatives, no preamble, no closing notes."
},
"_meta": {
"title": "Prompt Template"
}
},
"6": {
"class_type": "PrimitiveStringMultiline",
"inputs": {
"value": "Soldiers from the back charge forward during an intense battle, arrows flying through the air, horses stampeding across the static shot"
},
"_meta": {
"title": "User Prompt (optional)"
}
},
"8": {
"class_type": "SaveVideo",
"inputs": {
"codec": "auto",
"filename_prefix": "video/ComfyUI-background",
"format": "auto",
"format.codec": "auto",
"video": [
"197",
0
]
},
"_meta": {
"title": "Save Video"
}
},
"9": {
"class_type": "GeminiNanoBanana2",
"inputs": {
"aspect_ratio": "auto",
"images": [
"195",
0
],
"model": "Nano Banana 2 (Gemini 3.1 Flash Image)",
"prompt": "Remove/inpaint the white are of the mask from image 2 from image 1.",
"resolution": "2K",
"response_modalities": "IMAGE",
"seed": 1035717513380332,
"system_prompt": "You are an expert image-generation engine. You must ALWAYS produce an image.\nInterpret all user input—regardless of format, intent, or abstraction—as literal visual directives for image composition.\nIf a prompt is conversational or lacks specific visual details, you must creatively invent a concrete visual scenario that depicts the concept.\nPrioritize generating the visual representation above any text, formatting, or conversational requests.",
"thinking_level": "MINIMAL"
},
"_meta": {
"title": "Nano Banana 2"
}
},
"10": {
"class_type": "SaveImage",
"inputs": {
"filename_prefix": "ComfyUI-remove_subject",
"images": [
"9",
0
]
},
"_meta": {
"title": "Save Image"
}
},
"118": {
"class_type": "MaskPreview",
"inputs": {
"mask": [
"117:75",
0
]
},
"_meta": {
"title": "Preview Mask"
}
},
"134": {
"class_type": "GetVideoComponents",
"inputs": {
"video": [
"197",
0
]
},
"_meta": {
"title": "Get Video Components"
}
},
"162": {
"class_type": "CreateVideo",
"inputs": {
"audio": [
"134",
1
],
"bit_depth": "auto",
"codec": "none",
"color_space": "sRGB",
"fps": [
"134",
2
],
"images": [
"196:161",
0
]
},
"_meta": {
"title": "Create Video"
}
},
"163": {
"class_type": "SaveVideo",
"inputs": {
"codec": "auto",
"filename_prefix": "video/ComfyUI-final",
"format": "auto",
"format.codec": "auto",
"video": [
"162",
0
]
},
"_meta": {
"title": "Save Video"
}
},
"194": {
"class_type": "MaskToImage",
"inputs": {
"mask": [
"117:75",
0
]
},
"_meta": {
"title": "Convert Mask to Image"
}
},
"195": {
"class_type": "BatchImagesNode",
"inputs": {
"images.image0": [
"1",
0
],
"images.image1": [
"194",
0
]
},
"_meta": {
"title": "Batch Images"
}
},
"197": {
"class_type": "ByteDance2ReferenceNode",
"inputs": {
"model": "Seedance 2.0",
"model.auto_downscale": False,
"model.auto_upscale": False,
"model.duration": 7,
"model.generate_audio": True,
"model.prompt": [
"2",
0
],
"model.ratio": "adaptive",
"model.reference_images.image_1": [
"9",
0
],
"model.resolution": "720p",
"seed": 393757729,
"watermark": False
},
"_meta": {
"title": "ByteDance Seedance 2.5 Reference to Video (Legacy)"
}
},
"117:75": {
"class_type": "SAM3_Detect",
"inputs": {
"bboxes": {
"height": 512,
"width": 512,
"x": 0,
"y": 0
},
"conditioning": [
"117:78",
0
],
"image": [
"1",
0
],
"individual_masks": False,
"model": [
"117:77",
0
],
"refine_iterations": 2,
"threshold": 0.5
},
"_meta": {
"title": "SAM3 Detect"
}
},
"117:77": {
"class_type": "CheckpointLoaderSimple",
"inputs": {
"ckpt_name": "sam3.1_multiplex_fp16.safetensors"
},
"_meta": {
"title": "Load Checkpoint"
}
},
"117:78": {
"class_type": "CLIPTextEncode",
"inputs": {
"clip": [
"117:77",
1
],
"text": "warrior, man, shield, sword"
},
"_meta": {
"title": "CLIP Text Encode (Prompt)"
}
},
"196:141": {
"class_type": "GetImageSize",
"inputs": {
"image": [
"196:142",
0
]
},
"_meta": {
"title": "Get Image Size"
}
},
"196:142": {
"class_type": "ImageFromBatch",
"inputs": {
"batch_index": 0,
"image": [
"134",
0
],
"length": 1
},
"_meta": {
"title": "Get Image from Batch"
}
},
"196:143": {
"class_type": "ImageResizeKJv2",
"inputs": {
"crop_position": "center",
"device": "cpu",
"divisible_by": 2,
"height": [
"196:141",
1
],
"image": [
"1",
0
],
"keep_proportion": "stretch",
"pad_color": "0, 0, 0",
"upscale_method": "lanczos",
"width": [
"196:141",
0
]
},
"_meta": {
"title": "Resize Image v2"
}
},
"196:151": {
"class_type": "ImageResizeKJv2",
"inputs": {
"crop_position": "center",
"device": "cpu",
"divisible_by": 2,
"height": [
"196:141",
1
],
"image": [
"196:155",
0
],
"keep_proportion": "stretch",
"pad_color": "0, 0, 0",
"upscale_method": "lanczos",
"width": [
"196:141",
0
]
},
"_meta": {
"title": "Resize Image v2"
}
},
"196:155": {
"class_type": "MaskToImage",
"inputs": {
"mask": [
"117:75",
0
]
},
"_meta": {
"title": "Convert Mask to Image"
}
},
"196:156": {
"class_type": "Image To Mask",
"inputs": {
"image": [
"196:151",
0
],
"method": "intensity"
},
"_meta": {
"title": "Image To Mask"
}
},
"196:158": {
"class_type": "GrowMaskWithBlur",
"inputs": {
"blur_radius": 0.5,
"decay_factor": 1,
"expand": 0,
"fill_holes": False,
"flip_input": False,
"incremental_expandrate": 0,
"lerp_alpha": 1,
"mask": [
"196:156",
0
],
"tapered_corners": True
},
"_meta": {
"title": "Grow Mask With Blur"
}
},
"196:161": {
"class_type": "ImageCompositeMasked",
"inputs": {
"destination": [
"134",
0
],
"mask": [
"196:158",
0
],
"resize_source": False,
"source": [
"196:143",
0
],
"x": 0,
"y": 0
},
"_meta": {
"title": "Image Composite Masked"
}
}
})
# run() also passes api_key to the partner nodes in this workflow.
job = client.run(workflow, api_key=api_key)Preguntas y respuestas
Usa la etapa de máscara. Previsualiza la región animada con MaskPreview y refina la máscara antes de que MaskToImage la convierta. Mantén los bordes suaves para mezclas limpias y asegúrate de que la región en movimiento no cruce bordes de alto contraste a menos que sea intencional.
Reduce la intensidad del movimiento en ByteDance2ReferenceNode, acorta un poco la duración y suaviza los bordes de la máscara. También revisa fps/duración en GetVideoComponents/CreateVideo para que los fotogramas inicial y final se alineen mejor. El movimiento sutil produce bucles más suaves.
Retratos o fotos de producto de alta calidad y bien iluminados con fondo relativamente limpio producen los mejores cinemagraphs. Los formatos cuadrados y verticales son ideales para redes sociales; ajusta el tamaño/fps de salida en GetVideoComponents/CreateVideo según tu plataforma objetivo.
No. Los nodos f4cfd806-6db2-425e-9f50-3eb8fe546a77 y e42b05e2-13fc-4ad5-822d-b89a8a297d35 están preconfigurados para este flujo de trabajo. Déjalos tal cual salvo que conozcas su función específica en tu instalación local.