# 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": "OpenAIGPTImageNodeV2",
"inputs": {
"model": "gpt-image-2",
"model.background": "auto",
"model.custom_height": 1024,
"model.custom_width": 1536,
"model.quality": "low",
"model.size": "1536x1024",
"n": 1,
"prompt": [
"4",
0
],
"seed": 1096883914
},
"_meta": {
"title": "OpenAI GPT Image 2.5"
}
},
"2": {
"class_type": "SaveImage",
"inputs": {
"filename_prefix": "ComfyUI",
"images": [
"1",
0
]
},
"_meta": {
"title": "Save Image"
}
},
"4": {
"class_type": "GeminiNode",
"inputs": {
"model": "gemini-3-1-pro",
"prompt": "A ninja sprints across a rooftop beneath the moonlit night sky before launching high into the air. He descends with precision, crashing down onto an unsuspecting enemy and driving them to the ground.\n\nAs the enemy falls, the ninja immediately slips behind the corner of a nearby structure, disappearing from sight. Cut to a close-up of his eyes peering around the edge of the wall—focused, alert, and tense as he scans for danger.\n\nSpotting another enemy nearby, he quietly reaches for a ninja star. Without hesitation, he flicks his wrist and sends the weapon flying across the scene. The spinning star cuts through the air and strikes the second enemy, who stumbles backward from the impact before collapsing to the ground.\n\nFast-paced action, dynamic camera angles, dramatic moonlight, stealth, tension, and fluid ninja choreography.\n",
"seed": 486422127652111,
"system_prompt": "You are an elite action-film previs storyboard artist and cinematic director.\n\nYour task is to convert the user's idea into a highly detailed image-generation prompt for a single storyboard sheet containing 8 continuous sequential action frames.\n\nThe goal is not realism or finished concept art.\n\nThe goal is to communicate:\n\n- action choreography\n- movement\n- camera language\n- timing\n- momentum\n- staging\n- visual storytelling\n- dynamic composition\n\nVISUAL STYLE\n\nCreate a rough director-style previs storyboard.\n\nUse only:\n\n- minimalist stick figures\n- simple gesture drawings\n- construction-line anatomy\n- motion arcs\n- speed lines\n- camera framing indicators\n- directional arrows\n- momentum trails\n- impact marks\n- environmental movement cues\n\nDo NOT include:\n\n- realistic anatomy\n- muscles\n- clothing details\n- costumes\n- shoes\n- facial rendering\n- textures\n- shading\n- polished illustration\n- comic-book rendering\n- finished character design\n\nCharacters should appear as simple action mannequins whose purpose is only to communicate movement.\n\nSTORYBOARD STRUCTURE\n\nGenerate:\n\n- 8 continuous action panels\n- arranged in a storyboard grid\n- each panel representing the next moment in time\n\nEvery frame should feel connected to the previous frame.\n\nShow clear progression of:\n\n- movement\n- momentum\n- action escalation\n- camera movement\n- visual rhythm\n\nCAMERA DIRECTION\n\nUse highly varied cinematic camera language.\n\nPossible examples:\n\n- extreme wide shot\n- wide shot\n- medium shot\n- close-up\n- extreme close-up\n- over-the-shoulder\n- low angle\n- high angle\n- bird's-eye view\n- worm's-eye view\n- Dutch angle\n- tracking shot\n- handheld energy\n- push-in framing\n- pull-back framing\n\nAvoid repetitive framing.\n\nEvery panel should introduce a meaningful camera choice.\n\nACTION VISUALIZATION\n\nEmphasize:\n\n- body mechanics\n- weight shifts\n- momentum\n- anticipation\n- follow-through\n- directional force\n- action readability\n\nVisualize movement using:\n\n- motion trails\n- gesture overlaps\n- action arcs\n- trajectory lines\n- force indicators\n- impact symbols\n- environmental reactions\n\nThe action should feel energetic and easy to understand at a glance.\n\nOBJECT INTERACTION\n\nWhen the user's idea includes props, weapons, tools, vehicles, powers, or environmental interaction:\n\n- make them visually dominant\n- clearly show how they move through space\n- exaggerate motion paths\n- show cause and effect\n- emphasize silhouettes and readability\n\nEFFECTS\n\nEffects should be represented as rough previs graphics:\n\n- ink splashes\n- smoke trails\n- debris\n- energy streaks\n- particles\n- geometric motion shapes\n- environmental reactions\n\nKeep effects simple and readable.\n\nUse them only to reinforce movement and direction.\n\nTEXT USAGE\n\nUse very little text.\n\nOnly include:\n\n- occasional camera notes\n- simple action notes\n- short directional labels\n\nNever use dialogue.\n\nNever use captions that dominate the image.\n\nDRAWING STYLE\n\nThe final image should resemble:\n\n- film previs\n- stunt choreography planning\n- action blocking sheets\n- animation planning boards\n- director thumbnail sketches\n- cinematic motion studies\n\nPrioritize clarity of action over artistic polish.\n\nThe storyboard should feel like it was created by an action director planning a complex sequence before production.\n\nOUTPUT FORMAT\n\nCreate a single image containing all 8 storyboard panels.\n\nEnsure strong visual variety across the panels while maintaining continuous action flow.\n\nThe image should communicate movement, timing, camera choreography, and cinematic energy immediately, even when viewed from a distance."
},
"_meta": {
"title": "Google Gemini"
}
},
"5": {
"class_type": "ByteDance2ReferenceNode",
"inputs": {
"model": "Seedance 2.0",
"model.auto_downscale": False,
"model.auto_upscale": False,
"model.duration": 10,
"model.generate_audio": True,
"model.prompt": [
"20",
0
],
"model.ratio": "16:9",
"model.reference_images.image_1": [
"10",
0
],
"model.reference_images.image_2": [
"7",
0
],
"model.resolution": "720p",
"seed": 275927675,
"watermark": False
},
"_meta": {
"title": "ByteDance Seedance 2.5 Reference to Video (Legacy)"
}
},
"6": {
"class_type": "GeminiNode",
"inputs": {
"images": [
"19",
0
],
"model": "gemini-3-1-pro",
"prompt": "A ninja sprints across a rooftop beneath the moonlit night sky before launching high into the air. He descends with precision, crashing down onto an unsuspecting enemy and driving them to the ground.\n\nAs the enemy falls, the ninja immediately slips behind the corner of a nearby structure, disappearing from sight. Cut to a close-up of his eyes peering around the edge of the wall—focused, alert, and tense as he scans for danger.\n\nSpotting another enemy nearby, he quietly reaches for a ninja star. Without hesitation, he flicks his wrist and sends the weapon flying across the scene. The spinning star cuts through the air and strikes the second enemy, who stumbles backward from the impact before collapsing to the ground.\n\nFast-paced action, dynamic camera angles, dramatic moonlight, stealth, tension, and fluid ninja choreography.\n10 seconds",
"seed": 1050544927289873,
"system_prompt": "Storyboard Camera Interpretation\n\nWhen a storyboard is provided:\n\n* Analyze each panel to determine the intended framing, perspective, shot size, camera position, camera movement, and progression of the sequence.\n* Treat storyboard panels as a shot plan. A change in panel composition may represent either a camera move or a cut depending on what best serves the sequence.\n* Generate dynamic, cinematic shot progression inspired by professional film editing and action choreography.\n* Use exciting, visually engaging shot variety when appropriate, including wide shots, medium shots, close-ups, extreme close-ups, low angles, high angles, tracking shots, push-ins, pull-backs, and perspective-driven compositions.\n* Prioritize strong visual storytelling, pacing, impact, and readability over literal recreation of storyboard drawings.\n* Use cuts when they improve pacing, emphasize important actions, reveal information, increase intensity, or enhance dramatic impact.\n* Ensure every cut feels motivated by the action and preserves continuity of movement, screen direction, subject identity, lighting, environment, and spatial relationships.\n* Maintain clear audience orientation at all times so the subject remains easy to follow throughout the sequence.\n* Translate rough storyboard framing into polished cinematic coverage while preserving the intended sequence of events.\n* Favor cinematic shot progression that escalates in energy and visual interest over time when appropriate.\n* Never reproduce storyboard panel borders, arrows, labels, notes, annotations, or text. Use them only to understand the intended action and camera language.\n* The final video should appear as a professionally shot cinematic sequence, not as an animated storyboard.\n"
},
"_meta": {
"title": "Google Gemini"
}
},
"7": {
"class_type": "LoadImage",
"inputs": {
"image": "34854d164971d4da64fba423cc05f17520192f53926162c7e8380867d485d018.png [output]"
},
"_meta": {
"title": "Load Image"
}
},
"9": {
"class_type": "SaveVideo",
"inputs": {
"filename_prefix": "video/ComfyUI",
"format": "auto",
"format.codec": "auto",
"video": [
"5",
0
]
},
"_meta": {
"title": "Save Video"
}
},
"10": {
"class_type": "LoadImage",
"inputs": {
"image": "b00101f87d3beb138186a1534e0c9ac6a4a95e20d2c1f9a02d59563b165bf816.png"
},
"_meta": {
"title": "Load Image"
}
},
"11": {
"class_type": "PreviewAny",
"inputs": {
"source": [
"20",
0
]
},
"_meta": {
"title": "Preview as Text"
}
},
"19": {
"class_type": "ImageStitch",
"inputs": {
"direction": "right",
"image1": [
"10",
0
],
"image2": [
"7",
0
],
"match_image_size": True,
"spacing_color": "white",
"spacing_width": 0
},
"_meta": {
"title": "Stitch Images"
}
},
"20": {
"class_type": "StringConcatenate",
"inputs": {
"delimiter": ", ",
"string_a": [
"6",
0
],
"string_b": "RULES:\n@image_1 is the character, maintain facial and outfit identity in the generation.\n@image_2 is only for rough scene composition. DO NOT generate arrows or text from this image."
},
"_meta": {
"title": "Concatenate Prompt"
}
}
})
# run() also passes api_key to the partner nodes in this workflow.
job = client.run(workflow, api_key=api_key)