巴黎写实视频完整制作流程

拿走我用 GPT-6 阿斯特拉制作超写实人工智能视频的完整工作流程吧…… 希格斯场刚刚发布了接口，所以我把它拿到实际场景里测试。 我生成了一段设在巴黎的视频日志风格视频……它看起来像朋友用手机拍摄的原始素材。 下面是逐步的工作流程： > 步骤 1：借用结构，不是像素 我从酒店和河景中截取了 3–5 个场景，然后从拼趣和真实视频日志中寻找人物灵感。 > 步骤 2：用 Gemini 分析参考 我把收集的所有内容都发给 Gemini……视频、音频、关键帧，它给出了逐镜头分析，我把分析转成指令。 > 步骤 3：更换世界，保留骨架 我对人物和环境做了很多改变，直到它感觉像属于我自己的东西，而不是复制品。 > 步骤 4：先制作参考图，这里才是迭代的地方 图片便宜，视频昂贵。 因此，我先制作角色设定页，为每个地点制作一张具有相同氛围的图，再制作视频开场的准确首帧，全部使用 Nano Banana Pro。 > 步骤 5：一次生成 30 秒视频，不是十段 3 秒片段 完整的带时间戳脚本通过接口交给 Seedance 2.5，一次生成。 以首帧为锚点，每个节拍一个动作，具体描述摄影机和音频，末尾添加身份锁定要求，让她的脸、服装和妆容保持固定。 Seedance 在内部处理切镜。 > 步骤 6：像导演一样修改，不是像赌徒一样 第一版剪辑效果不错，但还不完全符合想法。 我没有重新生成全部 30 秒，而是只把需要改变的场景，从各自的首帧重新生成短片段，再用 FFmpeg 将它们剪入主片。 每次只改变一个变量，绝不重新生成已经有效的部分。 > 步骤 7：像处理普通素材一样完成后期 我用 FFmpeg 从最佳素材中组装最终剪辑，添加免费素材库里的氛围音乐，就完成了。 让这套方法有效的规则： - 参考的结构才是资产，挖掘提示词比编写提示词更有效。 - 在图片上迭代，在视频上确定下来；首帧决定一个镜头的 80%。 - 每个带时间戳的节拍只安排一个动作，动作过多的节拍会被跳过或被模型自行编造。 - 明确锁定身份，否则它会在切镜之间漂移。 - 修改一个变量来修复失败镜头，重做场景，绝不重做整段视频。 - 最后的 10% 是乏味的剪辑：裁切、拼接、音乐、编码。 （保存这套工作流程）

steal my full workflow to create ultra-realistic AI videos with GPT-6 Astra... higgsfield just released their API, so i took it into real conditions i generated a vlog-style video set in Paris... it looks like raw footage a friend shot on a phone here's the workflow, step by step: > step 1: steal the structure, not the pixels i captured 3-5 scenes from a hotels and river views, then pulled character inspiration from pinterest and real vlogs > step 2: analyze the reference with Gemini i sent Gemini everything i'd collected... videos, audio, keyframes and it gave me a shot-by-shot breakdown, i turned that into instructions > step 3: swap the world, keep the bones i changed a lot on the character and the environment until it felt like something of my own, not a copy > step 4: build reference images first, this is where you iterate images are cheap, video is expensive so i built a character sheet, one image per location in the same mood, then the exact start frame the video opens on, all Nano Banana Pro > step 5: one 30-second video run, not ten 3-second clips the full timestamped script goes to Seedance 2.5 through the API as ONE generation start frame as the anchor, one action per beat, camera and audio described concretely, identity locks at the end so her face, outfit and makeup is locked in Seedance handles the cuts internally > step 6: revise like a director, not a gambler the first cut came back good but not right instead of re-rolling the whole 30 seconds i regenerated only the scenes that needed changes as short clips from their own start frames, then cut them into the master with ffmpeg change one variable at a time, never re-roll what already works > step 7: finish like normal footage i assembled the final cut with ffmpeg from the best takes, added ambient music from a free library and done the rules that made it work: - the reference's structure is the asset, prompt archaeology beats prompt writing - iterate on images, commit on video, the start frame decides 80% of the shot - one action per timestamped beat, crowded beats get skipped or invented - lock identity explicitly or it drifts between cuts - fix failed shots by changing one variable, redo scenes, never the whole video - the last 10% is boring editing: trim, splice, music, encode (save this workflow)