{"id":53656,"date":"2026-06-04T11:55:35","date_gmt":"2026-06-04T03:55:35","guid":{"rendered":"https:\/\/www.1ai.net\/?p=53656"},"modified":"2026-06-04T11:55:35","modified_gmt":"2026-06-04T03:55:35","slug":"%e4%b8%80%e9%94%ae%e7%94%9f%e6%88%90ai%e7%9f%ad%e7%89%87%ef%bc%8cgpt-%e5%88%86%e9%95%9cseedance-2-0%e5%85%a8%e6%b5%81%e7%a8%8b%e8%af%a6%e8%a7%a3%e6%95%99%e7%a8%8b","status":"publish","type":"post","link":"https:\/\/www.1ai.net\/en\/53656.html","title":{"rendered":"One key to generate an AI short, GPT lens +Seedance 2.0 full process solution"},"content":{"rendered":"

Recently a particularly interesting stream of work has been painted:GPT TO GENERATE ACTION LABELS FIRSTMirror<\/a>And feed Feedance 2.0 directly to a consistent short film\u3002<\/strong><\/p>\n

I've been testing one of them latelyAI Video<\/a>Workstream\u3002<\/p>\n

Used toAI short film<\/a>What's the worst headache<\/p>\n

It's not pretty enough. It's: moving around. The shot's out of control. The character is like this for a second. The next second is like an actor. It wasn't easy to produce a video, and the result wasn't exactly what it wanted. But recently I found a very useful combination:<\/p>\n

GPT + GPT Image + Seedance 2.0 first let GPT design the action lens\u3002<\/p>\n

Let GPT draw the specs with arrows. It's the last time you throw it to Seedance to generate a video\u3002<\/p>\n

The key is the direction of the action in the spectroscope, the trajectories of the lens, the movement of people, and the ability of Seedance to understand\u3002<\/p>\n

The resulting video was executed almost according to the lens 1:1\u3002<\/p>\n

THE VIDEO WAS FINALLY TURNED FROM A \"TICK CARD\" TO A \"DIRECTOR\"\u3002<\/p>\n

We'll tear the whole thing down to you today. It'll work\u3002<\/p>\n

STEP 1: GET GPT TO WRITE A SCRIPT OF THE TEXT<\/strong><\/p>\n

A lot of people went straight to the spectroscopy, and half of them found that the action logic was not right and the lens was broken\u3002<\/p>\n

The right thing to do is..Write a text version of the schedule first<\/strong>And make it visual\u3002<\/p>\n

How do you write a hint<\/p>\n

YOU NEED TO GET GPT INTO THE ROLE OF A MOVIE ACTION DESIGNER:<\/p>\n

Tips:<\/strong><\/p>\n

You're a film-class martial arts designer, and you're good at designing coherent, robust action lenses\u3002<\/span><\/p>\n

Please generate an action spectroscopy script of about 15 seconds, including:<\/span><\/p>\n

    \n
  1. \n
    Number and duration of each shot<\/span><\/section>\n<\/li>\n
  2. \n
    Speculation of the lens (character\/median view\/vision\/topography, etc.)<\/span><\/section>\n<\/li>\n
  3. \n
    Role position and track in the picture<\/span><\/section>\n<\/li>\n
  4. \n
    Camera motion (push\/push\/smash\/smash)<\/span><\/section>\n<\/li>\n
  5. \n
    Key action rhythm description<\/span><\/section>\n<\/li>\n<\/ol>\n

    Subject:<\/span><\/p>\n

    Role: [insert role description]<\/span><\/p>\n

    Style: [insert style keyword]<\/span><\/p>\n

    Give me a chestnut<\/p>\n

    FOR EXAMPLE, I WANT A 15-SECOND SHORT FILM OF THE OLD WIND'S GATE, AND GPT RETURNS TO A TEXT LENS LIKE THIS:<\/p>\n

      \n
    • camera 1 (2s)<\/strong>: Vision\/set, Mountain Gate Panorama, morning fog, role above step<\/li>\n<\/ul>\n
        \n
      • camera 2 (2s)<\/strong>\u00a0: Midview\/Stalking, role step up, dress up<\/li>\n<\/ul>\n
          \n
        • camera 3 (1.5s)<\/strong>: close up, the role pupils are shrunk, the dark fog is rising far away<\/li>\n<\/ul>\n
            \n
          • camera 4 (2s)<\/strong>: Vision\/push lens, black fog swallows the gates, oppression is full<\/li>\n<\/ul>\n
              \n
            • camera 5 (1.5s)<\/strong>: Midview\/opposite, the part pulls the sword around, the sword slashs through the dark<\/li>\n<\/ul>\n
                \n
              • camera 6 (2s)<\/strong>\u00a0: close-up\/slow motion, blades of black fog, fragments scattered<\/li>\n<\/ul>\n
                  \n
                • camera 7 (2s)<\/strong>\u00a0: Vision\/Link, part of the debris of independence, all around<\/li>\n<\/ul>\n
                    \n
                  • camera 8 (2s)<\/strong>\u00a0: close-up, character eyeballs display red light, picture set<\/li>\n<\/ul>\n

                    With this text, there is evidence for the second step\u3002<\/p>\n

                    \"317f841aj00tg39a100avd000ir00byp\"<\/p>\n

                    Step 2: Generate a spectroscopy with labels using GPT Image 2<\/strong><\/p>\n

                    Pure spectroscopy you can read, but Seedance can't\u3002<\/p>\n

                    So let GPT Image 2 put the word \"draw\" out..The point is to take the track<\/strong>.<\/p>\n

                    Cue word templates<\/p>\n

                    Tips:<\/strong><\/p>\n

                    Please generate a 3x4 film spectroscopy grid map (out of 12), each of which contains:<\/p>\n

                      \n
                    1. \n
                      Image content (fireman\/simplistic drawings, focus on action attitude and location)<\/section>\n<\/li>\n
                    2. \n
                      Red arrow indicates role direction<\/section>\n<\/li>\n
                    3. \n
                      Blue arrow points the camera track<\/section>\n<\/li>\n
                    4. \n
                      Number and length of lens per cell<\/section>\n<\/li>\n<\/ol>\n

                      Require:<\/p>\n

                        \n
                      • \n
                        All visual characters look the same (physical, clothing, identity)<\/section>\n<\/li>\n
                      • \n
                        The movement direction must be clear and the arrows must not overlap<\/section>\n<\/li>\n
                      • \n
                        The whole picture is clean, the lines are simple and powerful<\/section>\n<\/li>\n
                      • \n
                        style: film pre-visible spectroscope style<\/section>\n<\/li>\n<\/ul>\n

                        Text Script:<\/p>\n

                        [Pasting text spectroscopy generated by the first step]<\/p>\n

                        Why the matchmaker\/simple<\/p>\n

                        becauseTHE CORE OF THE SPECS IS NOT TO MAKE IT LOOK GOOD<\/strong>.<\/p>\n

                        Instead, a fine-drawing wind interferes with Seedance's identification of action -- it focuses on the details of the picture, not on the track. The matchesman plus arrow points out that Seedance understands better\u3002<\/p>\n

                        Actual effects<\/p>\n

                        GPT Image 2 produces a spectrograph about this:<\/p>\n

                        \"4238765ej00tg39ay00dad000ip00dep\"<\/p>\n

                          \n
                        • 12 grids, one shot each<\/li>\n
                        • The role is presented in simple matches<\/li>\n
                        • Red arrows indicate the movement and movement of people<\/li>\n
                        • Blue Arrow Marks the camera and pulls<\/li>\n
                        • Number of hours per grid lower right corner<\/li>\n<\/ul>\n

                          This is the input certificate for step three\u3002<\/p>\n

                          Step three: Feed Feedance 2.0, mirror directly to video<\/strong><\/p>\n

                          This is the best step\u3002<\/p>\n

                          Throws the spectroscopy generated by the previous step together with the role-setting diagram to Seedance 2.0, with a precise hint, to generate consistent video following the arrow direction and action logic in the spectroscopy\u3002<\/p>\n

                          Cue word templates<\/p>\n

                          Tips:<\/span><\/strong><\/p>\n

                          This is a role-setting diagram {Portrait}. Please follow the lens-by-scope action in this spectrograph\u3002<\/span><\/p>\n

                          Here's the text of the action script script:<\/span><\/p>\n

                          [Playing text lens for first step]<\/span><\/p>\n

                          Require:<\/span><\/p>\n

                            \n
                          • \n
                            Follow strictly the direction and trajectory of the lens as indicated in the spectrograph<\/span><\/section>\n<\/li>\n
                          • \n
                            The role looks and sets are consistent<\/span><\/section>\n<\/li>\n
                          • \n
                            It's natural to connect, not jump frame or mutation<\/section>\n<\/li>\n
                          • \n
                            The image style corresponds to the spectroscopy<\/section>\n<\/li>\n<\/ul>\n

                            Key Tips<\/p>\n

                              \n
                            1. The character set must be given<\/strong>:Seedance needs to know what the role looks like, or change the face of each frame<\/li>\n
                            2. The text specs are to be posted together<\/strong>: Pure spectroscopy may throw out the details. Text supplements lock the action logic<\/li>\n
                            3. Run the short clip test first<\/strong>: Don't run for 15 seconds at a time. Three to five seconds to try<\/li>\n<\/ol>\n

                              \"40c5a5d7j00tg3900c9000b7d000hn00csp\"<\/p>\n

                              \"54984624j00tg39k00bfd000ip00b9p\"<\/p>\n

                              Full Workstream Overview<\/p>\n

                              Three steps, a simple version:<\/p>\n

                              \n\n\n\n\n\n\n\n
                              \n
                              move<\/section>\n<\/th>\n
                              		\n
                              tool<\/section>\n<\/th>\n
                              		\n
                              enter<\/section>\n<\/th>\n
                              		\n
                              Output<\/section>\n<\/th>\n<\/tr>\n<\/thead>\n
                              \n
                              1 Writing Text Spectrum<\/section>\n<\/td>\n
                              		\n
                              GPT<\/section>\n<\/td>\n
                              		\n
                              scene + character + style requirements<\/section>\n<\/td>\n
                              		\n
                              Text Script with View\/Trace\/Time<\/section>\n<\/td>\n<\/tr>\n
                              \n
                              2 Draw tape speculation<\/section>\n<\/td>\n
                              		\n
                              GPT Image 2<\/section>\n<\/td>\n
                              		\n
                              Text Script<\/section>\n<\/td>\n
                              		\n
                              3x4 grid spectrometry (including motion arrows)<\/section>\n<\/td>\n<\/tr>\n
                              \n
                              3 Mirror Video<\/section>\n<\/td>\n
                              		\n
                              Seedance 2.0<\/section>\n<\/td>\n
                              		\n
                              Role + Mirror + Text script<\/section>\n<\/td>\n
                              		\n
                              Coherent Action Short Film<\/section>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/section>\n

                              It's common for rookies to roll over<\/p>\n

                                \n
                              1. Skip the first step straight<\/strong>: The logic of the movement is confused and the lens cannot be connected<\/li>\n
                              2. It's too delicate<\/strong>:Seedance's attention was distracted by the details of the scene, but the movement was lost<\/li>\n
                              3. No character charting<\/strong>It's not the same for each role. It's like changing an actor<\/li>\n
                              4. It's too short<\/strong>: Light Write \"Generate Video by this lens\" is not enough, locks the action logic again with text<\/li>\n<\/ol>\n

                                Final Thoughts<\/strong><\/p>\n

                                The most exciting thing about this work stream is..THE SCHEMATIC IS NO LONGER JUST A REFERENCE FOR PEOPLE, BUT A \u201cCONSTRUCTION DRAWING\u201d THAT CAN BE READ DIRECTLY BY THE AI VIDEO TOOL\u3002<\/strong><\/p>\n

                                The arrows and trajectories drawn by GPT Image 2, which Seedance can really identify and execute, are a key step from the \"random card\" to \"precision control\"\u3002<\/p>\n

                                The threshold of the tool is low and the process is not complex, but the effect is visible\u3002<\/p>\n

                                There are no particularly high thresholds for the entire process\u3002<\/p>\n

                                But the improvement in the quality of video is clear\u3002<\/p>\n

                                IF YOU'RE DOING THE AI SHORTS, THE AI COMICS, THE AI ADS OR THE AI MOVIE PREVIEWS\u3002<\/p>\n

                                This workstream is worth a try\u3002<\/p>","protected":false},"excerpt":{"rendered":"

                                A particularly interesting workstream has recently been drawn: GPT is used first to generate action-labeled spectroscopy, then feeds to Seedance 2.0 and directly produces a coherent short film. I've recently been tested for an amazing AI video stream. What's the worst thing that ever happened to an AI short? It's not pretty enough. It's: moving around. The shot's out of control. The character is like this for a second. The next second is like an actor. It wasn't easy to produce a video, and the result wasn't exactly what it wanted. But recently I found a very good combination: GPT + GPT Image + Seedance 2.0 to get GPT to design the action lens. Let GPT draw the specs with arrows. It's the last time you throw it to Seedance to generate a video. Top<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[149,144],"tags":[3264,956,5321,8574],"collection":[],"class_list":{"0":"post-53656","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"hentry","6":"category-jiaocheng","7":"category-baike","8":"tag-ai","11":"tag-8574"},"acf":[],"_links":{"self":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/53656","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/comments?post=53656"}],"version-history":[{"count":0,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/53656\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/media?parent=53656"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/categories?post=53656"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/tags?post=53656"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/collection?post=53656"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}