FL2VA Tutorial: MiniMax H3 First/Last Frame Video Workflow
FL2VA (First-Last Frame to Video Animation) is the practice of giving the model a first frame and a last frame, then letting MiniMax H3 generate the motion that connects them. In 墨川导演台 (MiniMax H3 Director), you build this workflow by generating both frames as images in the AI 制图 module, attaching them to a single shot card in the 剧本库, and running the shot through 项目库 — either one shot at a time or as a batch. The result is a clip that starts exactly where you designed it and lands exactly where you planned, which is the whole reason editors love this method for AI short dramas and commercial spots.
Below is the practical version: how to set it up, how to write prompts for each frame, and where consistency usually breaks.
---
What exactly is FL2VA, and why use it instead of text-to-video?
FL2VA means you control both ends of the motion instead of only describing it in words. Text-to-video gives the model freedom over the opening and closing composition; FL2VA removes that freedom and turns the model into an interpolator with style.
Why that matters:
- Editability. A clip that begins on a known composition cuts cleanly against the previous shot. No hunting for a usable frame.
- Continuity in series work. In AI short dramas, shot B must start where shot A ended. First/last frame pairs make that a design decision rather than a lucky accident.
- Fewer wasted generations. You see and approve both endpoints as stills before spending GPU time on video.
- Better prompt economy. You describe *motion* in the video prompt, not appearance — appearance is already locked in the images.
In 墨川导演台, the two frames live as assets in the 资产库 and as references on the shot card, so they stay attached to the project instead of floating in a downloads folder.
---
How do I generate a usable first frame and last frame?
Generate both frames in AI 制图 (text-to-image or image-to-image), then upload them as references on the shot card. Treat them as a matched pair, not two separate pictures.
Step by step:
- Write the shot in the 剧本库 first. Use the AI 编剧 conversation to produce a six-段 script and storyboard draft — it writes the draft into your script library, so the shot card already has intent, duration feel, and dialogue attached.
- Generate the first frame in AI 制图. Describe composition, lens feel, lighting direction, wardrobe, and expression. Keep the description tight — one clear subject, one clear action pose.
- Generate the last frame from the first frame. Use image-to-image with the first frame as the reference so character identity, wardrobe, and color grade carry over. Change only what the motion should change: body position, camera distance, light shift, prop state.
- Check the pair side by side. Ask yourself: *could a camera physically travel from A to B in the shot length I have in mind?* If the answer is no, pull the last frame closer to the first.
- Save both into the 资产库 as character/scene references so the same face and location can be reused across other shots in the project.
A useful habit: generate three candidate last frames and pick the one that is closest to the first frame in framing. Interpolation quality drops fast when the two frames disagree about where the camera is.
---
What should the prompts for each frame actually contain?
The first frame prompt describes *state*; the last frame prompt describes *the changed state*; the video prompt describes *only the transition*. Keeping these three jobs separate is the single biggest quality lever.
First frame prompt — build the "before":
- Subject: age, build, wardrobe, exact expression
- Framing: shot size (wide / medium / close), camera height, lens feel
- Light: direction, hardness, color temperature
- Environment: location, time of day, key background elements
- Style: film stock, grade, reference genre
Last frame prompt — change only the deltas:
- Body: "same character, now turned three-quarters toward camera, right hand raised"
- Camera: "same location, camera pushed in to medium close-up"
- Light: "same room, now lit by the practical lamp on the left"
- Prop/environment: "the door behind her is now open"
Video prompt — motion only:
- "She turns from the window toward the camera, hair moving with the turn, slow push-in, natural handheld micro-shake, no cuts."
Three rules that save time:
- Never re-describe the character's face in the last frame prompt. Let image-to-image carry identity; re-describing invites drift.
- Name the camera move explicitly. If you don't, the model invents one and your endpoints stop matching.
- Keep one dominant action per shot. Two actions in one FL2VA pair usually means neither reads clearly.
---
How do I run the shot in 墨川导演台 and keep characters consistent?
Attach both frames to the shot card in 项目库, then generate — single shot for testing, batch for the full sequence. Consistency comes from the 资产库 references, not from repeating adjectives in prompts.
The working loop:
- Open the project in 项目库 and select the shot card that carries your script and storyboard notes.
- Attach references: first frame, last frame, plus the character reference image and any scene or prop references from the 资产库.
- Paste the motion-only video prompt and set your shot length to match the pacing you planned in the script.
- Run a single-shot generation first. Watch the 项目库 version history — each run is kept as a version, so you can compare rather than overwrite.
- Check the AI 日志 / 系统志 if something fails. Generation logs and system status are visible in real time, which tells you whether the problem is your prompt or the backend.
- Batch generate the remaining shots once the first one looks right. Batch queuing runs through the ComfyUI workflow on your own GPU (local or cloud), so the reference images travel with each job.
- Lock the good version and move to audio: 音频库 for voice cloning, lip-sync, BGM and SFX, or AI 配音 (IndexTTS 2.5, multi-language with emotion control) if you're voicing dialogue.
Consistency checklist:
- Same character reference image across every shot in the scene
- Same scene reference for every angle in the same location
- Last frame of shot N ≈ first frame of shot N+1 (match framing, not just content)
- One wardrobe, one grade, one lens language per scene
If you're producing a series, keep each episode as its own project so version history stays readable.
---
Do I need cloud GPU for this, and how do I set it up?
No — 墨川导演台 runs in local deployment on your own GPU machine, in cloud deployment over SSH to a GPU server you rent, or via API 调用. Cloud is the usual choice for batch work.
Cloud deployment, as tested on 仙宫云:
- Register on 仙宫云 at
https://www.xiangongyun.com/register/B1H9M6— you must register fresh and enter invite code B1H9M6. - Complete verification, then send your 仙宫云 account ID (the long UUID form) to customer service.
- Once your test account is opened, the 导演台 private image becomes visible — deploy an instance from it.
- On the running instance, click SSH on the right panel and copy the host, port, and password.
- In the 导演台 云部署 page, fill in SSH Host / SSH Port / SSH password, then click 验证 SSH.
- After verification passes, click 启动云端 ComfyUI and wait for the green success indicator.
- Write your script and test a generation.
The same 云部署 page also supports AutoDL, 阿里云, 腾讯云, and custom SSH servers. Local tunnel ports are isolated per user, so multiple people on one machine don't collide. For hardware requirements and current setup details, check with the official site or 客服微信 ymcandai — those numbers change and shouldn't be guessed.
---
FAQ
Can I use FL2VA without a last frame?
Yes — a first frame alone works as image-to-video, and the model will invent the ending. You lose editability at the tail of the clip, which is why FL2VA is preferred for anything that has to cut against another shot.
Why does my character's face change between the two frames?
Almost always because identity was described in words instead of carried by a reference image. Generate the last frame via image-to-image from the first frame, and attach the character reference from the 资产库 to the shot card.
How long should a single FL2VA shot be?
Keep it to one clear action. If the motion needs more time than your planned shot length, split it into two shots with a matched pair of frames at the join.
---
Next step: pick one 4–6 shot sequence from an existing script, build matched frame pairs for each shot, and run them as a batch. Compare the batch against a text-to-video pass of the same sequence — the difference in cut quality is usually obvious on the first watch.
Official site: https://www.ymcdirector.com · Blog: https://blog.ymcdirector.com · Support: info@ymcdirector.com
客服微信: ymcandai · 淘宝: search 「杨墨川AI」 · 抖音 / 小红书 / 视频号: search 「杨墨川」