Why these clips hold together

Hands, held objects and precise poses are common failure points in generated video. First-and-last-frame conditioning gives the model two visual anchors instead of asking it to invent both ends of the shot. Filming the opening frame preserves the real hand and object at the start; designing the closing frame specifies the desired destination. The model still has to generate the movement between them, so neither anchor guarantees that the transition will remain physically or visually consistent.

The source creator demonstrates this workflow using a model labelled MiniMax H3 Max inside Magnific. MiniMax separately documents first-and-last-frame image conditioning for its MiniMax-H3 model. The underlying technique is transferable, but the two names should not be assumed to identify the same model configuration, and third-party hosts may expose different controls. The steps below follow the creator's Magnific workflow; the technical limits cited are from MiniMax's own H3 API documentation unless otherwise stated.

Step 1: Film the start, do not generate it

Hold the object the way you want it held and record a few seconds of yourself. Keep the framing you intend to publish, vertical 9:16 for a Reel, and keep the object clear of the frame edges. Then pause the video on the cleanest moment and take a screenshot. That screenshot is your Start Frame.

Two details decide the result here. First, light: the end frame will be matched to this frame's look more than to your prompt, so shoot in the light you want the finished clip to appear to have. Second, headroom: leave space around the object for the transformation to happen in, because the model has to travel from this picture to a different one inside the same frame.

Step 2: Design the end frame as a picture, not a wish

The End Frame in the source tutorial is a finished rendered scene, a bedroom nightstand composition, not a sketch and not a prompt. Build yours the same way: a complete still image, at the same aspect ratio as the start frame, showing exactly where the shot should land.

Match the two frames on everything except the one change you want. Same camera distance, same hand position if a hand is still in shot, similar brightness. Every difference you leave in besides the intended one is a second change the model has to invent, and invented changes are where these clips wobble.

MiniMax's documentation sets hard limits worth planning around: frame images can be 256 to 5,760 pixels per side, with an aspect ratio between 2:5 and 5:2. In first/last-frame mode the output follows your frames, so you do not get to pick a different aspect ratio at generation time. Decide the shape when you make the frames.

Step 3: Choose a generator with both frame inputs

In the source reel, Mustafa Alnajjar uses Magnific's Video Generator and selects an option labelled MiniMax H3 Max. Before spending credits, confirm that the model currently exposes both a Start Frame and an End Frame input. Upload the screenshot from Step 1 as the start and the designed still from Step 2 as the end.

If you use MiniMax's own H3 interface or another host, look for its first-and-last-frame image-to-video mode. Do not assume that every service offering H3 exposes the same input slots, duration settings or output resolutions.

Use first-and-last-frame generation rather than the separate reference-generation mode. MiniMax documents these as distinct workflows, and Runway specifically says its H3 keyframes cannot be combined with reference images, video or audio in the same request. Other hosts may implement their controls differently, so follow the input requirements of the service you use.

Start with a short, lower-resolution test to check whether the transition works. MiniMax's H3 API supports whole-second durations from 4 to 15 seconds and outputs at 768P or 2K; a third-party host may impose different limits. Once the movement is right, you can generate another version at 2K. MiniMax also documents a separate regeneration workflow for qualifying 768P H3 videos, so check whether your chosen platform supports that option before paying to create the shot again.

Step 4: Prompt the path, not the pictures

Both frames already fix the subject, the light and the palette, so a prompt that redescribes them wastes its influence. Spend the prompt on the one thing neither frame contains: how the scene gets from the first picture to the second.

The camera remains fixed as the hand holds the object in place. [static] The object gradually transforms into the final scene shown in the End Frame. The movement is continuous and physically plausible, with no cuts or sudden camera movement. Lighting and shadows change gradually to match the final image.

This is ALFW's example prompt, not a quotation from Mustafa Alnajjar's reel. It describes the movement between the two supplied images rather than attempting to recreate either image from text. MiniMax documents camera cues such as [static], [pan] and [zoom]; their effect remains guidance, not a guarantee. Add sound separately during editing unless your selected generator explicitly supports audio output for this mode.

Two choices in that prompt are doing work. Naming what stays still — the camera, the hand — matters as much as naming what moves, because H3 can plan multi-shot sequences on its own and a transformation needs the opposite instruction. And describing the light change matters because the two frames rarely match exactly, so the model needs permission to grade one into the other instead of jumping.

Step 5: Change one thing at a time

If the transition looks smeared or inconsistent, compare the Start Frame and End Frame before rewriting the prompt. Try reducing the visual distance between them: preserve the camera angle, object scale, hand position and lighting wherever possible.

If the transformation is too large for one short clip, consider an intermediate still and make two separate generations, then join them in an editor. If the lighting jumps, adjust the end image to resemble the opening frame more closely.

These are troubleshooting techniques, not guaranteed fixes. Save successful frame pairs and the prompts used with them so you can compare what changed between attempts.

What this cannot do

It cannot rescue a start frame you would not publish as a still. It cannot hold a face or a logo through a large transformation reliably, and it cannot extend a transformation past 15 seconds in one generation. It also depends on a host exposing both frame slots: some third-party H3 interfaces enable the first frame only, so check for a last-frame slot before paying for credits, rather than after.

Magnific reports that approximately one in six clips generated on its platform during a summer measurement period used both a first and a last frame. It also reports that most generations began with a user-supplied image. These are Magnific's own platform statistics, not independent industry-wide adoption figures. Their relevance here is limited but useful: first-and-last-frame generation is one established workflow among several, not a requirement for making AI video.

Sources

  1. [1] Instagram tutorial, Mustafa Alnajjar (@__mustafa.alnajjar), October 31, 2026 (the reel was read first-hand for this article: film yourself holding an object, screenshot it as the Start Frame, choose the option labelled MiniMax H3 Max in Magnific's Video Generator, upload the Start Frame and an End Frame — his is a rendered bedroom nightstand scene — and add a text prompt describing the motion; his example prompts were left in his comments and are not reproduced here)Read source
  2. [2] MiniMax API documentation, Video Generation, updated October 30, 2026 (read first-hand: Image-to-Video is “Prompt + first-frame image and/or last-frame image”; output is 768P or 2K at whole-second durations from 4 to 15 seconds; the first/last-frame entry takes 0, 1 or 2 images at 256 to 5,760 pixels per side with aspect ratios from 2:5 to 5:2; the prompt limit is 7,000 characters; camera instructions such as [pan], [zoom] and [static] go directly after the description they apply to; reference generation is documented as a separate mode; and a separate Video Regeneration workflow turns a qualifying 768P H3 video into 2K)Read source
  3. [3] MiniMax, “MiniMax H3 Is Now Open Source” (the H3-Base-FL2VA first-and-last-frame mode takes zero, one or two input images)Read source
  4. [4] Runway, MiniMax H3 (Hailuo 3) product page (read first-hand: “Keyframes are their own mode and cannot be combined with reference images, video or audio in the same generation” — Runway's rule for its own H3 hosting; Runway's hosted limits also differ from MiniMax's API, at 5 to 15 seconds and a 6,000-character prompt)Read source
  5. [5] Magnific, Video Generator documentation (read first-hand: the generator route, a model picker that includes MiniMax, and image-to-video with a start image plus an optional end image enabled by an End image toggle)Read source
  6. [6] Magnific, AI Video Generator page, updated October 22, 2026 (read first-hand: Magnific reports 79% of clips start from a user-supplied image — 62.3% from a start frame only, 16.8% from a first and a last frame, 20.4% from text alone — measured over one summer; the page also describes first-and-last-frame work as a storyboard workflow)Read source