Generative video usually makes the creator wait for the most expensive artifact before revealing how it interpreted the request. A prompt is submitted, a full clip renders, and only then does the team discover that the product shape, lighting, or background is wrong. Runway’s frame-first workflow moves that visual decision earlier.

Aleph 2.0 and Edit Studio, released on May 21, 2026, let a creator preview an intended edit as an image. Once the frame looks right, Aleph applies that target to the relevant parts of an existing video. The model supports up to 30 seconds of 1080p footage and can carry edits across multiple shots.

This does not turn video into a collection of independent still images. Motion and continuity remain model problems. The production improvement is that art direction, temporal synthesis, and final finishing become separate approval stages.

Why the first approval should be visual

Text is efficient for intent but ambiguous for appearance. “Make the package premium” could mean a darker label, metallic material, restrained typography, or dramatic lighting. A model may choose one valid interpretation that is completely wrong for the campaign.

A target frame makes the request inspectable. Brand reviewers can compare color, shape, product details, scene composition, and mood before authorizing a video render. The frame also gives Aleph a more concrete reference than prose alone.

This resembles a style frame in conventional production. Agencies have long approved key visuals before committing to animation, shooting, or finishing. Edit Studio brings a similar gate into generative modification: type the change, see it on an image, refine it, then generate.

The benefit is fewer avoidable full renders. It does not guarantee fewer total attempts; an approved image can still fail when propagated through motion. Measure accepted-video rate, not only accepted-frame rate.

Selecting the right frame

The most attractive frame is not always the most informative one. A useful target should reveal the subject at a scale and angle that exposes the requested change. If a shoe rotates, a straight-on frame cannot define the heel. If a hand covers the product, an unobstructed frame says little about the later interaction.

Before creating the preview, scan the source for:

For a difficult clip, one target frame may be insufficient as a creative specification even if the interface uses one frame to initiate the edit. Keep additional reference images and shot notes for review.

Define change and preservation together

A generative edit specification should contain two lists: what changes and what stays fixed. The first might say “replace the green bottle with the approved blue bottle.” The second should protect the actor, hand position, table, lighting direction, camera motion, timing, and all existing text.

Aleph 2.0 is designed for localized edits and input-video preservation. Explicit preservation criteria help reviewers detect collateral drift. They also prevent a visually pleasing output from passing when it changes a legally approved claim or recognizable person.

Create a compact acceptance rubric:

Area Acceptance question
Target Is the intended object, color, or environment correct?
Identity Are faces and bodies unchanged unless authorized?
Action Does original timing and physical interaction remain coherent?
Scene Are unedited objects, lighting, and camera behavior preserved?
Text Are labels and claims exact and legible?
Continuity Is the edit consistent across frames and shots?

Runway’s preservation claims should be tested against this rubric rather than treated as a guarantee.

The multi-shot opportunity

Aleph can apply an edit across relevant shots in a clip containing cuts. For campaigns, that can be more valuable than an isolated perfect shot. A product variant may appear in a hand, on a table, and in a closing beauty shot. Editing each independently creates inconsistency and more review work.

Multi-shot generation needs shot-level reporting. Mark where the object should change and where it should not. Review the first and last frames around every cut. A model may carry the edit into a shot where the subject is absent or fail to recognize the same subject after a radical angle change.

If one shot repeatedly fails, split it rather than regenerating the entire sequence. The 30-second limit is a maximum, not an optimal unit of work. Smaller jobs can isolate difficult transitions and reduce the cost of accepting good sections.

Version the creative state

Frame-first work produces several related artifacts: original video, chosen source frame, prompt, preview variants, approved target frame, generated videos, and review notes. Store them as a versioned set.

A useful record includes:

Without this record, a team can approve a frame from one attempt and accidentally publish video from another. Versioning also makes it possible to reproduce a campaign variant or explain what was synthetically altered.

Limits of a still-frame approval

A still image cannot prove temporal coherence. It cannot show whether hair flickers, a hand passes through a product, shadows jump, or an object changes shape during a camera move. It also cannot validate audio synchronization or the timing of a visual claim.

Inspect the final video at normal speed, reduced speed, and around cuts. Compare it with the source side by side. Automated frame sampling can help find drift, but a human should review identity changes, brand meaning, and physical plausibility.

Edit Studio is available on paid Runway plans in the desktop web app. Usage follows current plan and credit terms. Because every full regeneration can consume time and credits, reject flawed target frames early and avoid vague “make it better” loops.

Frame-first does not mean one-tool production

Aleph’s role is generative visual modification. Other operations may have different control requirements. A campaign could need a separate character replacement, exact clip order, approved soundtrack, fixed logo placement, or delivery encoding. Combining all changes into one open-ended generation makes failures harder to locate.

Keep stages modular. An output should move forward only after the relevant review passes. If an edit changes a person or character, require consent and a dedicated identity review. If a clip assembly is deterministic, verify duration and order automatically.

Downstream character work and assembly

Medux provides external media operations that Claude or Codex can call through separately configured MCP access. It is not integrated into Aleph or Edit Studio. The Claude MCP character-replacement tutorial uses a source video and reference image, uploads them, creates an asynchronous task, polls status, and downloads the result.

That operation is higher risk than changing a product color. A production sequence should not stack it onto Aleph without a new approval. One possible order is:

  1. approve the source, rights, and target frame for the Aleph environment or product edit;
  2. generate and inspect the Aleph output;
  3. if authorized, submit the accepted clip and reference image for character replacement;
  4. store each task ID and never repeat a billable job after a lost response;
  5. inspect identity, motion, and disclosure requirements;
  6. assemble accepted clips in a deterministic final stage.

The Medux Codex MCP video merge tutorial shows that final assembly pattern: upload approved MP4 clips, create one ordered merge task, poll, download, and verify the result. This is a natural complement to frame-first editing because it does not ask another generative model to reinterpret the visual work.

Frame-first editing is valuable less because one image controls everything than because it creates a better decision point. The target look can be approved before temporal generation, the video can be approved before identity changes, and every creative stage can be approved before deterministic finishing. That structure makes generative production easier to diagnose, govern, and repeat.