Runway Aleph 2.0 vs Luma Ray 3.2: Editing vs Controlled Generation

Runway Aleph 2.0 and Luma Ray 3.2 both give creators more control than a single text prompt, but they begin from different production questions. Aleph asks how to turn footage you already value into a revised version. Ray asks how to direct a generated or transformed sequence through explicit visual beats.

Runway launched Aleph 2.0 on May 21, 2026, and Luma released Ray 3.2 on June 9. Both target professional short-form video and support 1080p output. Comparing them only by visual quality misses the more useful distinction: what the creator can supply as control and what must remain unchanged.

Aleph: preserve the source, change the target

Aleph 2.0 accepts an existing clip up to 30 seconds and attempts localized editing. In Edit Studio, the user describes a change, previews it on a still frame, refines that frame, and then asks Aleph to propagate the look through the relevant video.

The source supplies the performance, action, composition, camera movement, timing, and cuts. Common edits include changing a product or colorway, replacing a background, altering season or lighting, removing a distraction, or restyling the footage. Multi-shot support can carry an edit across relevant cuts rather than requiring a separate job per shot.

The preservation requirement is the hard part. The requested object may change correctly while a face, reflection, logo, or background detail drifts. Aleph’s product promise is stronger localization, not pixel-identical protection. Reviewers must compare the whole output with the source.

Ray: direct the sequence through keyframes

Ray 3.2 supports up to 16 keyframes inside one clip. A creator can define visual states across time and use them to guide pacing, motion, camera paths, and narrative beats. That control is closer to a storyboard distributed through the shot than to one starting image and a sentence.

Luma documents clips up to 20 seconds at 1080p, enhanced performance tracking, expressive facial performance for as many as eight faces, native HDR generation, and 16-bit EXR export. Those formats matter in compositing and color pipelines because the generated asset can carry more range and precision than a normal delivery MP4.

Ray also includes Reframe for adapting aspect ratios, extending the frame, or changing a background while preserving original lighting. Video-to-video workflows can transform existing footage. It is therefore too simple to call Ray “generation only.” Its center of gravity is directed synthesis, but its control surface overlaps with editing.

Frame target versus sequence plan

Aleph’s frame-first method answers, “What should this edited element look like?” Ray’s multi-keyframe method answers, “What should happen at several points through this shot?”

Choose Aleph when one approved moment, performance, or cut should survive and the change is localized. Choose Ray when movement between planned states is the creative task. Examples:

This table is a workflow recommendation, not a quality verdict. A source, prompt, subject, and delivery format can change the result.

For a mixed campaign, the tools can also be sequential rather than exclusive. A team might generate a storyboard-led insert with Ray, place it into an existing edit, then use Aleph for a localized environmental variation. Each handoff should use a reviewed master so one model does not amplify an unnoticed error from the other.

Duration and output are not interchangeable

Aleph’s 30-second limit is longer, but it modifies supplied footage. Ray’s 20-second limit covers generated or transformed shots and includes professional HDR and EXR options. A longer limit does not mean a model is better for long-form work; both still operate on short clips that must be assembled.

Delivery requirements should guide the choice. A social team may value an immediate 1080p visual variant. A VFX team may prefer EXR frames for compositing even if that output requires more storage, rendering, and finishing.

Ray’s public API page lists separate prices by task, duration, and resolution. As of July 2026, five-second text- or image-to-video was shown from $0.15 at 540p to $1.20 at 1080p, and ten seconds from $0.45 to $3.60. HDR was listed at twice SDR pricing and HDR plus EXR at three times. Rates are approximate, rate limits apply, and the pay-as-you-go tier has no latency SLA.

Runway provides Aleph through paid plans and Edit Studio; its launch announcement does not give one universal per-edit price. Use live plan information and actual credit records rather than forcing a misleading dollar comparison.

Access and automation

Aleph launched in Runway’s desktop web Edit Studio. Its frame preview and interactive refinement suit a human creative loop. Runway also lists Aleph among its broader developer platform models, but teams should verify which controls are exposed in the specific surface they plan to use.

Luma made Ray 3.2’s full control surface available through an API at launch. That is attractive for internal tools, render pipelines, and repeatable generation. API access does not remove the need for asynchronous job handling, storage, moderation, timeouts, and review.

An API workflow should store keyframes, prompts, source hashes, model version, output settings, task ID, and cost. Without those records, reproducing an approved shot becomes guesswork.

Failure modes differ

Aleph can fail by changing too much or applying the requested edit inconsistently. Review the untouched scene, transitions, object edges, shadows, text, and identities.

Ray can fail in the transitions between keyframes. A sequence may hit the endpoints while producing implausible motion in between. Performance transfer can lose gesture nuance; multiple faces can swap attributes; camera paths can accelerate unexpectedly. HDR output may be technically valid but still need grading.

For both, test difficult moments rather than only a representative thumbnail. View the whole clip, inspect frame boundaries, and compare against source or storyboard references.

Rights, identity, and disclosure

Editing a product in owned campaign footage is different from changing a person’s appearance or performance. Obtain rights for source media and references, secure consent for identity changes, and record what was altered. Do not use preservation claims to imply that a synthetic output is an untouched record.

Generated assets should pass brand, legal, and factual review. An AI-created package can contain incorrect text, and a localized scene can imply availability or endorsements that do not exist in a market.

One finishing stage for both approaches

Regardless of which model creates the approved visual shot, campaigns usually need several clips assembled and an exact soundtrack applied. Those are deterministic production steps rather than further visual invention.

Medux offers external MCP workflows for merging video clips with Claude and adding an audio track. There is no claimed native integration with Runway or Luma. The common handoff is simply an exported video file that an authorized Claude workflow uploads to a separate service.

A model-neutral pipeline can:

1. generate or edit shots in Aleph or Ray;
2. review each shot for visual continuity, rights, and disclosure;
3. normalize accepted files to the intended dimensions and frame rate;
4. merge them in the approved narrative order;
5. add a mastered soundtrack or narration;
6. verify output duration, audio presence, and file integrity;
7. retain source, generation settings, task IDs, and approvals.

This shared finishing layer also makes a fair comparison easier. Teams can evaluate Aleph and Ray on the creative shot without letting different assembly or audio processes distort the result.

Aleph 2.0 and Ray 3.2 are not substitutes in every job. Aleph protects the value already captured in footage; Ray expands the ability to direct what has not yet been shot. The right model begins with the production constraint, not the leaderboard.