Cover image for Wan 2.7, LTX-2.3, and Seedance 2.5 Reframe Generative Control
ANALYSIS ReelStack Editorial August 16, 2026

Wan 2.7, LTX-2.3, and Seedance 2.5 Reframe Generative Control

A wave of flagship model updates introduces native multi-image grids, 30-second continuous generations, and direct first-to-last-frame conditioning.

TL;DR

Alibaba released Wan 2.7 with native 9-grid multi-image conditioning, while ByteDance announced Seedance 2.5 with 30-second clips and 50 multimodal reference inputs (Digital Applied). Together with Lightricks deploying LTX-2.3 and Kling adding dual-frame constraints, video generation has shifted from prompt gambling to deterministic keyframe sequencing.

What Happened

Alibaba rolled out Wan 2.7 as the newest entry in its open video line (Pinggy). The update introduces first-and-last-frame control, a 9-grid multi-image input interface, instruction-based video editing, and a 5,000-character prompt capacity. The 9-grid input surface allows directors to pass multi-angle character sheets and location references directly into diffusion passes without relying on external control nets.

Simultaneously, ByteDance unveiled Seedance 2.5 in enterprise beta ahead of its planned public rollout (Mind Studio). The architecture expands sequence duration to 30 continuous seconds, ingests up to 50 multimodal reference inputs per project, and provides localized in-canvas redraw tools. This release builds directly on the previous Seedance 2.0 release, which expanded distribution through Jimeng AI and CapCut (NBC News).

Lightricks launched LTX-2.3 to update its open weights video pipeline (Pinggy). The model includes a rebuilt VAE for reduced motion artifacting, a text connector enlarged fourfold for prompt parsing, an upgraded HiFi-GAN vocoder rendering 24kHz stereo audio, and native vertical rendering up to 1080x1920 pixels. The LTX team also introduced LTX Elements, letting artists composite character, wardrobe, and environment layers before final diffusion (YouTube).

Kling updated its web interface with direct start-frame and end-frame uploads (YouTube), addressing interpolation stability across camera transitions. At the open-source pipeline layer, ComfyUI issued an update on June 3, 2026, fixing the Seedance 2.0 1080p first/last-frame stretch bug while integrating Bria transparent video background nodes with native alpha channel export in SaveWEBM (ComfyUI Changelog).

Why This Matters

Generative video production previously stalled on temporal consistency and short duration limits. Creators spent significant compute generating 4-second fragments, manually painting frames in third-party software, and re-injecting them as single image prompts. The combination of Wan 2.7's 9-grid input and Seedance 2.5's 50-reference window converts prompt-based guessing into controlled visual reference matching.

First-frame to last-frame interpolation is now standard across both commercial platforms like Kling and open models like Wan 2.7. When directors lock both ends of a sequence, camera movement becomes predictable rather than chaotic. ComfyUI integrating alpha channel video output directly from Bria nodes means compositors can extract moving subjects straight to clean transparency without intermediate chroma-key passes or rotoscoping artifacts (ComfyUI Changelog).

Audio integration has also advanced. Rather than relying entirely on downstream sound design tools, LTX-2.3 renders synchronized 24kHz stereo audio alongside video frames using its HiFi-GAN vocoder pipeline (Pinggy). This reduces the latency between visual motion generation and sound effect placement.

For AI Filmmakers

Directors working on narrative shorts can now establish shot continuity without multi-model rigging. By organizing pre-production character turnaround sheets with the Cinematic Sheet Composer, creators can supply consistent 9-grid reference frames directly into models supporting multi-image conditioning like Wan 2.7.

Precise camera control no longer requires heavy negative prompting. With Kling 2.5 and Wan 2.7 locking head and tail frames, you can set precise framing at both ends and specify the exact trajectory using the Camera Movement Builder. This eliminates unwanted morphing across long pans, pedestal moves, and orbit shots.

Editors and VFX leads benefit immediately from native alpha video. If you run a local ComfyUI instance, updating to the latest build gives you the Bria transparent video node (ComfyUI Changelog), outputting WebM files with built-in alpha channels for immediate layering over background plates in Premiere Pro or DaVinci Resolve.

What To Do Now

  1. Update your ComfyUI installation to pull the June 3 release, which resolves Seedance 2.0 aspect-ratio stretching and adds Bria alpha nodes (ComfyUI Changelog).

  2. Prepare clean 9-image reference sheets featuring distinct character angles and wardrobe details for testing Wan 2.7 multi-image conditioning.

  3. Test LTX-2.3 in native vertical 1080x1920 format for mobile delivery pipelines to evaluate its 24kHz audio vocoder outputs (Pinggy).

  4. Revisit shot lists to replace stitched 4-second micro-shots with single 15-to-30-second continuous generations in Seedance 2.5 test builds (Digital Applied).

Do NOT burn compute credits running unconditioned text-to-video generations for complex scenes when dual-frame or multi-reference inputs are available on the same platform.

The Bigger Picture

The generative video landscape in 2026 has transitioned from raw visual synthesis to structured production pipelines. The simultaneous expansion of prompt context windows to 5,000 characters in Wan 2.7, extended 30-second timelines in Seedance 2.5, and multi-element layering in LTX indicates that frontier labs are optimizing for editorial control rather than novelty clips (Digital Applied).

Sources & further reading click to expand
Powered by ReelStack

Help keep this running

Your tip funds servers, models, and the time it takes to ship new tools faster. Set any amount below — every bit helps.