LTX-2.5 Gets Day-0 ComfyUI Support With Dynamic Compute And Multi-Shot Sequencing

What happened

LTX-2.5 received Day-0 support in ComfyUI on August 12, 2026, requiring an update to version 0.32.0 or later. The release pairs open-weight checkpoints with paired cloud tiers that natively handle multi-shot sequencing, synchronized audio, and dynamic compute allocation.

Context

LTX models progressively scaled past the four-second generation barrier, but uniform compute distribution forced a trade-off between quality and render time. Studios adopted ComfyUI to bypass API rate limits and retain direct checkpoint control. This update closes that gap by introducing dynamic allocation and native multi-shot sequencing within a single pipeline pass.

How it works

The architecture replaces standard VAE decoding with a Diffusion Video Decoder. Diffusion Fidelity Rendering allocates compute dynamically based on scene complexity instead of distributing it uniformly across frames. It generates structural motion and framing in an 8x temporally compressed latent space alongside high-fidelity keyframes, then routes both into a dedicated pixel-diffusion stage for the final output. A custom Gemma 4 12B text encoder handles prompt parsing, while an experimental duration head predicts clip length before diffusion begins, removing manual duration tuning from generation nodes. The model natively supports up to 4K resolution, synchronized audio and video streams, and frame rates up to 50fps. Open weights for the dev variant and a reworked distilled variant are hosted at Lightricks/LTX-2.5.

Our read

The press materials emphasize resolution and frame rate, but the actual workflow shift happens in how compute is billed and distributed. Dynamic allocation means static shots will consume less inference budget than complex pans, flattening the cost curve that usually penalizes detail-heavy generations. Native multi-shot support lets a single inference pass output connected sequences, which cuts post-production stitching time significantly. The distilled variant targets constrained hardware by trading some fidelity for prompt adherence and motion retention, though exact VRAM savings remain unlisted. The biggest friction point will be the custom Gemma 4 12B encoder. Standard LTX workflows used off-the-shelf text models that ran reliably on consumer GPUs. We do not know if this requires a separate high-VRAM allocation or runs fused into the main pipeline. Until hardware recommendations are published, studios relying on mid-tier cards will need to profile the text encoder alongside the diffusion weights before committing to local fine-tuning.

What this changes

Update ComfyUI to v0.32.0 or later before importing any LTX-2.5 checkpoint or workflow template. Import the Dev or Distilled weights directly for local deployment to bypass third-party API dependencies. If running constrained hardware, load the distilled variant and monitor VRAM allocation during text encoding, since exact system RAM requirements are unlisted. The auto-duration head means manual duration parameters in generation nodes can be removed from custom workflows. For batch production, expect variable compute costs per shot based on scene complexity rather than fixed pricing tiers.

License

The sources do not state a licence for the LTX-2.5 weights, and the reader should check the model card before building anything commercial on it. Until Lightricks publishes terms, assume research-only or restrictive usage applies to any derivative fine-tunes.

Key takeaways

• Day-0 ComfyUI support requires v0.32.0+ and pairs with open weights at Lightricks/LTX-2.5.
• Diffusion Fidelity Rendering dynamically allocates compute by scene complexity, altering batch costing for static versus detailed shots.
• Native multi-shot generation outputs connected sequences in a single pass, reducing post-production stitching time.
• Exact VRAM requirements and the governing licence are not stated; verify hardware compatibility and terms before commercial deployment.

Sources

1. LTX-2.5 Day-0 Support in ComfyUI — tier 1