LTX-2.5 shipped on August 12, 2026, and within 24 hours every LTX 2.3 user is asking the same question: do I need to re-download 60+ GB and rebuild my workflows, or does 2.3 still do the job?
Most upgrade comparisons answer with clips: same prompt, two renders, pick a winner. That ignores the actual cost of switching, which is not render quality — it is that 2.3 and 2.5 are not interchangeable. The files differ, the LoRAs are version-locked, the text encoder changed, and the workflow templates are new. An upgrade is a migration, not a swap.
This comparison is built on primary sources only: the LTX-2.5 model card, the LTX-2 GitHub README, the ltx.io LTX-2.5 page, and the ComfyUI LTX-2.5 tutorial. Checked August 13, 2026 — one day after release.
The three questions that decide the upgrade
Frame-by-frame preference changes with every point release. These three things do not:
- Do you need multishot? LTX-2.5's headline feature is native multishot — one generation producing connected shots that hold character, environment, lighting, voice, and style across cuts. LTX 2.3 generates a single continuous shot. If your work is single-clip B-roll, this feature buys you nothing.
- What is your migration cost? Every file is version-specific. LoRAs trained on 2.3 do not work on 2.5 without verification, and the reverse is also true. Your ComfyUI workflow needs the new templates. Your GGUF/quantized builds are gone.
- What does "faster" mean on your hardware? The speed claims are real but measured on datacenter GPUs. On a consumer card, the win comes from the 8-step distilled schedule and the int8 builds — which have their own hardware notes.
1. Multishot is the only true feature gap
The official model card describes LTX-2.5's multishot capability directly: multiple connected shots in a single pass, holding character identity, environment, lighting, voice, and visual style across cuts. Previous versions — including 2.3 — produce a single continuous shot per generation.
That changes what you can do in one run:
- Before 2.5: a "scene" meant one continuous take. Multiple shots meant multiple generations, then manual stitching, and hoping identity survived the rerolls.
- With 2.5: one prompt can describe a sequence of connected shots and the model holds continuity across them. Voice and visual style persist across cuts.
For anyone making narratives, ads with multiple beats, or any output that cuts — this is the entire upgrade argument. For silent, single-shot B-roll, it does not matter at all.
The official page also claims "cleaner motion" and "better prompt adherence" via the new Gemma 4 12B text encoder and a prompt enhancer that expands short prompts into detailed instructions. These are quality claims from the vendor; the multishot capability is a structural change you can verify yourself.
2. The migration cost is real: files, LoRAs, and workflows
This is the part most upgrade posts skip because it is boring right up until you hit it.
Files are not interchangeable. The LTX-2 repo is explicit: "Files are not interchangeable between the two models, and a LoRA only works with the model it was trained on." LTX 2.3 ships as single bundled checkpoints (transformer + VAEs + text projection in one file, with the Gemma 3 encoder downloaded separately). LTX 2.5 ships as a split, Comfy-aligned pack — one .safetensors per component: transformer, text encoder, video VAE, audio VAE, duration head, spatial upscaler, temporal upscaler, distilled LoRA.
LoRAs are version-locked — mostly. The model card says the large majority of LoRAs and IC-LoRAs trained on LTX-2.3 run on LTX-2.5 without changes, with a small number of exceptions — and tells you to validate your adapters before production use. The GitHub README is more cautious, stating a LoRA only works with the model it was trained on. Both statements are official; they describe the same situation at different risk tolerances. If you have a stack of 2.3 LoRAs, plan to test each one.
The text encoder changed. 2.3 used Gemma 3; 2.5 uses a custom Gemma 4 12B fine-tuned for LTX, with the text projection bundled. The README warns that Google's stock Gemma 4 release is not a substitute — loading checks the encoder's version against what the checkpoint was trained with (gemma4-12b-ltx-v1). You must download the LTX-specific encoder.
Quantized builds are version-specific. Your 2.3 GGUF/FP8/int8 files do not carry over. LTX-2.5 offers its own paths: ComfyUI-only int8 builds (comfy-int8-convrot), an NVFP4 build for Blackwell GPUs, and runtime FP8 casting (--quantization fp8-cast).
ComfyUI needs the new templates. ComfyUI ships three native LTX-2.5 workflows (T2V, I2V, FLF2V) that reference the split files. Your saved 2.3 workflows will not resolve the 2.5 files automatically.
The honest math: upgrading means re-downloading roughly 66 GiB (the official Quick Start's five-file set), getting gated-repo access approved on Hugging Face, downloading the new encoder and VAEs, and rebuilding workflows. That is a weekend, not an afternoon.
3. Speed: what the numbers actually measure
The ltx.io page claims a 10-second 720p clip in 6.8 seconds on their on-prem hardware (2× GB200 at steady state) — faster than real time. That number is about datacenter GPUs, and it is not what you will see on a single consumer card.
What scales to consumer hardware is the pipeline structure:
- The distilled checkpoint runs a fixed 8-step schedule with CFG=1 (8 steps in stage 1, 4 in stage 2).
- ComfyUI's default LTX-2.5 workflow uses the int8 + convrot build of the distilled transformer, which is what makes 8 GB-class GPUs viable (community reports of int8 running on 8 GB VRAM appeared within hours of release; treat as preliminary).
- The diffusion video decoder is higher quality but heavier; a lighter convolutional decoder exists for speed. On LTX-2.5 you get to choose.
The vendor's own artifact benchmark (98 prompts, automated scoring) ranks LTX 2.5 Pro first at 0.28 visible glitches per clip and LTX 2.5 Fast second at 0.39, with LTX 2.3 Pro at 0.74. It is vendor-published and preliminary — but it is the only standardized artifact comparison that exists, and it is internally consistent.
Don't re-download 66 GB to find out whether the upgrade matters to you. Generate with LTX 2.5 in the browser at ltx23.app and judge multishot and quality on your own prompts before you commit a weekend to the local migration.
Where each one is the obvious pick
Stay on LTX 2.3 when:
- Your work is single-shot clips and you never cut between shots
- You have many 2.3 LoRAs in production that you have not verified on 2.5
- Your hardware is marginal and your 2.3 quantized setup is already tuned
- You have no need for the new encoder, decoder, or duration head
Upgrade to LTX 2.5 when:
- You need multishot continuity (narratives, multi-beat ads, any output with cuts)
- Complex multi-subject prompts are failing — the Gemma 4 12B encoder is the official fix
- You want the improved distilled model's quality at 8 steps
- You are starting fresh anyway — new project, new workflows, nothing to migrate
Honestly? For many users the answer is "try it before you migrate." The weights are free under the community license, the demo is free, and the decision cost is a weekend of downloads rather than money. Just do the migration deliberately — verify your LoRAs, rebuild the workflows, and keep your 2.3 folder intact until 2.5 has proven itself on your exact prompts.
What this comparison deliberately does not claim
No fresh side-by-side quality scores — LTX-2.5 is one day old and third-party benchmarks do not exist yet. The artifact and speed figures quoted are vendor-published (ltx.io), measured on datacenter hardware, and labeled as preliminary by the vendor. Treat them as directional, and verify on your own GPU and prompts. Everything structural — file layout, encoder change, LoRA versioning, pipeline step counts, multishot capability — is from official documentation and is the reliable part of this comparison.
FAQ
Can I use LTX 2.3 LoRAs on LTX 2.5? The model card says the large majority run without changes but tells you to validate adapters before production use; the GitHub README states a LoRA only works with the model it was trained on. Test each one.
Do I need to download new files? Yes. LTX-2.5 is a split pack (~66 GiB for the Quick Start set) and 2.3 files are not interchangeable with 2.5 files. The Gemma 4 12B encoder is LTX-specific and required.
Is LTX 2.5 faster than LTX 2.3? The vendor claims 6.8 s for a 10-second clip on 2× GB200. On consumer hardware, the practical speedups are the 8-step distilled schedule (CFG=1) and int8 builds. Compare on your own hardware.
Does LTX 2.5 do multishot? Yes — multiple connected shots in a single pass, holding character, environment, lighting, voice, and style across cuts. LTX 2.3 produces a single continuous shot.
Is LTX 2.5 free? Open weights under the LTX-2.x Community License; free commercial use under $10M annual revenue. Same license family as 2.3.
Bottom line
Upgrade for multishot and the new encoder and decoder — those are structural. Everything else (speed, quality) is worth verifying on your own hardware before you spend a weekend migrating 66 GB, re-approving gated access, and re-verifying every LoRA. Start by testing LTX 2.5 in your browser at ltx23.app, then decide if the migration is worth it.
Sources
Verified August 13, 2026 against primary documentation:
- Lightricks/LTX-2.5 — Hugging Face model card — what's new (multishot, DFR, diffusion decoder, Gemma 4 12B, prompt enhancer, duration predictor, improved distilled), split file layout, constraints, license
- Lightricks/LTX-2 — GitHub repository README — "files not interchangeable" statement, LoRA versioning, Quick Start (~66 GiB), Gemma 4 12B version check, pipeline step counts, int8/FP8/NVFP4 paths
- LTX-2.5 official model page (ltx.io) — capability claims, speed measurement (6.8 s, 2× GB200), artifact benchmark, API pricing, $10M license threshold
- ComfyUI official LTX-2.5 tutorial — three native workflows, model storage paths, gated-repo warning
- LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler — Hugging Face — detailing IC-LoRA for the DFR refinement stage
Speed and artifact figures are vendor-published (ltx.io), measured on datacenter GPUs, and labeled preliminary by the vendor. Licensing summaries are not legal advice; read the license text in the repository before making a commercial deployment decision.
