What Is Model Distillation? Dev vs Distilled Explained

Model distillation is the practice of training a faster student model to reproduce a larger teacher model's output in far fewer computation steps.

In generative video that becomes immediately practical. The LTX-2.5 release publishes two transformer checkpoints, a dev transformer and a distilled transformer, and the distilled one runs on a fixed 8-sigma schedule instead of the 40 steps a guided two-stage run baselines at.

This post answers the question that brings most people here: which of the two you should download, and what you give up by picking the fast one. The capabilities overview covers what the model does; this covers which file does it.

What is model distillation?

Model distillation compresses the sampling process, not just the weights.

A diffusion model denoises from noise over a schedule of steps, and each step is a full forward pass through the transformer. Distillation trains a variant that lands on a comparable result using a short, fixed schedule, so the expensive part of generation runs a handful of times rather than dozens.

The technique is called knowledge distillation in the literature and appears across modalities. What makes it matter for video is arithmetic: a video run denoises a patchified sequence covering every frame at once.

Why the step count dominates

Step count dominates cost because every step pays for the whole sequence.

The transformer documented in ltx-core is asymmetric and dual-stream, processing video and audio tokens together across 48 shared blocks, and the denoising loop runs that full stack on every step.

Cutting a 40-step schedule to 8 is therefore not a 5x saving on a small number. It is a 5x saving on the most expensive operation in the run.

That is why the repository names the distilled path as its quick-start route, and why it documents gradient estimation as a way to bring a guided run from 40 inference steps down to 20 to 30 while maintaining quality.

The two transformers

LTX-2.5 ships its weights as one file per component, so you download only the parts your pipeline needs. Two of those files are transformers, and they are not interchangeable. The release notes track what changed at launch.

ltx-2.5-22b-dev-transformer-bf16.safetensors is the full, trainable model. It is what the guided pipelines run in stage 1, where classifier-free guidance and spatio-temporal guidance are available.

ltx-2.5-22b-distilled-transformer-bf16.safetensors is the few-step variant, running a fixed 8-step schedule with CFG at 1.0.

Both sit alongside the same required text encoder, gemma4-12b-with-proj-ltx-2.5-bf16.safetensors, a Gemma 4 12B encoder fine-tuned for LTX. Google's stock Gemma 4 release is not a substitute, and the encoder is what carries LTX-2.5's prompt adherence.

Plan disk accordingly. A 22B transformer in bf16 and a 12B encoder in bf16 account for most of it before the VAEs and upscalers are counted, so the quick-start set runs to roughly 66 GiB.

How many steps does each use?

The pipelines reference states the distilled schedule directly: two-stage generation with 8 predefined sigmas, 8 steps in stage 1 and 4 steps in stage 2, with no guidance required.

Read that as a fixed schedule rather than a contradiction. The 8 sigmas define the stage-1 denoise; stage 2 runs a shorter refinement pass on top.

For guided runs the reference point is the optimization documentation, which describes reducing inference steps from 40 to 20 or 30 using gradient estimation. That places the ungraded baseline at 40.

TI2VidTwoStagesHQPipeline sits between them. It runs the same two-stage structure with the res_2s second-order sampler instead of Euler, documented as typically allowing fewer steps for comparable quality.

Which pipelines take which

Pipeline compatibility should drive your download, because a mismatched checkpoint fails at load rather than degrading gracefully.

Distilled only. DistilledPipeline, ICLoraPipeline, DubItPipeline, HDRICLoraPipeline and DFRPipeline all require the distilled checkpoint. The IC-LoRA family is the reason: those adapters were trained against it, and the documentation states plainly that ICLoraPipeline can only be used with a distilled model. That covers most video-to-video work, HDR and EXR output, and dubbing, which sits inside LTX-2.5's joint audio and video generation.

Full model. TI2VidTwoStagesPipeline and TI2VidTwoStagesHQPipeline for text-to-video, KeyframeInterpolationPipeline for image-to-video interpolation, A2VidPipelineTwoStage for audio-to-video, plus TI2VidOneStagePipeline and T2AOneStagePipeline. These run the full model in stage 1, which is what makes CFG and STG available on those paths.

Either. RetakePipeline is the flexible one. It can run the full model with CFG guidance or the distilled model with a fixed sigma schedule, depending on what you have loaded, which makes it the entry point for targeted video editing.

One caveat on TI2VidOneStagePipeline: the documentation labels it educational and routes production work to the two-stage pipelines, so treat it as a reference implementation rather than a speed option.

Distillation is also an adapter

This is the detail most likely to cost you a failed run, and the source of a common misconception.

The guided two-stage pipelines do not escape distillation by running the dev transformer. TI2VidTwoStagesPipeline upsamples to 2x resolution in stage 2 with distilled LoRA refinement, TI2VidTwoStagesHQPipeline does the same, and A2VidPipelineTwoStage refines the video in stage 2 using a distilled LoRA while holding the audio fixed.

That adapter is ltx-2.5-22b-distilled-lora-450-bf16.safetensors, published in the LTX-2.5 repository and labelled for dev-transformer workflows.

Note where it applies. It is a stage-2 refinement adapter, not something fused into stage 1. Stage 1 is where the full model and the guidance parameters live; stage 2 is where distillation comes back in to make the upsample cheap.

So distillation in LTX-2.5 is two things: a standalone checkpoint, and an adapter the guided pipelines fuse onto the dev transformer for their second stage. Downloading the dev transformer alone does not give you a working two-stage run. The LoRA adapters guide covers how adapters load, and LoRA training on the LTX Trainer covers producing your own.

Distilled vs dev, side by side

Four axes separate them.

Step schedule. Distilled runs a fixed 8 sigmas with no guidance. Dev baselines at 40 steps, dropping to 20 or 30 with gradient estimation, or fewer again on the res_2s sampler.

Pipeline access. Distilled unlocks the IC-LoRA family, Dub-It, HDR video-to-video and DFR. Dev unlocks the guided two-stage paths, keyframe interpolation and audio-to-video, along with the CFG and STG parameters. If you need deterministic structural control rather than either, motion and camera control runs through IC-LoRA and therefore through the distilled checkpoint.

Extra artifacts. Distilled runs need the spatial upscaler. Dev two-stage runs need the spatial upscaler and the distilled LoRA. Either way, ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors is required for any multi-stage pipeline.

Memory envelope. Both are 22B transformers, so the checkpoint footprint is comparable. The practical difference is the step count: fewer forward passes means less time on a constrained card, which is why the distilled path is where low-VRAM configurations start.

Does distilled mean lower quality?

Not in the way most people assume.

The documentation frames the choice as speed against capability. DistilledPipeline is described as fastest inference with good quality, and some of what you gain on the dev path is fidelity while some is simply access to pipelines that only run there.

Production quality is not automatically a dev question. DFRPipeline, the Diffusion Fidelity Rendering path with a detailing pass and optional temporal 2x and 4x refinement, is documented as the production-quality route and runs on the same distilled transformer. It takes --detailing-lora rather than --distilled-lora, and that detailing adapter is the same one behind LTX-2.5's generative upscaling.

LTX-2.5 also ships a substantially improved distilled model in its launch capability set. Treating "distilled" as shorthand for "preview quality" is the most common misreading of the trade-off.

Fast, Pro, dev, distilled

Fast and Pro are the model tier names on the hosted LTX API, where requests specify "model": "ltx-2-5-fast" or "model": "ltx-2-5-pro". Dev and distilled are the open-source variant names.

The older ltx-2-fast and ltx-2-pro IDs were removed in August 2026 and now return an error, so any guide still showing them is out of date.

These are two naming systems describing a similar split between a faster tier and a higher-fidelity tier, and they do not share parameter schemas. A step count, sampler choice or quantization policy from the open-source pipelines has no equivalent field in a hosted request, and per-second SKU pricing has no equivalent in a local run.

Check which surface a guide is written for before copying settings across.

How should you choose?

Choose on the pipeline you need first, the hardware you have second, and quality expectations third. The first two are hard constraints and the third is adjustable.

Start with the pipeline. If your workflow needs ICLoraPipeline for video-to-video or DubItPipeline for dubbing, the distilled transformer is not a preference. It is the required checkpoint.

Check the memory envelope. The open-source repository targets NVIDIA GPUs with 80GB or more of VRAM, tested against CUDA 12.7 or newer, with LTX's LoRA guide recommending 13 or newer. Distilled variants are supported on 32GB with quantization, where --quantization fp8-cast --offload cpu is the documented starting point. Block streaming exists for machines that cannot hold all the weights at once.

Then decide on fidelity. Run distilled to lock composition and motion, and move to a guided two-stage run once the shot is approved. Iterating on 8 steps and rendering finals on 20 to 40 is the allocation the documented step counts point to. Prompt construction affects this more than step count does past a point, so the LTX-2.5 prompt guide is worth reading before you start tuning schedules.

If none of that fits your hardware, the hosted API removes the VRAM question entirely, and the LTX Playground runs the model with nothing installed.

Common mistakes

Three artifact-pairing errors account for most failed runs.

Forgetting the distilled LoRA on a dev run. The two-stage pipelines need ltx-2.5-22b-distilled-lora-450-bf16.safetensors for stage 2. The dev transformer alone is not enough.

Skipping the spatial upscaler. Any multi-stage pipeline requires ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors. Without it there is no second stage to run.

Assuming adapters do not carry across versions. This one is worth correcting because the opposite advice is common. LTX's model card states that 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. Test yours before production rather than retraining on the assumption it will fail. The community library is worth checking before you train anything new.

One more: expecting dev step counts to help a distilled run. The distilled schedule is fixed at 8 sigmas, not a starting point to tune upward.

Summary

Distillation trains a faster student model to reproduce a teacher's output on a short, fixed schedule. LTX-2.5 publishes it two ways: as ltx-2.5-22b-distilled-transformer-bf16.safetensors, and as ltx-2.5-22b-distilled-lora-450-bf16.safetensors, which the guided pipelines fuse into stage 2 on top of the dev transformer.

The distilled path runs 8 predefined sigmas, 8 steps in stage 1 and 4 in stage 2. A guided run baselines at 40 and drops to 20 or 30 with gradient estimation.

Pipeline compatibility decides your download. DistilledPipeline, DFRPipeline, ICLoraPipeline, DubItPipeline and HDRICLoraPipeline require the distilled checkpoint. The guided two-stage, keyframe, audio-to-video and text-to-audio pipelines run the full model. RetakePipeline takes either.

Fast and Pro are the hosted tier names for a similar split, currently ltx-2-5-fast and ltx-2-5-pro, and their parameters are not interchangeable with the open-source ones.

Both transformers are published under the LTX-2.x Community License and available from the open-source model page. Check the LTX-2.5 model page for the current capability set, and the developer program if you are building on this at scale.