LTX 2.5 Distilled vs Dev: The Two Checkpoints, Picked for You

LTX-2.5 ships two transformers, and the choice between them is not "fast vs pretty" — it is which pipelines and workflows you are running. The distilled checkpoint runs a fixed 8-step schedule with no guidance; the dev checkpoint is the full trainable model that the two-stage pipelines use for quality. Pick the wrong one for a pipeline and it errors out or silently runs the wrong configuration.

The official model card states the split plainly: the distilled DiT uses a "Fixed 8-step schedule, CFG=1," and the dev DiT is the "Full / trainable DiT." The GitHub README maps them to pipelines: DistilledPipeline, ICLoraPipeline, and DubItPipeline expect the distilled model; the guided two-stage pipelines run the dev model (with the distilled LoRA applied for the stage-2 refinement). Checked August 13, 2026.

The two checkpoints

The headline detail: the distilled checkpoint is not "a faster version of the dev model." It is a fixed-schedule model — 8 steps, CFG=1 — meaning it is a different inference contract. The 8-step schedule is what makes it fast, and it is also why some control pipelines (IC-LoRA, DubIt) are documented as distilled-only: the fixed schedule is part of how they work.

Which pipelines require which

The README's pipeline list is explicit about expectations:

• DistilledPipeline — "Fastest inference with 8 predefined sigmas (recommended)" — uses the distilled transformer
• ICLoraPipeline — video-to-video and image-to-video transformations — distilled model
• DubItPipeline — dubbing with speaker identity and lip matching — distilled model
• DFRPipeline — detail-fidelity rendering (keyframes + spatial detailing, optional temporal 2x/4x) — dev-based
• TI2VidTwoStages / TI2VidTwoStagesHQ — production two-stage with 2x upsampling — dev-based, and these are the pipelines that require the distilled LoRA (ltx-2.5-22b-distilled-lora-450) for the stage-2 refinement pass
• TI2VidOneStage — single-stage prototyping — dev-based
• KeyframeInterpolation, A2Vid, Retake, HDRICLora — specialized pipelines

The practical pattern: single-pass generation and reference-based control run distilled; multi-stage quality pipelines run dev + distilled LoRA. That is the sentence most guides omit, and it is the one that prevents config errors.

What "8 steps, CFG=1" means for your iteration loop

The distilled model's fixed schedule changes the economics of iteration:

• 8 denoising steps in stage 1, 4 in stage 2 (per the README's optimization notes on DistilledPipeline) — with no guidance required
• No CFG to tune, no guidance scale to search — one less knob, and the knob that usually costs the most wall-clock time
• The release's speed claims ("better results, less compute"; 6.8 s for a 10-second clip on 2× GB200, per the official page) are built around the distilled path

The vendor's own framing for the release notes the improved distilled model as a headline item: it retains much more of the full model's visual quality, prompt adherence, and motion consistency than previous distilled builds. The quality gap between distilled and dev has narrowed — which is precisely the "distilled vs dev" decision many users no longer need to make: if your pipeline accepts distilled, distilled is the recommended default.

Picking in ComfyUI

The three official ComfyUI workflows (T2V, I2V, FLF2V) all use the distilled int8-convrot build — ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors in models/diffusion_models/. If you load the dev int8 build into those workflows expecting an upgrade, you get a pipeline mismatch, not a quality boost: the workflows are wired for the distilled schedule.

The dev int8 build (ltx-2.5-22b-dev-transformer-comfy-int8-convrot.safetensors) exists for dev-based workflows that want the int8 memory profile — but as of the official docs, the native templates are distilled. If you want dev-based two-stage quality in ComfyUI, you are assembling that workflow yourself.

Choosing for your use case

Two more notes for the decision:

• Download size is a real factor. The distilled and dev transformers are both 22B bf16 files — roughly 44 GB each. The official Quick Start pulls only the distilled one (~66 GiB total for the five-file set). If you do not need two-stage quality, you do not need the dev file at all.
• int8 variants exist for both, ComfyUI-only. The model card labels both comfy-int8-convrot builds "ComfyUI only — not for ltx-pipelines / PyTorch." The NVFP4 distilled build is the Blackwell path (ComfyUI, or ltx-pipelines with --quantization nvfp4-prequant).

Decide on the model before the checkpoints. Generate with LTX 2.5 in your browser at ltx23.app — the distilled path, quality included — and check whether you even need the dev file's two-stage margin before downloading 44 GB more.

FAQ

Is the distilled LTX 2.5 checkpoint better than the dev one? Not better — different contract. Distilled is fixed 8-step, CFG=1, fast, and required by DistilledPipeline/ICLoraPipeline/DubItPipeline. Dev is the full trainable model used by the two-stage pipelines (with the distilled LoRA). Quality has converged per the vendor's release notes, but the pipeline determines which one you load.

Can I use the dev checkpoint in the official ComfyUI workflows? The official T2V/I2V/FLF2V workflows are wired for the distilled int8-convrot build. Loading dev into them is a pipeline mismatch, not an upgrade.

Does the distilled checkpoint support fine-tuning? No — dev is the trainable transformer. For deeper fine-tuning foundations, Lightricks also publishes the Pre-Trained checkpoint.

Why do two-stage pipelines need a distilled LoRA? The dev-based two-stage flows (TI2VidTwoStages, etc.) run the full model in stage 1 and apply the distilled LoRA (ltx-2.5-22b-distilled-lora-450) for the stage-2 refinement pass. It is infrastructure, not optional flavor.

Which file should I download first? The distilled bf16 transformer — it runs the Quick Start and the official workflows. Add the dev file only if you need the two-stage pipelines.

Bottom line

LTX 2.5's distilled vs dev decision is pipeline-driven: 8-step fixed-schedule distilled for single-pass generation, control, and the official ComfyUI workflows; full dev (plus the distilled LoRA) for two-stage quality and fine-tuning. Start with distilled — it is the recommended default and the only file the Quick Start needs — and reach for dev when a pipeline or your training workflow demands it. Try the distilled quality in your browser at ltx23.app before committing disk space to both 44 GB files.

Sources

Verified August 13, 2026 against primary documentation:

1. Lightricks/LTX-2.5 — Hugging Face model card — distilled (fixed 8-step, CFG=1) vs dev (full/trainable) descriptions, int8/NVFP4 variants, improved distilled model claim
2. Lightricks/LTX-2 — GitHub README — pipeline-to-checkpoint mapping, distilled LoRA requirement, 8+4 step counts, Quick Start file set (~66 GiB)
3. ComfyUI official LTX-2.5 tutorial — official workflows using the distilled int8-convrot build, model placement
4. LTX-2.5 official model page (ltx.io) — speed measurement, "better results, less compute" framing
5. Lightricks/LTX-2.5-Pre-Trained — Hugging Face — pre-training checkpoint for fine-tuning foundations

Speed figures are vendor-published and measured on datacenter GPUs; verify quality and speed on your own hardware.