LTX 2.3 Dev vs Distilled: Which 22B Model Should You Run?

virtuavixen No Comments

LTX 2.3 ships in two main flavours: 22B Distilled and 22B Dev. They share the same parameter count and the same VAEs, but they're trained differently and behave very differently at inference. Choosing the wrong one is the most common reason new LTX 2.3 users get garbage output. This guide explains exactly when to use which.

If you'd rather not pick, our VirtuaVixen Studio uses the distilled model by default — it's faster and what most production NSFW workflows ship with. Available in your browser with daily free tokens. The ComfyUI Workflow Pack defaults to distilled too, with optional Dev workflow templates for advanced users.

Quick Verdict

  • Use Distilled if: you want fast inference (9 steps), you're using ComfyUI's standard workflows, you don't want to mess with CFG values.
  • Use Dev if: you need fine-grained control via CFG, you're building experimental pipelines, you don't mind 3–4× longer generation times.

Side-by-Side

Property22B Distilled22B Dev
Sampling steps8–12 (9 typical)25–35 (30 typical)
CFG (prompt strength)1.0 (fixed)4–7 (tunable)
Generation time (12s clip, 4090)~8 minutes~30 minutes
Quality at default settingsExcellentSlightly better detail
Prompt-followingGoodStronger (CFG-driven)
Distill LoRA neededYes (built into model)No
NSFW LoRA compatibilityExcellentExcellent
Best forProduction, batch generationExperimental, tightly-controlled output

What Distilled Actually Means

Distillation is a training technique where a “student” model learns to approximate the output of a “teacher” model in fewer steps. The LTX 2.3 distilled checkpoint was trained to produce in 9 steps what the dev model produces in 30. The trade-off: the distilled model loses some flexibility — it works best at CFG=1 and a narrow range of step counts, but it's much faster.

For 95% of users, distilled is the right default. It's the model that ships in our Studio and what every NSFW workflow we publish targets.

When to Use Dev

The Dev model is worth the extra time when:

  • You need very tight prompt adherence — Dev's tunable CFG lets you push prompt strength up to 7 for hard-to-render concepts. The distilled model can't do this.
  • You're building research pipelines or training LoRAs — Dev is the standard for fine-tuning. See how to train an LTX 2.3 LoRA.
  • You're chasing maximum quality at the cost of speed — Dev at 35 steps produces marginally sharper detail than distilled at 9.
  • You need negative-prompt strength control — only meaningful at higher CFG, which distilled doesn't support.

The Distilled LoRA Trick

You'll see a separate file in the LTX 2.3 release: ltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16.safetensors. This is the distillation LoRA — it converts a Dev checkpoint into something that behaves like Distilled at low step counts.

If you've already downloaded the Dev checkpoint and don't want to re-download Distilled, load Dev + the distilled LoRA at strength 0.6. Runs at 9 steps with CFG=1, same as native Distilled. We use this in some of our older workflows for backwards compatibility.

Common Failure Modes

  • “My distilled output is burned/oversaturated” — you set CFG above 1.0. Drop to 1.0.
  • “My dev output is washed out” — CFG too low. Try 4–6.
  • “I can't tell which model I'm using” — check the filename of the checkpoint loaded in your UNETLoader. Distilled has “distilled” in the name; Dev usually says “dev” or just the model number.
  • “Distilled model takes 30 minutes per clip” — you set steps too high. Distilled wants 9, not 30.
  • “Dev model produces blurry output at 9 steps” — Dev needs 25–35 steps. Step count is the wrong dimension to optimise here.

Which One Should You Download First?

Distilled. Almost every user gets more value from it: faster iteration, better defaults, broader workflow compatibility. Once you're comfortable with the model, download Dev if you specifically need its capabilities.

Try Both in the Studio

Our Studio defaults to the distilled model in every LTX 2.3 workflow — no decisions, no setup. If you want to compare Dev vs Distilled side-by-side without downloading both, the Studio is the fastest way to test.

For local control, the Workflow Pack ships the distilled checkpoint as the default + optional Dev workflow JSONs that are pre-configured with the right step count and CFG values. Saves you the trial-and-error of dialing in either model.

Related Reading

Leave a comment

Are you 18 or older?

You must be 18 years or older to access this website.

👑 AI Studio ×

Categories