LTX 2.3 has a lot of dials — sampler, scheduler, steps, sigmas, CFG, motion amplitude, image compression, guide strength. Get them right and the model produces sharp, smoothly-animated output. Get them wrong and you get blurry frames, jittery motion, or completely fried generations. This guide covers the settings that matter, with the values we use across our production workflows.

Every value below is what we run in VirtuaVixen Studio for our LTX 2.3 Cinema workflows. If you'd rather not tune anything, the Studio runs them at these defaults free with daily tokens. Or grab the ComfyUI Workflow Pack to get the same configs as JSON — every node pre-tuned, with Discord support if you want to tweak.

Best Sampler for LTX 2.3

Use euler_ancestral_cfg_pp. It's the sampler the LTX team designed for the model and what every production workflow ships with. Other samplers technically work but produce inferior results:

• euler_ancestral_cfg_pp — recommended. Smooth motion, good prompt adherence, no tearing.
• euler_ancestral — works but loses some sharpness compared to the cfg_pp variant.
• dpmpp_2m — sometimes used in non-distilled workflows. Slower, marginal gain.
• lcm / lcm_custom_noise — only for the LCM-distilled checkpoint, not the standard 22B.
• uni_pc / uni_pc_bh2 — produces visible artifacts on LTX 2.3.

Best Scheduler

Two options that both work well, picked by workflow type:

• linear_quadratic with BasicScheduler — best for FML keyframe and dialogue-heavy workflows. Smooth ramp, good detail at the end.
• LTXVScheduler with stretch=true, terminal=0.1 — best for sex-cut Cinema and continuous-motion workflows. The terminal value controls how the schedule lands at the final timestep.

Avoid karras, exponential, and simple — they produce visibly degraded LTX 2.3 output.

Steps

For the distilled 22B model (what most people use):

• 8 steps — fast, slight motion smoothness sacrificed. Good for rough drafts.
• 9 steps — sweet spot. Our production default for FML and Cinema workflows.
• 12 steps — sharper detail, costs another ~30% of generation time. Use for hero shots.
• 16+ steps — diminishing returns. Don't bother.

For the dev (non-distilled) model: 25–35 steps. Different model, completely different schedule. See Dev vs Distilled.

CFG (Prompt Strength)

• Distilled model: CFG = 1.0. Hard rule. Higher values produce burned, oversaturated output.
• Dev model: CFG = 4–7. Standard SD-style behavior.

The Studio's “Cinema” advanced panel exposes CFG from 0.5–6 — but for 99% of users on the distilled model, leave it at 1.0.

Best Resolution for LTX 2.3

• Longest side 1080 — production default. Generation runs at this size, then a 2× spatial upscaler doubles it post-VAE-decode. Final output 2160 longest side.
• Longest side 768 — for low-VRAM setups. Output is softer but generation is much faster and fits on 16 GB cards.
• Longest side 832 — what WAN 2.2 typically runs at. LTX accepts it but you're not using the model's strengths.

LTX 2.3 was trained on 1080-class data. Below 768, output quality drops noticeably.

Why Your Output Is Blurry

• Steps too low (under 8) — increase to 9 or 12.
• Wrong sampler — switch to euler_ancestral_cfg_pp.
• Source image too low-res — your input image gets resized to match LTX's working resolution. If your source is 512×512, the result is upscaled blur.
• Image compression too high — the LTXVPreprocess node has an img_compression input. Default 18 is fine; values above 30 introduce visible JPEG artifacts.
• VAE Decode tiled overlap too low — for 1080+ output, set spatial_overlap=8 and temporal_overlap=4 on VAEDecodeTiled. Lower values cause tile seams.

Manual Sigmas

Some users on Reddit have shared “manual sigmas” — handcrafted denoising schedules that override the scheduler. These exist mainly to squeeze marginal quality gains out of fewer steps. For most users it's not worth the trouble; the linear_quadratic scheduler at 9 steps already produces good output. If you want to experiment, the Discord has shared sigma sets for specific use cases.

Production Defaults Summary

Skip the Tuning

If you don't want to tune anything, our Studio runs every LTX 2.3 workflow at these production-validated settings. The Cinema panel exposes only the dials worth tweaking (steps, CFG, negative prompt). For local power users, the ComfyUI Workflow Pack ships every workflow with these defaults already configured — paste, generate, modify only if you want to.

Related Reading

• LTX 2.3 Prompting Guide: Tips, Examples & Negative Prompts
• LTX 2.3 Quants Compared: GGUF, FP8, FP4, NVFP4
• LTX 2.3 Dev vs Distilled: Which 22B Model Should You Run?
• LTX 2.3 ComfyUI Workflow Guide
• LTX 2.3 GPU and VRAM Requirements