LTX-Video in ComfyUI: Near Real-Time AI Video Generation on a Single GPU
Most open video models make you wait. LTX-Video (LTXV), from Lightricks, takes the opposite stance: it is a DiT-based (Diffusion Transformer) video model engineered so aggressively for speed that short clips can render in seconds on a single consumer GPU, fast enough that iterating on a shot feels closer to image generation than to traditional video rendering.
If Wan 2.2 is the quality-first choice, LTX-Video is the iteration-first choice. Here is how to run it in ComfyUI and get clean results.
Visual guide: Make a video experiment reproducible
How to use this map
- Identify the exact model release and template. A setting from another LTX family or a distilled variant should not silently become the default for this workflow.
- Keep target frames, dimensions and playback FPS in the experiment record. Playback duration and generation runtime are different measurements.
- Describe one action and one camera movement for the first test. This makes it easier to identify subject drift or motion failure than a long multi-shot prompt.
- Inspect several frames and the encoded video, not just its opening still. Record temporal artifacts, runtime and memory before making a claim about practical throughput.
Stop and check: Near real-time is not established by a model name or a cover image. Report the hardware and complete timing boundary for any speed result.
Next practical guides: Wan 2.2 Image-to-Video in ComfyUI: Setup and Motion Checks · Wan 2.2 Lightning in ComfyUI: Dual-Expert Sampling and Fixes
Implementation reference: Official LTX-Video project. Use the matching version-specific example alongside this editorial map.
Why Speed Changes the Workflow
When a clip takes two minutes, you plan carefully and accept the first non-broken result. When it takes ten seconds, you explore — a dozen seeds, several prompt phrasings, different start frames. LTX-Video's architecture is built around that loop:
It is a single transformer model (no dual-expert split), so the graph is simple.
It supports text-to-video and image-to-video in the same family.
It runs in far less VRAM than the 14B video models, which is why it shows up in so many "AI video on a laptop GPU" tutorials.
The tradeoff is honest: at its fastest settings LTXV trades some motion coherence and fine detail for throughput. Used well — especially for image-to-video where a strong start frame anchors the result — it is remarkably good.
Installation
Update ComfyUI, then install ComfyUI-LTXVideo via ComfyUI Manager (search "LTXVideo") or by cloning it into ComfyUI/custom_nodes/.
Download the LTX-Video checkpoint into ComfyUI/models/checkpoints/ (LTXV ships as a packaged checkpoint; newer distilled releases are noticeably faster — grab the latest 0.9.x).
Text encoder: LTXV uses a T5-XXL encoder. If you already run FLUX you likely have it in ComfyUI/models/text_encoders/.
Load the Workflow > Browse Templates > LTX-Video example, or import a template JSON from the ComfyUI-LTXVideo repo.
Building the Graph
Text-to-Video
Load Checkpoint (LTXV) — supplies model, CLIP/T5, and VAE.
CLIP Text Encode — positive and negative prompts.
EmptyLTXVLatentVideo (or the template's latent node) — set width, height, and frame count. LTXV is happiest at its trained resolutions; start around 768x512 and 97 frames.
LTXV sampler nodes — the template wires a scheduler and sampler tuned for the model. Keep steps modest (the distilled variants target very low step counts).
VAE Decode > Video Combine / Save Video, fps around 24–30.
Image-to-Video
Swap the empty latent for an LTXV image-to-video conditioning node and feed it your Load Image. This is the mode most people should start with — the start frame carries composition and identity, so the model only has to invent motion. Describe only what moves in the prompt.
Prompting LTX-Video
LTXV, like other video models, wants motion described in prose, and it is particularly sensitive to prompt detail — thin prompts tend to produce static or jittery clips.
Lead with the subject and a clear action, then the camera move, then environment.
Be explicit about camera behavior: "slow push-in," "orbit left," "locked-off tripod shot."
Add short atmosphere cues ("dust motes drifting," "gentle wind in the grass") — LTXV uses them to generate believable secondary motion.
Keep the negative prompt populated with the usual video offenders: "static, blurry, distorted, warping, jitter."
Optimization: Speed, VRAM, Quality
Frame count is your biggest cost lever. Prototype at ~65 frames, then extend once the shot works.
Resolution: stay near the model's native training sizes; overshooting causes warping and costs disproportionate VRAM. Upscale afterward instead.
STG / guidance: LTXV workflows often expose Spatiotemporal Guidance controls — nudge these up if motion looks unstable, down if the clip looks over-processed.
VRAM: LTXV runs on 8–12 GB comfortably at moderate settings; use FP8/GGUF weights and --lowvram on the smallest cards.
Finishing: the highest-leverage post step is frame interpolation (RIFE) to smooth motion, followed by a light upscale. Because LTXV is so cheap to sample, many creators generate several candidates and only upscale the winner.
A Developer Note on Throughput
Because a clip is so fast, LTXV pairs naturally with batch automation: script a list of prompts, queue them through the ComfyUI API overnight, and review a contact sheet in the morning. The economics are completely different from heavyweight video models — you can afford to overshoot and cull.
Conclusion
LTX-Video is the model to reach for when your bottleneck is iteration, not final polish: storyboards, animatics, social clips, and rapid concept exploration. Start in image-to-video mode with a strong first frame, describe the motion clearly, keep frame counts low while you dial in the shot, and lean on interpolation and upscaling to finish. It will not replace a quality-first model for hero shots — but for everything upstream of the hero shot, it is hard to beat.
Downloadable Workflow & References
|
