by.waclaw.online / spark / video / 01

What a Generated Clip Actually Costs Here

The lead article: which open video models are real, what they weigh, how long they take on this hardware, and which pipeline matches which of your goals.

Series 3, article 1 of 4 · written 31 July 2026

Set the expectation first, because everything else follows from it: a few seconds of video takes a few minutes to generate on this box. A short 720p clip from a well-configured LTX-2 workflow lands at just over three minutes per iteration once the model is loaded. Heavier workflows — higher resolution, more frames, a big Wan configuration — run 15 to 25 minutes. There is no setting that turns this into a real-time tool.

What you get in exchange is that the models fit at all, and fit without the offloading contortions a 24 GB card needs. A 22-billion-parameter video transformer plus its text encoders plus an audio decoder is an awkward fit on consumer hardware and an unremarkable one here. If your working style is "queue five variations and go and do something else," this machine suits it well. If it is "nudge a slider and look again," it will frustrate you.

The models that matter

LTX-2.3 — the one to start with

Lightricks open-sourced LTX-2 on 6 January 2026, and the 2.3 release followed in March at 22B parameters. Two checkpoints are published: a development model and a distilled variant that runs on a short fixed schedule of denoising steps and is dramatically faster.

Its distinguishing feature is that it is an audio-video foundation model: it generates synchronised sound and picture in a single diffusion pass rather than dubbing sound onto finished frames. That is not a convenience — it means ambient audio, footsteps and speech are generated with knowledge of the image, and it is what makes the LipDub pipeline possible. It ships nine pipelines including text-to-video, image-to-video, audio-to-video, video-to-video, keyframe interpolation and lip dubbing, with a maintained ComfyUI integration.

It is also pre-bundled in the GB10-specific ComfyUI image the next article uses, at around 42 GB for the transformer. For a machine with your language requirements and your goals, this is the default choice.

Wan 2.2 — when the humans have to be photoreal

Alibaba's Wan 2.2 remains the reference point for photorealistic human subjects: facial detail, skin texture and hair hold up better than the alternatives. The open family is:

VariantSizeWhat it does
TI2V-5B5BText and image to video, 720p at 24 fps. The efficient one — around 9 minutes for 5 seconds on a 4090-class card.
T2V-A14B / I2V-A14B27B MoEThe quality tier. Designed for 80 GB-class accelerators; fits here, slowly.
S2V-14B14BSpeech-to-video — audio-driven cinematic generation. Directly relevant to article 4.
Animate-14B14BCharacter animation and replacement.

The important caveat: Wan 2.2 has no native audio. Sound is a separate pass. And the later versions are not what the marketing implies — Wan 2.5 and 2.6 shipped through Alibaba Cloud APIs without public weights, which is a hard stop for an on-box setup. Claims circulating about a fully open Wan 2.7 should be checked directly against the publisher's own repositories before you plan around them; a good deal of the content asserting it is machine-generated and the official GitHub organisation does not currently carry such a repository.

The general lesson about video model news. Video generation attracts more search-optimised invention than any other corner of this field, because the audience is large and the release cadence is fast. Model names, parameter counts and "open weights" claims are all routinely fabricated. Before downloading anything, check the publisher's own GitHub organisation or model hub. If a model's weights are not there, they are not open, whatever a blog says.

HunyuanVideo 1.5 — the physics option

Tencent's model is the one people reach for when motion has to be physically convincing: fluids, smoke, cloth, object interaction. Version 1.5 slimmed to 8.3B and runs in roughly 14 GB with offloading, making it by far the lightest of the three. Also no native audio.

Matching pipeline to purpose

You named three goals. Each has a clear best answer, and they are not the same answer.

GoalPipelineWhy
Short cinematic B-roll LTX-2.3 text-to-video, distilled weights Audio in the same pass means atmosphere without a second stage. The distilled checkpoint is where the wall-clock time becomes bearable for iteration. Article 3.
Animating stills LTX-2.3 image-to-video Cheapest per second of anything here — you supply the hardest part of the image, so the model spends its budget on motion. Article 3.
Talking-head presenter LTX-2.3 LipDub, or InfiniteTalk for long-form Both are driven by an audio track, which you already produce in series 2. LipDub re-generates mouth movement on existing footage; InfiniteTalk animates a still portrait for arbitrarily long speech. Article 4.
(Photoreal humans, no speech) Wan 2.2 I2V, audio added separately Worth keeping installed if faces are the point and the extra minutes are acceptable.

The memory picture

Video models are not one file. A working LTX-2.3 setup means the transformer, one or more text encoders, a VAE and an audio decoder resident at once. Indicative sizes from the GB10 ComfyUI bundle:

ComponentSize
LTX-2.3 22B transformer~42 GB
Flux 2 Dev image model (for stills to animate)~35.5 GB
ACE-Step v1.5 XL Turbo (music, bundled)~10 GB
Text encoders, VAEs, supporting weightstens of GB across the set
Whole bundle on disk~285 GB

Two consequences. First, this is why the hub page insists on a plan for model storage — 285 GB for one application is most of a 1 TB drive before you have downloaded a single LLM. Second, the resident set during generation is large enough that ComfyUI's memory behaviour on unified memory stops being a detail and becomes the thing that decides whether your workflow completes. That is the entire subject of the next article.

Timings you can plan around

Order-of-magnitude figures for this hardware. Your numbers will vary with resolution, frame count, step count and which checkpoint you use — measure your own and keep a note.

JobRough wall clock
Short 720p clip, LTX-2 distilled, warm~3 minutes per iteration
Same, first run of the sessionSubstantially longer — CUDA kernel compilation alone accounts for a reported ~3× penalty
Heavier Wan 2.2 workflow15–25 minutes
Model load from external storageMinutes, per checkpoint change
Lip-sync pass over existing footageScales with duration; a two-stage refinement roughly doubles it

The practical implication is a working method rather than a setting: iterate at low resolution and short duration, then re-run the keeper at full quality. A 480p, two-second draft tells you whether the composition, motion and prompt are right, and costs a fraction of the final render. Every minute spent iterating at final quality is a minute wasted on a clip you are going to discard.

Warm-up is real and it is not a bug. The first generation after starting ComfyUI compiles kernels and populates caches. Reported speedups of about 3× on subsequent runs are normal. Do not benchmark your setup on run one, and do not conclude the box is broken because the first clip took quarter of an hour.

Before you publish anything with a face in it

The same regulation flagged in the audio series applies here with more force. From 2 August 2026, Article 50 of the EU AI Act requires AI-generated or manipulated video to be marked machine-readably and disclosed, with a reinforced obligation for deepfakes — content resembling a real person that would plausibly be taken as authentic. Intent to deceive is not required for the obligation to attach. A talking-head video of a real person, generated from a cloned voice, is the central case the rule was written for.

For private experimentation on your own machine this is not a concern. The moment a clip is published or sent to someone, it is. Build the labelling into the pipeline rather than remembering it later — article 4 shows where it fits.

Sources