Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

Author Makes a Narrative Video on Home LTX 2.3 (22B Parameters) Without Subscriptions

Paid cloud video generators are pushing creators to look for alternatives, and author Pavel Vlasov tested a home setup. LTX 2.3 is a local 22-billion-parameter video model (a distilled, accelerated version) that runs on a home computer. He used it to create a story about a girl and a dragon, though one scene had to come from the cloud-based MiniMax H3. Local generation can handle a project like this, but it costs time and a lot of fiddling with prompts.

AI-processed from TG: @cgevent (Метаверсище и ИИще); edited by Hamidun News
Author Makes a Narrative Video on Home LTX 2.3 (22B Parameters) Without Subscriptions
Source: TG: @cgevent (Метаверсище и ИИще). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @cgevent (Метаверсище и ИИще) ↗
Editor-in-Chief
Key metric #1LTX failed to render the scene of the dragon flying away, so that episode was made in MiniMax H3
Prompts for local models are fussierthe author rewrites them through DeepSeek
Pros of localno censorship, no subscriptions, no internet needed; the downside is time and flexibility

Cloud video generators keep moving to paid tiers one after another, and the question of whether you can do without them has stopped being theoretical. Pavel Vlasov, author of a project in the "Neuro-Burn" column of the "Metaverse and AI" channel, tested this in practice. He made a short atmospheric story about a girl and a dragon entirely on a home computer, using the local video model LTX 2.3 with 22 billion parameters. This means the model runs on his own hardware rather than on someone else's servers, and he doesn't pay for each second of video.

What the video is made from

The foundation is LTX 2.3 in its distilled version (Distilled 1.1). Distillation makes the model lighter and faster, usually at the cost of a small loss in quality. On top of it, the author added a style LoRA called Luxe_Sensual_Style. A LoRA is a small add-on that shifts the look of the image without retraining the entire model. It is activated with a special trigger word. According to the developers' description, this add-on pushes a "glossy" look: soft cinematic lighting, natural skin, and the texture of lace and silk in the spirit of fashion photography. Vlasov honestly admits that it is hard to assess the LoRA's contribution precisely: "something changed," but he does not claim to know how much.

The whole video did not come out of a single model. The fragment where the dragon and the girl fly toward the lake, LTX 2.3 did not understand: the idea of a "flying away" object did not work for it. The author generated this segment in the cloud-based MiniMax H3 (the FL2VA Pruned 20B variant). In his assessment, MiniMax wins on quality, but he deliberately kept LTX as the main model because it was fast enough.

Next came the usual editor's work. The fragments were not always consecutive, so at some point the story had to be assembled from what was available and missing pieces had to be drawn in to make the idea look complete. He created the atmosphere, rhythm and music in CapCut. A significant part of the "style" was created not by the neural network but by the editing.

Prompts: the main cost of local generation

The most practical observation in the post concerns prompts. For local models they are far more critical than for cloud services. A text that works perfectly in one model may fail completely in another: it needs a different structure and its own specifics. This holds even for different versions of the same model.

Vlasov sees three options:

  1. Read the prompt-writing documentation. This is accurate, but it eats up time that could go toward creativity and settings.
  2. Use the built-in prompt "enhancer" in wan2GP or in third-party services.
  3. Ask a chatbot: Alice, ChatGPT, DeepSeek, and so on.

He chose DeepSeek himself. By his account, this saves time and lets him quickly rewrite a prompt if the result does not satisfy him. For teams planning local video generation, the takeaway is simple: build a separate step into the process where a chatbot rewrites the task for the specific video model, rather than relying on a universal text.

Final polish and honest caveats

Before publishing, the author ran the video through the Media Flow tab with the Detailer process on LTX 2.5 Distilled 22B (his name for the model in the post). Unlike a regular upscaler, which simply stretches pixels, Detailer redraws small details such as pores, micro-textures and hairs based on the source video. Vlasov considers the result more effective than Topaz Video and DLSS.

This is one author's subjective impression. The post contains no metrics, comparative tests or processing speed, so the superiority over Topaz and DLSS should be checked on your own material. There is also no data on what hardware was used or how long generation took.

The author's bottom line: local models can already compete with cloud ones in intended perception and quality. They lose on time and flexibility, but win on the absence of censorship, subscription costs, independence from the internet, and the ability to fine-tune and experiment. For independent creators and small studios this is a workable compromise: there are fewer ready-made "buttons," but control and budget stay with you. And for rare complex scenes, like the one with the flying dragon, it makes sense to keep a cloud model on hand as a backup.

Story

Key Questions Answered

What the video is made from?

The foundation is LTX 2.3 in its distilled version (Distilled 1.1). Distillation makes the model lighter and faster, usually at the cost of a small loss in quality. On top of it, the author added a style LoRA called Luxe_Sensual_Style. A LoRA is a small add-on that shifts the look of the image without retraining the entire model. It is activated with a special trigger word.

What are the key details on Prompts: the main cost of local generation?

The most practical observation in the post concerns prompts. For local models they are far more critical than for cloud services. A text that works perfectly in one model may fail completely in another: it needs a different structure and its own specifics. This holds even for different versions of the same model. Vlasov sees three options: Read the prompt-writing documentation. This is accurate, but it eats up time that could go toward creativity and settings.

What are the key details on Final polish and honest caveats?

Before publishing, the author ran the video through the Media Flow tab with the Detailer process on LTX 2.5 Distilled 22B (his name for the model in the post). Unlike a regular upscaler, which simply stretches pixels, Detailer redraws small details such as pores, micro-textures and hairs based on the source video. Vlasov considers the result more effective than Topaz Video and DLSS. This is one author's subjective impression.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…