📺 crtanim — CRT / Retro Terminal Video LoRA for LTX‑2.3 22B

An open-source LoRA for LTX‑2.3 22B that locks generations into a real late-80s / early-90s CRT-monitor look — scanlines, phosphor glow, chromatic aberration, barrel curvature, dithering, signal noise, choppy limited-frame motion.

Trained on fal. Runs on fal, ComfyUI, or any other LTX‑2.3 22B runtime that accepts a standard .safetensors LoRA file.

Trigger word: crtanim, (first token of every prompt).

🚀 One-click test on fal: playground

---

Checkpoints

File	Steps	Best for
crtanim_4000.safetensors	4 000	Looser, more expressive motion
crtanim_10000.safetensors	10 000	Cleaner text, sharper structure (recommended)

Both files are already in ComfyUI key format — drop into ComfyUI/models/loras/, no conversion needed.

Direct URLs:

• 4 000: https://v3b.fal.media/files/b/0a96b375/yYgKwLBgP9cIx7yJMT0Iy_lora_weights_step_04000.safetensors
• 10 000: https://v3b.fal.media/files/b/0a96eb30/QratjdCIjYTgJe6jB47OY_lora_weights_step_10000.safetensors

---

Showcase

Range (6k vs 8k vs 10k training steps)

Terminal text rendering (10k is the one)

---

Usage — fal

curl --request POST \
  --url https://fal.run/fal-ai/ltx-2.3-22b/text-to-video/lora \
  --header "Authorization: Key $FAL_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "prompt": "crtanim, a dense CRT terminal typing \"$ claude --continue\" in glowing green pixel font, scanlines, phosphor glow, low choppy frame rate.",
    "loras": [{"path": "https://v3b.fal.media/files/b/0a96eb30/QratjdCIjYTgJe6jB47OY_lora_weights_step_10000.safetensors", "scale": 1.0}],
    "video_size": {"width": 1024, "height": 1024},
    "num_frames": 121,
    "fps": 24,
    "enable_prompt_expansion": false,
    "negative_prompt": "blurry, low quality, jpeg artifacts, compression artifacts, watermark, signature"
  }'

Important settings:

• enable_prompt_expansion: false — keeps your prompt intact
• Override negative_prompt — the default negative includes "text, on screen text, titles" which breaks readable text
• video_size 1024×1024 matches training ratio

Works on every LTX‑2.3 22B endpoint: text-to-video/lora, image-to-video/lora, distilled/text-to-video/lora.

---

Prompt recipe

Always start with crtanim, then describe, in order: CRT aesthetic → color palette → animation style → subject → composition → literal text (in quotes) → mood.

Example:

crtanim, a CRT aesthetic with scanlines, green phosphor glow, bloom, chromatic aberration, a terminal typing out exactly "$ ls -la" character by character in glowing green blocky pixel font, a blinking cursor below, static centered composition, low choppy frame rate, hacker mood.

---

Caveats

• Experimental first release — text rendering and motion can be hit-or-miss.
• On-screen text works best with short strings (1–3 words / short commands).
• The LoRA pulls toward static framing; ask for camera movement explicitly if you want it.
• Always override the default negative prompt if you want readable text.

---

Training

• Base: LTX‑2.3 22B · Trainer: fal-ai/ltx23-video-trainer
• Aspect 1:1 · LR 2e‑4
• Dataset: small hand-curated set of CRT / pixel-art clips

License

Apache 2.0.

Credits

By @lovisdotio. Trained on fal.