OmerHagage/ltx2-ume-pixelart-lora

LTX-2.3 LoRA — Modern Pixel Art (Ume style)

A rank-64 LoRA fine-tuned on top of LTX-2.3 22B that biases generations toward the "modern pixel art" style of the UmeAiRT/Ume-Modern_Pixel_Art dataset. Trigger word: ume-pixelart.

Trigger word

Include ume-pixelart near the start of the prompt to activate the style. Omitting the trigger should yield a regular LTX-2.3 generation — the LoRA is gated on this token.

Example prompts:

• ume-pixelart, a tiny knight in silver armor walks through a forest clearing at golden hour, modern pixel art with crisp pixel edges and a limited palette
• ume-pixelart, a cozy ramen shop interior at night, warm lanterns flickering and steam rising gently from a bowl on the counter, modern pixel art with dithered shadows

Examples

Trigger word gating

The same generation with the same LoRA loaded — only the prompt differs (with vs without the ume-pixelart trigger). Without the trigger the model produces a standard LTX-2.3 generation; with it, the pixel-art style activates. Both were rendered with the high-quality two-stage pipeline from ltx-pipelines.

With ume-pixelart trigger	Without trigger

Base model vs base + LoRA (10 prompts)

Each pair below uses the same prompt, same seed, same generation settings — only the LoRA differs. Generated with scripts/compare_lora.py at 768×768×41.

1 — Knight in a forest clearing

ume-pixelart, a tiny knight in silver armor walks slowly through a forest clearing at golden hour, modern pixel art with crisp edges and a limited palette

Base model	+ LoRA

2 — Cozy ramen shop at night

ume-pixelart, a cozy ramen shop interior at night, warm lanterns flickering and steam rising gently from a bowl on the counter, modern pixel art with dithered shadows

Base model	+ LoRA

3 — Astronaut on a violet alien planet

ume-pixelart, an astronaut plants a flag on a violet alien planet as two moons drift across the sky, modern pixel art, vivid saturated colors

Base model	+ LoRA

4 — Black cat on a rainy windowsill

ume-pixelart, a black cat sleeps on a windowsill while rain streaks down the glass behind it, modern pixel art with soft lighting

Base model	+ LoRA

5 — Wizard reading in a stone tower

ume-pixelart, a wizard in a tall blue hat reads an ancient tome at a candlelit desk in a stone tower, modern pixel art, warm orange light

Base model	+ LoRA

6 — Sailing ship in a storm

ume-pixelart, a sailing ship cuts through stormy waves under a lightning-streaked sky, modern pixel art with dramatic contrast

Base model	+ LoRA

7 — Cherry blossoms along a canal

ume-pixelart, a row of cherry blossom trees along a quiet canal in spring, petals drifting in the breeze, modern pixel art, pastel palette

Base model	+ LoRA

8 — Samurai at a bamboo grove

ume-pixelart, a samurai stands at the edge of a bamboo grove, kimono blowing in the wind, modern pixel art with strong silhouettes

Base model	+ LoRA

9 — Cyberpunk neon city street

ume-pixelart, a futuristic neon-lit city street at night seen from above, rain reflecting purple and cyan signs, modern pixel art, cyberpunk palette

Base model	+ LoRA

10 — Child running across a sunflower field

ume-pixelart, a child runs across a sunflower field with a kite trailing behind them, modern pixel art, bright midday sun

Base model	+ LoRA

Training summary

Base model	LTX-2.3 22B
Dataset	UmeAiRT/Ume-Modern_Pixel_Art — 100 static JPGs, auto-captioned
Trigger word	ume-pixelart
Training resolution	768 × 768 × 1 (image-only / single-frame LoRA)
LoRA rank / alpha	64 / 64
Target modules	to_k, to_q, to_v, to_out.0 (attention only)
Steps	2000 (~20 epochs at batch=1)
Learning rate	1e-4
Optimizer	AdamW
Mixed precision	bf16
Toolkit	Lightricks/LTX-2 ltx-trainer

The LoRA was trained on stills but transfers to video at inference time — the spatial-style biases learned during training are applied identically at every video frame, while the base model's motion circuitry handles temporal coherence.

Files

• lora_weights_step_02000.safetensors — the final checkpoint (rank 64, bf16)
• comparison/prompt_NN/{base,lora}.mp4 — 10 paired clips: same prompt, same seed, with vs without the LoRA applied. Generated via scripts/compare_lora.py from the trainer toolkit.
• samples/hq_with_trigger.mp4 — high-quality two-stage pipeline output with the ume-pixelart trigger in the prompt.
• samples/hq_no_trigger.mp4 — same LoRA loaded, without the trigger word in the prompt. Demonstrates that the LoRA's effect is gated on the trigger.

License

The LoRA weights in this repository are released under the MIT License — you are free to use, copy, modify, merge, publish, distribute, sublicense, and sell them, provided the MIT notice is preserved.

Note that the LoRA is derived from and requires the Lightricks/LTX-2.3 base model. Use of the base model is governed by its own license; nothing in this MIT grant overrides that. To use this LoRA you must comply with both this MIT license (for the LoRA delta) and the LTX-2.3 base-model license (for the base weights).

The training images are MIT-licensed per the upstream UmeAiRT/Ume-Modern_Pixel_Art dataset.