ArtificialGuyBR

Home / Blog / Training LoRAs

ObjectTypography LoRA for Wan2.2 T2V 14B

Train your videos to spell words with crops, light, neon, and rust using this Wan2.2 LoRA.

1 sources cited Training LoRAs

On this page

  1. What Is ObjectTypography?
  2. How It Was Trained
  3. Trigger Words and Prompt Format
  4. How to Use in ComfyUI
  5. Recommended Settings
  6. Common Errors and Fixes
  7. FAQ
  8. Sources

ObjectTypography LoRA for Wan2.2 T2V 14B

TL;DR:

What Is ObjectTypography?

ObjectTypography is a creative technique where words and letters are not printed on a surface — they are formed by objects themselves. Crops at different growth stages spell "GROW". Dusty sunlight beams spell "DECAY". Skateboard chalk trails spell "GRIND". Neon signs spell "GLOW". Crushed cars spell "RUST". Lightning bolts spell "STRIKE".

The ObjectTypo Redmond WAN2 T2V 14B LoRA fine-tunes Wan-AI/Wan2.2-T2V-A14B so the model can compose these scenes on its own, taking a word's meaning and turning it into a physical arrangement of elements in the frame.

How It Was Trained

The model was trained on prompts that describe each letter being constructed from scene elements. Each example follows the same pattern: a cinematic camera angle, a description of how each letter is formed, the materials and lighting, and an atmospheric adjective (pastoral, post-apocalyptic, cyberpunk, etc.).

The LoRA ships as two weight files:

This gives you control over the noise schedule — use the high-noise variant when your workflow introduces randomness early, and the low-noise variant for schedules that add noise later.

Trigger Words and Prompt Format

Always start your prompt with the trigger phrase:

ObjectTypography. <description of how a word is formed by objects>

Example prompts from the model card:

  1. GROW — Aerial drone shot revealing the word "GROW" formed by crop growth stages: freshly plowed soil for 'G', young sprouts for 'R', mature wheat for 'O', being harvested for 'W'.
  1. DECAY — Wide shot of an abandoned warehouse where sunlight beams through broken roof panels spell "DECAY".
  1. GRIND — Bird's eye view of a skate park where chalk trails and scuff marks spell "GRIND".
  1. GLOW — Rain-soaked brick wall at night where vintage neon signs in different colors spell "GLOW".
  1. RUST — Massive junkyard with crushed vintage cars stacked to spell "RUST" in 30-foot tall letters.
  1. STRIKE — Lightning during a thunderstorm forming "STRIKE" across the sky in a single moment.

How to Use in ComfyUI

  1. Download both .safetensors files and place them in ComfyUI/models/loras/.
  2. Load your base Wan2.2-T2V-14B model in ComfyUI.
  3. Add a Load LoRA node and select the ObjectTypo variant.
  4. Connect the LoRA output to your KSampler.
  5. Set strength_model to 0.8–1.0 for a clear object-typography effect.

Common Errors and Fixes

Error | Cause | Fix | Source

|---|---|---|---

LoRA does nothing | Strength too low | Increase to 0.8+ | HF model card Garbled letters | Prompt missing trigger | Always start with ObjectTypography. | HF model card Wrong variant | Noise mismatch | Match variant to sampler schedule | Repo file names

FAQ

Do I need both LoRA files? You can use either, but having both lets you switch based on your generation pipeline's noise strategy.

Can I use this for still images? The base concept works in principle, but this LoRA is trained specifically for text-to-video.

What camera styles does it handle? Aerial drone shots, wide shots, bird's eye views, high-angle shots, and wide-angle shots are all covered in the training examples.

Sources