Main prompt structure: A chubby man with glasses. He is wearing a blue polo shirt. He says in Japanese "Kon'nichiwa, hajimemashite". Medium shot. Anime style.

Note: You can further enhance the prompt, LTX2 likes long detailed prompts.

Recommended models:
- Lora tested on LTX2-19b-DEV-FP8 + LTX2 Distilled Lora + abliterated Gemma3 text encoder but it should work on all versions of LTX2/Gemma3

Workflow examples for I2V and T2V + explanations on parameters can be found here: https://civarchive.com/articles/26140/

Recommended generation settings:
- Sampler: (1st and 2nd stage): Euler
- Resolution: wide aspect ratio (e.g. 1280x704 / 1920x1024) is preferred, other resolutions like 1:1 square (1024x1024) or portrait (768x1024) will also work but may cause problems
- Frame rate: I recommend 25 by default, however, using higher frame rate can increase motion and reduce blurry artifacts
- CFG: 1 Steps: 8

Note: CFG/Steps will depend on what kind of workflows, models and Loras you use (e.g. Distilled). Refer to gallery videos comments for examples and parameter recommendation from the linked article.

Prompt tips (green - positive, red - negative):
- Anime style - required to keep the style, you can make him fully realistic if you remove this tag or use Realistic instead
- For correct Japanese dialogue use He says in Japanese " " and put text as romaji in the quotes
- Shot types: Close-up shot, Medium close-up shot, Medium wide shot, Wide shot
- He is wearing a grey cap, a white and green striped polo shirt over a blue undershirt - alternative clothes

I2V support:

- The Lora works with I2V, however, do not expect amazing results. It was not trained with I2V focus and LTX2 right now does not perform great with flat 2D styles either.

Most of the I2V source images can be found on my Seiichi Kinoshita Illustrious page.

Known issues:

• The Lora is currently biased towards close-up/medium shots by default. Solution: provide more details about what the character is wearing (e.g. trousers, shoes) to force wider shots or change resolution to square/portrait

• Will occasionally generate in realistic style. Solution: either change seed or add "Anime style animation" or "Animated in Anime style" in the prompt

• In weird/unusual angles will sometimes lose character consistency and become skinny Solution: add fat/obese to describe him

Training info:

• Trained on 82 videos and 8 images using the Musubi LTX2 fork: https://github.com/AkaneTendo25/musubi-tuner

• Training buckets used - images: 960x512 / 864x864 / 1152x640 || videos: 736x416 / 800x448

• Frame length: between 25 and 121 frame videos

• Training time: 7000 steps total, Lora taken from 5300 epoch

• Training parameters below extracted from Lora metadata:

"ss_logit_mean": "0.0",
"ss_steps": "5300",
"ss_logit_std": "1.0",
"ss_max_grad_norm": "1.0",
"ss_discrete_flow_shift": "1.0",
"ss_fp8_base": "True",
"ss_base_model_version": "ltx2_v1",
"ss_optimizer": "bitsandbytes.optim.adamw.AdamW8bit(weight_decay=0.0001)",
"ss_max_train_steps": "7000",
"ss_network_dropout": "0.1",
"ss_gradient_accumulation_steps": "1",
"ss_mixed_precision": "bf16",
"ss_sd_model_name": "ltx-2-19b-dev.safetensors",
"ss_guidance_scale": "1.0",
"ss_full_bf16": "False",
"ss_vae_name": "ltx-2-19b-dev.safetensors",
"ss_gradient_checkpointing_cpu_offload": "False",
"ss_network_dim": "32",
"ss_network_alpha": "32.0",
"ss_lr_scheduler": "constant_with_warmup",
"ss_timestep_sampling": "sigmoid",
"ss_full_fp16": "False",
"ss_epoch": "58",
"ss_sigmoid_scale": "1.0",
"ss_lr_warmup_steps": "100",
"modelspec.title": "NewEraV14Audio",
"ss_learning_rate": "0.0001",
"ss_gradient_checkpointing": "True",