Anima RenderStyle v1 (style LoRA)

Style LoRA for Anima (CircleStone Labs, Cosmos-Predict2-2B DiT). Trained with kohya-ss/sd-scripts anima_train_network.py on Anima-Base v1.0.

Dataset: RenderStyle v2 (SFW, 249 imgs) + non-muscular men set (19 imgs) = 268 images, natural-language captions, no trigger word (always-on style).

Training config

base: anima-base-v1.0.safetensors + qwen_3_06b_base TE + qwen_image_vae
network: networks.lora_anima, dim=32, alpha=32, DiT-only (--network_train_unet_only, Qwen3 TE + LLM adapter frozen)
lr=2e-5, AdamW, constant; timestep_sampling=sigmoid; bf16; gradient checkpointing
resolution 1024 (bucketed 512–2048), batch 1, 16 epochs, save every 2 epochs
final avr_loss: 0.08640

Checkpoints

anima-renderstyle-v1-000002.safetensors
anima-renderstyle-v1-000004.safetensors
anima-renderstyle-v1-000006.safetensors
anima-renderstyle-v1-000008.safetensors
anima-renderstyle-v1-000010.safetensors
anima-renderstyle-v1-000012.safetensors
anima-renderstyle-v1-000014.safetensors
anima-renderstyle-v1.safetensors

Pick a checkpoint visually. Load in ComfyUI with Anima nodes; DiT-only LoRA needs no sd-scripts->Comfy conversion.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Zaytron40k/anima-renderstyle-v1

Base model

nvidia/Cosmos-Predict2-2B-Text2Image

Finetuned

circlestone-labs/Anima

Adapter

(91)

this model