anyze/wan2.1-anime-lora

Wan 2.1 T2V-1.3B anime LoRA: a general anime style, plus two example characters

A general-purpose anime style LoRA for Wan 2.1 T2V-1.3B: add it to any prompt and the small model draws anime (cel shading, clean black outlines, TV-anime colour) instead of live action, CGI or a painted render. It works for any anime scene, character, setting or genre you describe. Two character LoRAs come with it as examples of stacking a character on the style.

file	what it does	trigger	rank
wan2.1-t2v-1.3b-anime-style.safetensors	general anime style for any prompt: cel shading, black outlines, TV-anime colour; keeps the small model away from live action and CGI	bldln style anime	32
wan2.1-t2v-1.3b-character-zhou-ming.safetensors	example character: Zhou Ming (young, black hair, grey hoodie)	zhmng man	16
wan2.1-t2v-1.3b-character-duncan-abnomar.safetensors	example character: Captain Duncan Abnomar (black swept-back hair, short black beard, navy captain's coat)	dncn man	16

The style LoRA works on its own; you only need the character LoRAs for those two characters. The two characters are fan versions of the leads of the web novel Deep Sea Embers; the characters belong to their author. They are meant to be stacked on the style LoRA.

Use (ComfyUI)

workflow_api_t2v_style_character.json is the exact API graph we ran (load it with ComfyUI's API, or rebuild it in the editor from the node list). Files it expects:

folder	file	from
models/diffusion_models	wan2.1_t2v_1.3B_bf16.safetensors	Comfy-Org/Wan_2.1_ComfyUI_repackaged split_files/diffusion_models/
models/vae	wan_2.1_vae.safetensors	same repo, split_files/vae/
models/text_encoders	a UMT5-XXL encoder (we used umt5-xxl-encoder-Q5_K_M.gguf through ComfyUI-GGUF; umt5_xxl_fp8_e4m3fn_scaled.safetensors from the same Comfy-Org repo with the stock CLIPLoader is the usual choice, not tested here)
models/loras	Wan2_1_self_forcing_dmd_1_3B_lora_rank_32_fp16.safetensors	Kijai/WanVideo_comfy LoRAs/Wan2_1_self_forcing_1_3B/
models/loras	the LoRAs from this repo

The chain: base model -> Self-Forcing LoRA (strength 1) -> anime style (1) -> one character (1) -> ModelSamplingSD3 shift 5 -> KSampler 4 steps, cfg 1, euler / simple -> VAE decode. 832x480, 33 frames (2 s at 16 fps) or 49/81 frames for longer shots.

The Self-Forcing LoRA is what makes it fast: 4 steps at cfg 1 instead of 20+ steps at cfg 5. With cfg 1 the negative prompt has no effect.

Prompts that worked

Start with the style trigger, then the character trigger and a full description of the character every time (the LoRA holds the face; the words hold the costume):

bldln style anime, close-up of dncn man, a tall man in his early forties with thick black hair swept back, a short
majestic black beard and deep-set eyes, in a dark navy sea captain's coat with gold buttons and gold braid, glaring
at the camera, in a dim captain's cabin with wooden walls, warm lamplight.

Shots without the character: leave the character LoRA out of that run (set its strength to 0). A character LoRA acts on every person in the prompt: with it loaded, a second man in the shot takes the character's face.

Measured

• Speed: 832x480, 33 frames, 4 steps, this exact workflow in ComfyUI 0.37 (2026-11-07, 4 runs on an 8 GB GPU): 54 s for the first clip (models loading), 48-52 s for each one after.
• Same picture either way: the Self-Forcing LoRA loaded as a LoRA and the same LoRA merged into the base model give the same clip (mean pixel difference 2.5 of 255).
• The files are exact conversions of the trained weights: for all 300 modules of each LoRA, the weight change in this file matches the trained one (largest relative difference 2.4e-7).

Known limits

• Mirrors and reflections tend to come out as a second person; an over-the-shoulder framing ("the back of his head fills the left side, a mirror on the right shows his reflection") fixed it for us in 1 of 3 seeds.
• Fog and mist are often drawn as dark blobs or rocks.
• Hiding a face ("his face in shadow", "seen from behind") is not reliable: in one shot 6 takes all showed the face.
• The style LoRA pulls every man toward black hair and dark clothes.
• Two different characters in one shot (two character LoRAs at once) is untested.
• 1.3B model: hands and small props (pens, tools) break more than in the 14B models.

Training

musubi-tuner, Wan 2.1 T2V-1.3B base in fp8, AdamW8bit, flow shift 3, bf16. Style: rank 32, lr 2e-4, 4 epochs. Characters: rank 16, lr 1e-4, 8 epochs, 512x512 buckets, each image captioned with the trigger plus the view and expression. The full pipeline (training, shot generation, voices, sound effects and the cut) is in the GitHub repo: https://github.com/PiPercentHuman/PiPercentHiman_Shareables/tree/main/anime-lora-wan.

Licence

Weights in this repo: CC BY-NC 4.0 (non-commercial). Wan 2.1 is Apache 2.0; the Self-Forcing LoRA is a separate download under its own terms.