deepglugs/lotf-models

Legacy of the Fallen — character models

Character identity models for the cast of Legacy of the Fallen, published so community modders can generate art that matches the base game.

If you're writing a mod, see the game's modding guide.

What's here

Two model families. Each ships as a merged checkpoint (character training baked into the base model — no LoRA tag needed) and a runtime LoRA (same training, separate file, for stacking with other concept LoRAs).

File	Size	Base	Type
illustrious_lotf_v1.safetensors	6.5 GB	waiNSFWIllustrious_v140	merged checkpoint
klein_lotf_v1_fp8.safetensors	8.5 GB	Flux.2 Klein 9B	merged checkpoint (fp8)
lora/lotf_sdxl_v1.safetensors	306 MB	Illustrious / SDXL	runtime LoRA
lora/lotf_v2_klein.safetensors	1.1 GB	Flux.2 Klein	runtime LoRA

Which do I want?

• Illustrious — fast, anime/illustrated style. The usual first pass.
• Klein — photoreal-leaning, and supports true masked inpainting.
• Prefer the merge for straight identity/style generation. Reach for the runtime LoRA only when stacking with other concept LoRAs the merge doesn't include (weight 1.0).

Triggers

All models use the same two-part trigger:

lotf, <character>, <description of the character>, ...

lotf selects the house style; the character token selects the likeness.

Character	Token	Canonical look — state these explicitly
Pegasus	pegasus	long blonde hair, two dark horse ears, large white feathered angel wings
Ceraphina	ceraphina	fair ivory skin, glowing pale blue eyes, long silver-white hair in a low ponytail with a thin gold band, slim red tribal streak face tattoo on both cheeks
Chrys (male)	chrys, 1boy	shoulder-length dark brown hair, blue-gray eyes, muscular tan athletic build, light stubble
Chrys (female/futa)	chrys, 1girl	long silver-white hair, blue-gray eyes, sun-tanned golden bronze skin, athletic build
Nashoba	nashoba	long white hair, fluffy white wolf ears, one large bushy white tail, fair skin, blue eyes
Kallirhoe	kallirhoe	—

Always describe the character; don't rely on the token alone. The tokens bind likeness, not every attribute, and several have base-model priors working against them.

Known traps

• pegasus collides with the "winged horse" prior. A bare trigger can render a literal horse. Anchor it: pegasus, a human woman with long blonde hair, two dark horse ears, large white feathered angel wings.
• nashoba can render an animal head on terse close-up prompts. Lead with nashoba, 1girl, woman, human face; negative animal head, wolf muzzle, snout. If the tail duplicates, add single bushy tail + negative multiple tails, extra tails.
• Ceraphina and female-Chrys blend together — both silver-haired. The reliable separator is skin tone (Ceraphina = fair ivory, Chrys-f = tanned bronze), not markings or hair.
• On the Illustrious models, Ceraphina renders tan-skinned with a short bob. Rescue with ceraphina, 1girl, solo, pale ivory white skin, fair skin, long silver-white hair + negative dark skin, tan, bronze skin, short hair.
• Multi-character scenes bleed features. Klein separates pairs better than Illustrious; for 3+ characters, render each solo and composite.

Matching the game's art

The base game renders at 2560×1080. Generate at or above that and downscale — never upscale into it.

Two production notes that save time:

• Background removal (rmbg and friends) deletes wings, tails, and other detached appendages that run off the edge of the frame, keeping only what it recognises as a person. Keep the whole silhouette in frame, or render on flat chroma green and key on colour instead.
• When combining a character pass with other generative edits, whichever step restores a flat chroma background must run last — a character/identity pass replaces flat chroma with a lit backdrop, after which the background and subject colour ranges overlap and keying becomes impossible.

Example workflows

Ready-to-run ComfyUI pipelines are in workflows/ — text-to-image and masked inpaint, for both merges, plus LoRA-stacking variants. They use only core ComfyUI nodes. See workflows/README.md for where each file goes and how to run one.

Klein also needs a VAE and text encoder

The Illustrious checkpoint is self-contained. The Klein workflows load the diffusion model, VAE, and text encoder separately, and two of those come from elsewhere:

• VAE (~336 MB) — Comfy-Org/flux2-dev → split_files/vae/flux2-vae.safetensors → ComfyUI/models/vae/
• Text encoder — Klein 9B uses a Qwen3 encoder. Official weights are at black-forest-labs/FLUX.2-klein-9B under text_encoder/, but in multi-shard diffusers layout that ComfyUI's CLIPLoader can't read — you need a single-file build. Community repackages exist in several precisions; search HuggingFace for flux2-klein-9b text encoder and pick one matching your VRAM.

The workflows reference these by filename; edit the clip_name / vae_name strings to match whatever you downloaded.

Usage

Merged checkpoint (no LoRA tag — trigger words alone):

lotf, pegasus, a human woman with long blonde hair, two dark horse ears,
large white feathered angel wings, standing, full body

Runtime LoRA (append the tag at the end of the prompt):

lotf, nashoba, 1girl, woman, human face, long white hair, white wolf ears,
one large bushy white tail, <lora:lotf_v2_klein:1.0>

License

These models depict characters from Legacy of the Fallen and are published for use in mods and fan work for that game. They are adult-content capable; you are responsible for what you generate and where you publish it. Not for use in training commercial models or for generating content depicting real people.