A character LoRA for Anima, trained on Heibei's own published images for thebounty "Creating my Character in Anima" (civitai.com/bounties/12716).Trigger: heibei WHAT IT IS FORAnima already draws Heibei fairly well from her booru tags alone — she is acombination of common tags, not an unknown identity. This LoRA was only keptbecause it measurably beats the naked base model on her: judged blind, mixed inunlabelled against base-model images, it won 77% of 31 duels, 95% CI [67%, 89%]. READ THE VERSION NOTES BEFORE JUDGING IT.There is a known issue with the reference prompts themselves that is not causedby this LoRA, and a one-line fix. It is documented below.

Heibei (Anima) - character LoRA (commission)

<pre><code>SETTINGS
 Base Anima-Base v1.0. Also works on Anima-Turbo, JANIMA and the
 all-in-one. Turbo is an additive distillation merged into Base, so
 a LoRA trained against Base carries over; the reverse is not true.
 Strength 1.0
 Sampler er_sde Scheduler beta
 CFG 4.5 Steps 30
 Size 832x1216 portrait, or 1216x832


KNOWN ISSUE — AND IT IS NOT THE LORA

 The reference prompts for this character describe a complete character sheet:
 identity, full outfit, simple background. Anima answers with the booru
 convention that goes with it — several views in one frame, chibi insets,
 sometimes a small second figure.

 This happens on the naked base model too, with nothing loaded.

 Add this to your negative prompt and it goes away:

 multiple views, reference sheet, character sheet, chibi, thought bubble,
 2girls, multiple girls

 Measured on paired seeds: 2 malformed images out of 6 without it, 0 out of 6
 with it. No cost to pose or composition.


IF AN IMAGE COMES OUT MALFORMED, CHANGE THE SEED — NOT THE STRENGTH

 Strength was tested blind at 0.5 / 0.7 / 0.9 / 1.1. None of the six
 comparisons separated: the spread between the best and worst strength was
 20 Elo points, while the spread WITHIN a single strength was 88 to 92. The
 seed matters about four times more than the strength.

 Some seeds fail at every strength — one produced a missing arm at 0.5, 0.9
 and 1.1 alike. Rerolling costs one image; hunting for the right strength
 costs an afternoon and fixes nothing.


COMING FROM ILLUSTRIOUS? THREE THINGS CHANGE

 - CFG 5 -&gt; 4.5, and Euler a -&gt; er_sde. Anima is a rectified-flow DiT, not
 SDXL; the ancestral samplers are not its regime.
 - 18 steps is a Turbo-only budget. On Base or JANIMA, use 30.
 - There is no clip skip. Anima encodes with a Qwen3 text encoder and an
 adapter, not CLIP. Any clip-skip value is ignored.


THE REFERENCE PROMPT CONTRADICTS ITSELF

 The short-hair prompt asks for "showing nape" together with "looking at
 viewer" and "sideway glance". Nape means her back is turned; looking at
 viewer means it is not. The model resolves it by drawing two figures.

 Every sample here uses the prompt exactly as written. If you want a single
 figure reliably, drop either "showing nape" or "looking at viewer".


HOW IT WAS MADE

 Trained with kohya-ss/sd-scripts (anima_train_network) against Anima-Base
 v1.0. unet-only, dim/alpha 4, AdamW8bit 1e-4, cosine with restarts, sigmoid
 timestep sampling, 1024, 800 steps, bf16, on a single 8 GB card.

 Text encoder not trained — measured, not assumed: the 196 Qwen3 LoRA modules
 cost 1.1 GB of VRAM and bought no measurable loss difference at 400 steps.
 Civitai's own Anima trainer ships ss_text_encoder_lr = 0.

 Dataset: 38 published images, perceptually deduplicated (one pair was 0 bits
 apart on a dHash and counted once). Hair length rebalanced with num_repeats
 rather than by deleting images, because a held-out test showed each split
 winning only where its own data was. The artist style tags are written INTO
 the captions on purpose: what you leave out of a caption is what the trigger
 absorbs.

 Selection: every candidate was judged blind against the naked base model,
 mixed in unlabelled.


WHAT IT DOES NOT DO

 - The beret ornament is baked in rather than following the scene. None of the
 public captions name it, so it learned one fixed motif.
 - Character and art style are not fully separable — 36 of the 38 public images
 use the same style stack, so there is nothing to contrast it against. A
 separate style LoRA is published alongside; see "Milk (Anima)".</code></pre>

Files with SHA256: e99eb12c3bf12f0cdddd3680b8a0f37eefcbc7ffff6f24a32462d66277394498