Merged kei, kana, and koharu into a single Anima (Qwen-DiT) LoRA and ran my first training on Blackwell (RTX 5090, sm_120). Hands-on log: the cu128 / torch2.8 / SDPA stack swap from the 4090, why the weakest character gets absorbed (caption asymmetry, not rank), and how trigger-only prompts separate three close-packed characters at ep143 without ControlNet.
rank128 + 20 two-character images killed the v1 ahoge bleed and body fusion on this Anima dual-character LoRA. Lap-sit stays a Qwen3 text-encoder limit; sweet spot is ep140.
One Anima (Qwen-Image DiT) LoRA, two characters, trained on RunPod: can they touch? Hugs and piggyback hold, lap-sit fuses; stacked limbs survive, interleaved break. Best at ep100, Turbo.
On Anima-Base, my character LoRA bent its legs even on standing. Adding upright references didn't fix it; cutting 36 posed full-body images did. Subtract, don't add.
Rebaked a WAI-Anima character LoRA onto upstream Anima-Base with off-distribution Gemini data. Trigger-only usable, face fidelity beats v1, intakes still cap out.
WAI-Anima LoRA trained only on images the model itself made. Distribution shift drops to ~zero, so the sweet spot hits epoch 20 not 150 (7.5x faster). What the trigger bakes in vs what still needs tags, plus pose/angle control.
With v3 captions kept as-is and only the training amount pushed up to Anima's official 12,000+ step recommendation, the direction hit rate went 100% at ep150-180, crashed to 0% at ep200, then partially recovered to 67% at ep227 — a non-monotonic curve. 600-720 exposures per training image is the sweet spot; over 800 triggers catastrophic forgetting. Learning rate 2e-5, ~11 hours / $10 of RunPod training plus a sweet-spot epoch scan.
Alibaba ATH's video generation model HappyHorse-1.0: API test status on Model Studio, open weights availability, Mac local inference reality, and which GPU to pick on RunPod.
Trained a WAI-Anima character LoRA on RunPod (AnimaLoraToolkit + sd-scripts) for $1.22, but at inference the side-ponytail direction won't shift with Danbooru tags or natural language — a directional bias from Anima base. Full verification record.
I dropped the nervous sample identified as the culprit last time, plus 5 others, and retrained the LoRA under otherwise identical conditions. The sweat drops on ep08 angry are gone, and as a bonus, ep06 produced the closed-mouth restrained anger that the previous training never managed to reproduce.
Training an Illustrious-XL LoRA on RunPod for around $1 by doing env setup on a $0.08/hr CPU Pod and renting the 4090 only for actual training. Network Volumes attach to both pods at the same time, so there's no idle GPU billing. Four sd-scripts gotchas hit on the way included.
Testing See-through for anime character PSD decomposition: 23 generated layers, front/back hair separation, hidden-area inpainting, and what LayerDiff + Marigold actually produced from a single illustration.
