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.
Tested on M1 Max 64GB ComfyUI: one character LoRA across 6 Anima derivatives, same prompt/seed. Trigger-only never stabilizes, RDBT bolts to beast-ears, structure tags fix it.
Anima is a Cosmos-based DiT, not SDXL, so one Anima LoRA loads on every derivative checkpoint. 20+ CivitAI Anima models sorted by type, aesthetic, and prompt adherence.
Tested on M1 Max ComfyUI: two WAI-Anima character LoRAs in one image. Side-by-side works, but only non-overlapping inpaint keeps the design pixel-exact; Qwen-Image-Edit elongates the side ponytail even with training tags, and overlapping interaction poses jam at skeleton extraction.
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.
Ran WAI-Anima v1.0 with a custom character LoRA on an M1 Mac to see if 2- and 3-character compositions actually hold up. Notes on what breaks and what holds at different LoRA weights, with practical settings that stay stable.
Tested on M1 Max 64GB: SetLatentNoiseMask returns unchanged images on Anima and Anima-Turbo with no error. LanPaint alone takes 32 min per image; adding Inpaint-CropAndStitch drops text inpaint to 2:31 and clothing replacement to about 7 min.
Tested on M1 Max 64GB ComfyUI: Anima-Base v1.0 matches preview3-base in speed; WAI-Anima kana LoRA hits 22% on light prompts but 67% with hood+robe+embroidery added.
Tested Klein 9B + 9B NSFW LoRA on M1 Max 64GB via mflux 0.17.5: 1m51s/512, 5m37s/1024 q4, 224/224 LoRA keys match, NSFW prompts uncensored, Japanese subjects work with helper tokens.
Klein 4B / 9B / Base LoRAs aren't cross-compatible — a 9B NSFW LoRA throws 'lora key not loaded' on mflux's 4B path. The variant map, what mflux runs today, and where the working hands-on test lives.
LoRAs trained directly on Z-Image-Turbo come out blurry at 8 steps but fine at 30 — the distilled trajectory breaks. How Ostris's zimage_turbo_training_adapter avoids this as training-time-only scaffolding, plus SDXL LoRA incompatibility and Z-Image caption notes.
