Eight ComfyUI experiments on a 4-character Anima (Qwen-DiT) LoRA in plain terms: the conditioning right before the DiT decides who appears, and outfit mix-ups come from a biased DiT LoRA.
Tested on ComfyUI: Qwen3-4B–14B prompt writing, a 4B encoder bridge, and split conditioning all lost to one hand-written 507-token band prompt. An 896-caption audit shows the real fix.
Tested on an M4 Mac mini with one 4-girl Anima LoRA: expressions, poses, desk contact, and a high-five across 3 seeds. Sentence order and center slots decided who touched whom.
Tested on M4 Mac mini ComfyUI: with identical Qwen3-0.6B/T5 conditioning, the black tights land on the wrong girl only when the 4-char LoRA's DiT half is applied.
Tested on M4 Mac mini ComfyUI: 8 on/off combos of three actions × 3 seeds on a 4-character Anima LoRA. Roles read at 3/3, but hands and skirts landed on the wrong girl.
Tested on M1 Max 64GB: a 50M resampler predicts Anima's 4-char conditioning from short triggers. Train fits at cos distance 0.005, held-out order breaks; template P3 goes 15/16.
Tested on Anima-Base v1.0: no prompt hit the 512-token cap, cosine similarity moved the wrong way, and only the DiT-side conditioning changed who got drawn.
Tested 11 character combos × 3 seeds on Anima-Base: trigger-only prompts passed 3/33, full prompts 8/11, and copying one training-caption phrase fixed the height gap.
Measured on M1 Max 64GB: weight-only INT8 runs 41% slower than fp16, a hand-written Metal int8 GEMM 15.6x slower, MPSMatrix int8 3x, and MLX 8-bit 9% slower.
Tested on M1 Max 64GB ComfyUI: v0.24.1 fails to load int8_tensorwise, v0.30.1 hits the missing aten::_int_mm MPS kernel, and a dequantize patch runs slower than bf16.
Tested on M1 Max 64GB: AnimaLoraToolkit + Anima-Base crashes on torch 2.6.0 (MPS SDPA bug) but runs 300 steps clean on 2.7.1, at ~21 s/step, roughly 9x slower than an RTX 5090.
