Kroma v0.3 sensei teacher (diffusers format)
A diffusers-layout repackaging of Kroma v0.3 "sensei" (kroma-sensei-booru-e6-teacher-velocity-v0.3), the Krea 2 finetune by lodestones that served as the teacher for the on-policy distilled kroma-v0.3.1-turbo-opd. The turbo-opd checkpoint's text-fusion and text-input weights are this checkpoint's weights verbatim, and its block weights were distilled from it, which makes this the closest non-distilled base for LoRAs that will be run on turbo-opd.
Transformer weights are converted 1:1 from lodestones/Kroma/kroma-sensei-booru-e6-teacher-velocity-v0.3.safetensors (original Krea
key layout, bf16 attention/MLP weights, float32 norms) into Krea2Transformer2DModel naming, with per-tensor dtypes matching the
official Krea 2 diffusers release (bf16 for attention/MLP projections, float32 for norms,
modulation tables, timestep and text input projections). Tokenizer, text encoder (Qwen3-VL),
VAE and scheduler are the standard Krea 2 components and are unchanged.
Intended use: a base_model_name for OneTrainer LoRA training. Per the Kroma model card, LoRAs
trained against this checkpoint load unchanged onto kroma-v0.3.1-turbo-opd for inference.
Conversion: direct key rename (blocks.N.attn.wq -> transformer_blocks.N.attn.to_q, etc.);
blocks.N.mod.lin reshaped from [6*hidden] to the [6, hidden] scale_shift_table; no
numerical changes beyond the dtype cast. Key map cross-checked against ComfyUI's
krea2_to_diffusers table and the diffusers model source.
License: Krea 2 Community License (see LICENSE.pdf). Same terms as the source checkpoint.
- Downloads last month
- 34