CoolFace
Modelpublic

orlecampos/maya-lora

sourceHugging Faceotherupdated 5d agoView on Hugging Face
0likes51downloads
Model Card

Maya (maya_rios) — identity LoRA v1 (4B)

Identity LoRA for the fictional adult character Maya. Trained on the FLUX.2-klein base 4B and intended to be applied to the distilled Klein model.

Files

  • —loras/maya_rios_v1_4b_000002250.safetensors — the elected checkpoint (step 2250), in ComfyUI layout (loras/), so it can be loaded directly by ComfyUI / Comfy Cloud.
  • —lora-maya-4b-v1/ — the full training archive: every 250-step sampler checkpoint, the final state (maya_rios_v1_4b.safetensors, step 2500), config.yaml, optimizer state and sample grids.

Usage

  • —Trigger token: maya_rios
  • —Recommended weight: 0.5 – 0.7, alongside the character's prompt canon (hair, skin, eyes and styling described in text). Lower weights keep the base model's realism; higher weights lock identity harder.
  • —Elected by inspecting the sampling grids, not by step number: intermediate steps 750/1500 collapse framing, and 2000/2250 are the solid ones.
  • —Train on base, run on distilled — do not apply this LoRA to the base model expecting distilled speed.

Training

ItemValue
BaseFLUX.2-klein base 4B (Apache-2.0)
Steps2500 (batch 2, 1024 px)
Rank / alpha128 / 64
Learning rate1e-4, flowmatch scheduler, 8-bit optimizer, bf16 + qfloat8
Dataset88 image+caption pairs — thirds of close-up / mid-body / full-body, 2784x4192

The 9B variant of Klein is non-commercial, which is why this LoRA was trained on the 4B base.

Notes

  • —The character is a fictional adult; do not use this model to depict real people or minors.
  • —Published for the owner's own pipeline (Comfy Cloud). Public visibility does not grant third parties a license for commercial use.