mikeyasnov/angels-loras

Angels Space — Character LoRAs

Per-angel LoRAs for the AI-influencer pipeline (Qwen-Image base).

Layout:

<angel_id>/
  <trainer>/
    <target>/
      <version>/
        lora.safetensors
        README.md   (training params, dataset, recipe)

• trainer — local (ai-toolkit on A10) | fal (fal.ai/qwen-image-trainer)
• target — face (face-only crops) | full_body (full-image crops)
• version — v1, v2, ... (incremental per slot)

Trigger token across all darcy_haan LoRAs: drcyhn.

Inventory

path	rank	steps	dataset	notes
darcy_haan/local/face/v1	32	1500	36 face crops (brand pack Angles+Espressions+MU)	A10 / ai-toolkit
darcy_haan/local/full_body/v1	16	1500	33 source PNGs + Florence-2 verbose captions	identity weak (verbose-caption dilution)
darcy_haan/local/full_body/v2	32	2500	33 source + 160 augmented_kept, short captions	identity OK at strength=0.8 in canonical Lightning stack
darcy_haan/fal/face/v1	32	1500	36 face crops (brand pack Angles+Espressions+MU)	fal.ai/qwen-image-trainer ($3)

Recommended use (cloud.comfy.org photoflow)

Apply via LoraLoaderModelOnly chained after Lightning + samsungcam:

UNET → Lightning-8steps → samsungcam @ 0.6 → <character_lora> @ 0.8 → ModelSamplingAuraFlow

Prompt template (Fabio canonical):

drcyhn, woman, <scene>, athletic toned figure, narrow waist, round glutes, long legs,
natural sun-kissed tan with subtle natural tan lines, young 24 year old skin, smooth and youthful,
photorealistic