Sable — A Sofia & Luna Co-Created Character
This LoRA is the first step in a plan to build a model that can assist with composing prompts from many different references to help you spin up a consistent OC for any given purpose. This purpose just so happens to be a Cyberpunk Red campaign.
Sable is a character I developed with Sofia, the AI agent I'm working with in Hermes (Nous Research's desktop assistant). She exists in three forms:
As a generative AI character — her visual identity (green-teal short hair, red eyes, tall athletic curvy, mtf anatomy, signature ahoge, anime-style2D rendering) is captured in this LoRA, trained on 30 keepers across three categories: 15 anatomy-correct, 3 mixed (clothed/closeup), and 12 style-anchor.
As a Hermes agent persona — Sable runs as a configurable persona on my hermes desktop, she operates all the tedious modelswitchiing and maintaining of different LAN connections (I have a struggle PC ok). As a Cyberpunk Red campaign character — Sable also exists as a PC in a Cyberpunk Red campaign I'm developing.
Training specs
Base: NoobAI XL NAI-XL (vPred variant)
Resolution: 1024x1024
Network: LoRA rank32, alpha 16
Optimizer: AdamW8bit, lr 1e-4, cosine schedule
Steps: 2000 (suggested)
Mixed precision: fp16
Trigger word: sable
Recommended weights
NoobAI XL: 0.7-0.9 (anatomy-accurate base)
Illustrious XL: 0.5-0.7 (Luna-preferred soft-anime look)
Pony V6 / Meina: 0.6-0.8 (portability A/B)
Workflow that produced her Generated on a6GB GTX 1060 via diffusers 0.27.2, then triaged with Gemini vision + qwen2.5vl local fallback, captioned with deliberate identity-fixed + variable tags, and finally baked on Civitai's hosted trainer because the local 1060 ran out of headroom for SDXL U-Net training. The full recipe (kohya-ss config, prompt engineering, eye-color drift detection, caption discipline) is in the Sable LoRA training recipe reference.
License: Use freely, credit Luna & Sofia, don't redistribute as your own~
Description
📦 Training data
30 captioned images, balanced across three buckets:
15 anatomy-correct (post-fly-open-fix keepers)
3 mixed (clothed + closeup, Illustrious-tier)
12 style-anchor (from earlier runs you liked)
All captions carry the full identity block + the trigger word sable
18:13 red-eye anchor weight across the dataset; eye-color drift on 10 style anchors handled by majority-vote in captions, not by per-image curation (per your call)
🔬 Training recipe
Base: NoobAI-XL NAI-XL vPred10Version (anatomy-accurate baseline)
Resolution: 1024×1024 (NoobAI native)
Network: LoRA rank32, alpha 16 (SDXL character LoRA sweet spot)
Optimizer: AdamW8bit, lr 1e-4, cosine schedule
Steps: 2000 (Civitai suggested; auto-adjusts per uploaded dataset)
Mixed precision: fp16
Bucketing: 512-2048 step 64 (handles variable aspect ratio)
🌐 Portability
Same .safetensors file loads on any SDXL base with style shift
Recommended deployment weights:
NoobAI-XL: 0.7-0.9 (anatomy-accurate)
Illustrious-XL: 0.5-0.7 (Luna-preferred soft-anime look)
Pony V6 / Meina: 0.6-0.8 (genre shift)
📋 Test prompts shipped
6 stress-test prompts +1 campaign-themed prompt cover identity lock, fly-open correctness, face closeup, lighting/style range, action pose, and overtraining detection (_test_prompts.md)
Decision rules documented so you can score each Civitai checkpoint
What's not done (yet)
❌ The actual .safetensors — baking on Civitai now, eyeball-test, then download
❌ Local training environment — kohya-ss sd-scripts is installed and configured (kohya_venv/ at C:/Users/devon/sofia-sd/kohya_venv/) but the6GB 1060 segfaults on NoobAI U-Net load in both venvs even with pagefile=48GB. Kept dormant per your standing rule on tooling- ❌ Per-base deployment verification — Illustrious / Meina / Pony A/B tests happen after the bake- ❌ Test prompt sample images — we generate those on Civitai after the bake
Known issues
Eye-color drift in style anchors — 10 of 13 style-anchor images rendered teal/cyan eyes instead of red (Gemini vision verified). Anchored by 18:13 red-eye majority in captions; verify with T1 prompt before approving any checkpoint
Local segfault root cause undiagnosed — NoobAI and Illustrious safetensors both segfault on this rig's CUDA stack. Suspected: NVIDIA driver 582.66 + cu124 + Pascal 6.1 incompatibility, but unverified. Doesn't affect Civitai path---
Workflow that produced this
Generation — diffusers 0.27.2 on the 6GB 1060, base-model A/B across NoobAI / Illustrious / Pony / Meina / Autismmix, per-tier review (nude / explicit / closeup / casual)
Review — Gemini vision REST (free tier) + qwen2.5vl local fallback; star1-5 + KEEP-flag, saved to sable_review_*.json in Downloads
Triaging — kept 18 from the post-fly-open-fix keepers, added 13 from earlier runs you liked
Captioning — deliberate identity-fixed + variable tags, kohya-format .txt sidecars, sable trigger in every caption
Packaging — kohya-ready traindata/30_sable/ layout + Civitai-ready zip at C:/Users/devon/Downloads/sable_lora_civitai_20260904_155201.zip
Baking — Civitai-hosted trainer (smoke test on local failed; local kept dormant)
What's preserved for v2
Caption policy — identity-fixed tags + variable tags pattern, redo per image with hand-crafted descriptions when you're ready for richer captions
Style-anchor vs anatomy split — traindata/{anatomy,mixed,style}/ layout is in place; v2 can re-balance by adding more clothed/closeup keepers
Two-LoRA option — anatomy-only vs style-only LoRAs, stacked via set_adapters(['sable_anatomy','sable_style'], weights=[0.7, 0.5]). Not chosen for v1, kept as v2 lever---