# Pallas Cat (Otocolobus manul) — SDXL v1
A LoRA for generating photorealistic Pallas cats — a rare wild
felid native to Central Asia, frequently mis-rendered by base
models as tiger, snow leopard, or generic gray cat.
## Usage
- Trigger: pallas cat, Otocolobus manul
- Recommended weight: 0.7 – 0.9
- Best results: front-facing portraits, sitting poses,
natural wildlife photography contexts
- Known limitation: side-profile and dynamic poses can produce
"anatomical chimeras" (front-face on side-body) — see the v2
notes in the linked article. v2 will address this.
## Sample prompts
pallas cat, Otocolobus manul, sitting on rock, looking at camera,
soft natural daylight, wildlife photography, photorealistic
## Training details
- Base: SDXL 1.0
- Dataset: 92 photographs (zoo livestream,
photographer portfolios)
- Rank 8, Alpha 1, 10 epochs, AdamW8bit
- Captioning strategy: variable elements only — trigger word absorbs
permanent species features
## Full write-up
Three findings I noticed during this training run — including a
chimera failure mode that more epochs don't fix — are written up at:
[Part 2 LinkedIn Article link]
## Attribution
Full photographer credits: [GitHub ATTRIBUTION.md link]
## v2 coming
v2 with diversified pose data is in training. Will be published
under this same model.
Description
v2 — Side-Profile Update
v2 adds 52 side-profile and dynamic-motion images on top of the v1
dataset (134 total) to address the front-face overfitting issue
documented in Part 2 of the Pallas Cat AI Pipeline Study
(Finding #8 — overfit chimera, front-face-on-side-body).
### What v2 improves over v1
- Side-profile and 3/4 angle generations are anatomically correct
- Reduced front-face chimera at later epochs (epoch 4 final)
- Better motion poses: walking, jumping, turning
### Where v1 is still better
- Front-facing tight portrait shots (v1 was overfit on these,
which paradoxically makes them more polished)
- Anything where you want the v1 "tight ruff close-up" baseline look
v1 is not deprecated — both versions remain available and
trade in opposite directions.
### Known v2 limitations
- Multi-tail artifacts at certain camera angles (~5–10% of generations)
- Dynamic motion drifts at CFG > 5.5
- Best results at strength 0.75–0.85
### Recommended settings
- Base: SDXL 1.0
- LoRA strength: 0.75–0.85
- Sampler: DPM++ 2M Karras
- Steps: 28–32
- CFG: 3.5–4.5
- Resolution: 832×1216 (vertical) or 1024×1024 (square)
### Trigger
pallas cat, Otocolobus manul
(same as v1 — no migration needed)
### Part 3 of the series uses v2 in a 5-way video model comparison.
Article and full eval grid linked in the project repo.