houseofboern/pose-pack-krea2-v1

Pose Pack — Krea 2 Raw pose LoRA (v1)

Puts a named body position onto whoever else is in the prompt. It carries no identity — stack it with a character LoRA, which supplies the face and body.

Trained on 58 real Instagram photos, each captioned so the pose is a promptable knob.

Files

Path	What it is
HOW-TO-PROMPT.md	⭐ Read this first. All 58 poses with the exact wording to paste
checkpoints/	8 checkpoints, every 250 steps (250 → 2000)
dataset/	the 58 training images + their .txt captions
samples/	6 fixed-seed test prompts rendered at every checkpoint
posepack_krea2_v1.yaml	the exact ai-toolkit config
SAMPLE-PLAN.md	how the run was judged, and why those 6 samples

How to use it

1. Load this LoRA and your character LoRA together.
2. Pick a pose from HOW-TO-PROMPT.md. Copy its BODY and HANDS lines.
3. Write the prompt in 12-slot order. BODY goes at the end of slot 2, HANDS is slot 6.
4. Fill the rest yourself — subject, outfit, hair, light, background. Those are the free knobs.
5. Paste the pose wording verbatim. Reword it and you are asking for a pose it never saw.

Aim for ~1,100–1,800 characters. A short prompt does not fire the register.

⛔ Never write: quality words (realistic, 8k, high quality), opinion words (beautiful, sultry), the word or, or anything absent from the picture (no jewellery).

Which checkpoint

Start at 1000 or 1250. Judged by eye on the sample sheet, the poses were landing by step 1000 and nothing changed dramatically after 1250. Later checkpoints are included so you can check for yourself — the final pick is an eye call, not a step number.

Training recipe

Copied wholesale from two demonstrated Krea 2 Raw runs (ahnparker_krea2_v1, and a character run at 1024 / rank 32 / lr 1e-4). Nothing was tuned by reasoning.

Setting	Value
Base	krea/Krea-2-Raw (arch: krea2)
Rank / alpha	32 / 32
LR / optimizer	1e-4 / adamw8bit, flowmatch
Resolution	1024, batch 1, gradient checkpointing OFF
timestep_type / content_or_style	linear / balanced
Caches	cache_latents_to_disk, cache_text_embeddings, unload_text_encoder all ON
Trigger word	none — this is an always-on register LoRA
Steps run	2000 of a planned 3000, stopped early on the samples
Hardware	1× B200 @ 1000W, ~1.2 it/s at 1024, ~29 min

Honest findings from this run

• Two training images was enough for the mirror-selfie pose to land clearly. The worry that low-count poses would be too thin did not hold at that count.
• One training image was not enough. The single-image pose (bent forward over a counter) came out cooked. If you want a pose reliably, get more than one example of it.
• Poses were landing by step 1000 and little changed by 2000, on 58 images across ~28 pose groups. Hence stopping at 2000.
• The control sample (same prompt, pose wording deleted) stayed clearly different from the posed version — that is the evidence the captions are actually binding.

Caveats

• 58 images, most poses appearing only once or twice. This is a v1 probe, not a finished tool.
• Expect this LoRA and a character LoRA to pull against each other. Sweep the strengths.
• Captions were written by a vision model from contact sheets, then audited (57/58 clean on a banned-word/slot audit). They are good, not perfect.