headswap-qwen-edit-2511

Head-replacement LoRA for Qwen-Image-Edit-2511

Trains the two-image edit path to take a child's face and hairstyle from a reference photograph and draw that child into an existing illustrated page, preserving the page's art style, background, composition, pose and clothing.

Inputs

Order matters: scene first, photo second (the convention BFS Head V3/V4/V5 use).

Prompt used in training, and the sentence the adapter is addressed by:

Replace the child in the first image with the child from the second image,
keeping the same face, hair colour, hair length and hairstyle as the second
image. Keep the background, composition, lighting, pose, clothing, hands and
all other characters unchanged.

Render at cfg 4.0 with a " " negative prompt, 25-40 steps, 1344x768.

Checkpoints

lavuribooks_headswap_v2.safetensors is the final (1500 steps); the numbered files are intermediate saves at 250-step intervals, kept so a checkpoint can be chosen after the fact rather than guessed in advance.

Training

LoRA rank 32, lr 1e-4, adamw8bit, bf16 without quantisation, batch 1, 1500 steps on an RTX PRO 6000. 640 triplets from 31 target pages, split by character so validation never shares an identity with training.

Known limitations

• Clothing from the reference can leak into the scene.
• A reference that is itself a scene (a child photographed in a room) is sometimes treated as the canvas rather than as a reference.

Both are data gaps rather than tuning problems.