Lanternwake turns characters, animals, props and scenery into a Gothic cut-paper theatre: layered matte cardstock, fibrous cut edges, folded shapes and crisp cast shadows. Its palette combines ink black, warm ivory, burnt amber and midnight blue.
Use with Krea 2 Turbo. Trigger: Lanternwake Halloween theatre. Start at strength 0.8; 1.0 gives stronger paper rendering. Tested in ComfyUI at 1440 × 1440, 8 steps, CFG 1, Euler ancestral / beta. Apply to the diffusion model only.
Trained on Krea-2-Raw with 48 curated synthetic images and 8 held-out scenes, including four girl portraits and four adult-woman portraits. Rank/alpha 32, 1024 buckets, 768 updates; BF16 final export. Training data was generated with built-in imagegen and visually reviewed.
Characters and scenic subjects show stronger paper rendering than smooth objects such as a teacup. Ordinary prompts can gain pumpkins, lanterns or night scenery. Omitting the trigger reduced added night scenery in the tested comparison, but did not fully remove Halloween motifs. Faces are stylized and may look younger.
The gallery PNGs preserve their native pixels and contain generation metadata. The accompanying short film is generated locally with MiniMax H3 from the owl PNG; Lanternwake creates the reference image, while H3 generates motion and audio. Video prompt and settings are supplied with the video.
Description
Final selected 768-update export. Ordinary LoRA, rank/alpha 32, trained on Krea-2-Raw; use with Krea 2 Turbo. BF16, 228,615,912 bytes. SHA-256: fcc7e35aa5bc2aa94a4d949dbc3a76a6b3980f33fa35f5574a51ba8f931feeb3.
48 reviewed training images + 8 held-out scenes, 16 epochs, 768 optimizer updates. 130 native evaluation images checked across checkpoints, two seeds, strengths 0/0.5/0.8/1.0, trigger removal and new girl/adult-woman portraits. Start at 0.8; model-only LoRA, no text-encoder adapter.