Krea 2 Scanline Attention

Krea 2 Scanline Attention

ComfyUI custom node that swaps the attention in every Krea 2 DiT block for the scanline attention pattern used to train the scanline-distilled LoRA: text tokens are a global prefix, and each image token attends only to image tokens within window // 2 latent-token rows. Pair it with the scanline-distilled Krea 2 LoRA for good results.

Nodes

• Krea 2 Scanline Attention (advanced/model_patches): MODEL -> MODEL
  • enabled — when off, the model is returned unpatched.
  • window — odd number of latent rows each image token attends to (default 11).
  • tile_aligned — admit every valid pair inside an active 128x128 tile, matching the trained graph.

Non-Krea 2 models raise an error. Calls with extra tokens (reference latents) fall back to dense attention.

Requirements

• ComfyUI >= 0.32.0 (Krea 2 model support, comfy-kitchen attention override contract)
• CUDA (uses torch.nn.attention.flex_attention)

Install

ComfyUI-Manager: search for "Krea 2 Scanline", or comfy node install krea2-scanline.

An example text-to-image workflow is in workflows/krea2_scanline_t2i.json.