Krea2 Identity Edit (source patch) ⚡

Here's the thing about the Krea 2 Identity Edit LoRA: Krea 2's native forward pass only ever builds [text | target]. There is no built-in way to prepend a clean copy of your source photo. The LoRA, though, was trained on [text | source | target] - the source sitting there as an extra image frame the model cross-attends to while it generates the edit. This node is the wrapper that rebuilds that sequence at runtime.

It's the appearance half of the identity-edit pair. Its partner, Krea2 Identity Edit (grounded encode), feeds the same photo into Qwen3-VL alongside your prompt for the semantic half ("the man on the left"). Together they reproduce the training recipe; separately they underperform. This node takes model + a latent of your photo and wraps the diffusion model's forward so the source gets prepended as clean in-context tokens at RoPE frame 1 - the exact [text | source(frame=1) | target(frame=0)] layout from training (ai-toolkit's predict_velocity_edit). It's implemented through ComfyUI's real ModelPatcher extension point, not a from-scratch sampler, so CFG and the sampler stay fully intact.

The inputs that matter

• model and source_latent - required. The latent comes from a stock VAEEncode of your photo. Minimal graph: LoadImage → VAEEncode → here.
• vae + source_image - the recommended upgrade. This is the blur-proof pixel-space path: the node fits the raw image to the target grid in pixel space instead of resizing an already-VAE-encoded latent, which is what causes the classic fuzzy-reference look when your photo's aspect ratio doesn't match the output. fit_mode="fit" (default) is the training-matched geometry; crop (legacy) is for v1/v1.1 weights.
• target_latent - recommended whenever you use the pixel path. Wire in the same latent that feeds KSampler.latent_image. Without it, the node VAE-encodes the source on the first sampling step, which can yank the VAE onto the GPU mid-sampling and evict part of the diffusion model on VRAM-tight setups - the rest of the run silently streams from CPU. The console tells you which path you got.
• ref_boost / ref_boost_a - reference-fidelity dials (1.0 = off). Higher pulls harder toward the reference's appearance; the optimum is model-specific. In two-ref workflows, ref_boost is the subject and ref_boost_a the scene.
• source_latent_b / source_image_b - a second reference at RoPE frame 2 for two-input edits (person + scene).
• ref_boost_mask - optionally boost just a region, e.g. the face.

Output is a MODEL, which flows into your KSampler (with the source-patch applied to model and the grounded-encode CONDITIONINGs feeding positive/negative).

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
pip install --upgrade gguf

Restart, then download the Identity Edit LoRA separately - krea2_identity_edit_v1_2.safetensors from conradlocke/krea2-identity-edit - along with the Krea2 model, Qwen3-VL-4B encoder, and VAE. Nodes live under 🤖 CCTech/Krea2.

Gotchas

Wire target_latent or live with the slowdown. Match input/output aspect ratios - mismatched ones are the #1 cause of that "bad photoshop job" look. And the LoRA must actually be loaded upstream (stock LoraLoaderModelOnly at ~1.0); with no LoRA, the extra context tokens are inert, so nothing breaks but nothing changes either.