Krea2 Edit (source patch)

Krea's official edit model still hasn't shipped, so right now the best "edit a photo with a sentence" tool for Krea 2 is a community one: the Krea 2 Identity Edit LoRA by conradlocke, with this node pack supplying the plumbing. Krea2EditModelPatch is the appearance half of that plumbing. The source image you're editing gets VAE-encoded and prepended as in-context tokens inside the diffusion model, so the model treats "keep this person" as something it structurally can't forget - not as a prompt clause it might ignore.

It's the same family of trick as Flux.1 Kontext: reference images consumed in-context rather than injected as adapter embeddings the way IP-Adapter or PuLID do, here wrapped around Krea 2's single-stream DiT and matched to how the LoRA was trained. Krea2EditGroundedEncode handles the semantic half (the text encoder reads the image while reading your instruction) - you need both nodes, and this one goes between your LoRA loader and the sampler.

How it works

The node clones your model and installs a wrapper on the diffusion-model patcher. On each denoise step the VAE-encoded source is scaled into the same latent space as the target and prepended as clean tokens at RoPE frame 1 - a second reference (two-image edits) lands at frame 2. ref_boost then multiplies target→reference attention to pull harder toward (or loosen from) the reference's appearance.

You feed the source one of two ways. The latent path is just source_latent from a VAEEncode - dead simple, but if the source resolution doesn't match the output, the resampling gets sloppy and reference quality turns blurry. The pixel path (vae + source_image) crops and resizes in pixel space, then encodes internally. It's the blur-proof option, it makes mismatched aspect ratios safe, and it's the one people end up on.

The inputs that matter

• model (required) - your Krea 2 model with the Identity Edit LoRA already applied via LoraLoaderModelOnly. Patch on top of a model without the LoRA and nothing will behave.
• source_latent (required) - VAEEncode of the image being edited.
• vae + source_image (optional, recommended) - the pixel path. Required if you want fit_mode: fit to do anything.
• target_latent (optional, recommended with the pixel path) - wire the same latent that feeds KSampler.latent_image. More on this below; it's the one that bites people.
• fit_mode (default fit) - how a source fits a mismatched output aspect ratio. fit = the training-matched centered resample (v1.2). crop (legacy) = the v1/v1.1 center-crop geometry; use it only with older weights.
• ref_boost / ref_boost_a (default 1.0) - the fidelity dials. >1 pulls harder toward the reference's appearance, <1 loosens. ref_boost hits the last reference (the subject in two-ref work), ref_boost_a hits the first (the scene).
• ref_boost_mask (optional) - a region of the reference to boost, e.g. just the face.

Output is a MODEL - wire it into KSampler.model.

The target_latent gotcha

The pixel path has to VAE-encode the source at the output resolution. Without target_latent the node doesn't know that resolution until sampling starts, so the encode lands on the first step - mid-run, with the diffusion model already resident. ComfyUI frees room by partially offloading whatever's loaded, sampler included, and nothing loads it back. Every remaining step then streams weights from CPU. Wiring target_latent moves the encode to node-execution time and restores the sane VAEEncode → KSampler order. The console tells you which path you're on - look for pre-encoding sources at target ... (before sampling...) (good) versus a NOTE: connect 'target_latent' (bad). Measure over 20+ steps, not 1: this is a per-step penalty, invisible on a single-step test run.

Install

ComfyUI Manager → search "Krea 2 Identity Edit", or:

cd ComfyUI/custom_nodes
git clone https://github.com/lbouaraba/comfyui-krea2edit
# restart ComfyUI

No Python dependencies. You do need the whole Krea 2 stack - the model (Raw or Turbo), the Qwen3-VL 4B text encoder (~8 GB, the thing people forget), the Qwen-Image VAE, and krea2_identity_edit_v1_2.safetensors from HuggingFace. The README's fast path is Turbo, 8 steps, CFG 1.

Common issues

• Blurry reference / lost identity - you're on the latent path with a resolution mismatch. Connect vae + source_image; at matched resolution the two paths agree.
• fit_mode: fit does nothing - it silently requires the pixel path. The console warns when it falls back.
• Removals re-render the subject - Turbo at CFG 1 usually redraws instead of deleting. Use Raw at CFG 3, ~20 steps, with a grounded negative.
• Ghosting / doubling on outpaint - keep dimensions on the /16 grid, and prefer euler over er_sde for outpainting (the SDE noise wrecks the reference-copy channel).
• Generate ≤ 2 MP - above the trained range, source content bleeds and subjects duplicate. And "zoom out" is genuinely tricky on current weights; plan around it.