Krea 2 Identity Edit v1.2 LoRA ships three ComfyUI workflows tuned by one ref_boost dial

Krea 2 gains a three-mode identity editor

Krea 2's identity editing LoRA has reached v1.2, and a set of three ComfyUI workflows built on it has been published covering single-image editing, masked local edits, and two-image reference transfers. According to a dev.to walkthrough that inspected the saved graphs rather than running them, the notable design choice is that one LoRA serves three different input paths, each changing where the model draws its reference from.

All three workflows share the same spine: the krea2_turbo_fp8_scaled base model loaded through a UNETLoader, the krea2_identity_edit_v1_2_fp16 LoRA applied at strength 1, the qwen3vl_4b_fp8_scaled text encoder set to type krea2, and decoding through a qwen_image HDR VAE. The sampler settings are identical in all three files — 10 steps, CFG 1, euler with a simple scheduler, full denoise. Notes embedded in the graphs add that switching to the raw bf16 model instead of the turbo variant means 40 steps and CFG 3–4, at which point negative prompts start to matter. The author's practical warning: do not mix the turbo LoRA with raw-model sampling settings.

A fidelity dial called ref_boost

The node that carries the actual editing behaviour is Krea2EditModelPatch, and its first setting is a dial named ref_boost. The graph notes are unusually specific about it: 1.0 reproduces the old v1.1 behaviour, 4.0 is the recommended preset, and very high values over-copy the reference until removals and replacements stop working entirely. The single-image workflow ships at 4; the mask and two-image workflows ship at 3.

Instructions are passed through a Krea2EditGroundedEncode node and written as plain-English sentences rather than tag lists. The saved examples pair a direct action — adding a pair of thin black cat-eye sunglasses, for instance — with an explicit list of what must not change: face, hairstyle, clothing, background, lighting.

What each mode does

The single-image editor, at 24 nodes and 26 links, looks like a plain one-image editor but ships with a bypassed group labelled as an optional second reference. Enabling it turns the same graph into a scene-plus-person compositor, with one ordering rule: the scene goes in slot 1 and the person in slot 2. Getting that backwards produces the opposite of the intended result, with no error message to flag it.

The mask editor, at 22 nodes and 27 links, is aimed at product and detail work. It adds a mask input plus a GrowMask node with expand 6 and taper enabled, dilating the mask before it gates the edit — skip that step, the author notes, and seams are the usual result. The saved state uses an encode size of 768 and ref_boost 3. The mask tells the model where it may write; the instruction sentence still has to say what may not move.

The two-image editor, at 22 nodes and 31 links, is the dedicated path for transferring identity or wardrobe from a second source. Its saved example does three things at once: it identifies the subject by position (the woman in Image A), enumerates each garment being transferred, and lists the invariants — facial identity, hairstyle, original scene. The graph notes recommend inputs around 1 megapixel, and at or below roughly 1.5 MP for images containing two people, which the author frames as a practical constraint rather than a hard technical limit.

Read the fine print

Two caveats are worth carrying away. First, the dev.to author ran no inference for the write-up — everything is read from the saved graphs, so the node notes describe intended behaviour, not measured output. Second, while the three workflows are free to download from Hugging Face, ModelScope and Civitai, the weights, LoRA and example media are not bundled and carry their own terms; the author explicitly asserts no blanket commercial or redistribution licence. The graphs also depend on custom nodes, including Krea2EditGroundedEncode and Krea2EditModelPatch, being present in your install at compatible versions.

Why it matters

Identity-preserving editing remains one of the harder problems in image generation, and most tools solve only one slice of it — global edits, local edits, or reference transfer. Shipping all three modes on a single LoRA, with one dial that trades reference fidelity against editability, gives ComfyUI users a coherent setup instead of three separate ones. The documented ref_boost guidance and the built-in mask growing are the kind of practical detail that usually has to be learned the hard way. The open question, which the source itself flags, is output quality: nothing in the published graphs demonstrates it, so v1.2's behaviour should be treated as intended rather than measured until independently tested.