Krea 2 Identity Edit v1.2 的三種 ComfyUI 編輯流程共用 ref_boost 保真度調節

Krea 2 shipped an editing model, and the interesting part is not the demo images — it is that the same LoRA now covers three different working modes, and each one changes where the model gets its reference from.

Three ComfyUI graphs ship with this episode, all built on the Krea 2 Identity Edit v1.2 LoRA. They share a base model, a text encoder and a VAE; what differs is the input path and the constraint: a plain single-image editor, a masked single-image editor, and a two-image reference editor.

No inference was run for this release. Everything below is read from the saved graphs, not from a benchmark I executed.

The shared spine

All three files load krea2_turbo_fp8_scaled.safetensors through a UNETLoader, apply krea2_identity_edit_v1_2_fp16.safetensors at strength 1 through LoraLoaderModelOnly, use qwen3vl_4b_fp8_scaled.safetensors with type krea2 as the text encoder, and decode through qwen_image_HDR_vae_fp32_comfy.safetensors.

The KSampler is identical in all three: 10 steps, CFG 1, euler, simple, denoise 1. That is the turbo-matched setup; the workflow's own note says switching to krea2_raw_bf16 instead of turbo means 40 steps and CFG 3–4, and then the negative prompt starts mattering.

The node that carries the actual editing behaviour is Krea2EditModelPatch, and its first widget is the fidelity dial — ref_boost. The note in the graph is unusually specific about it: 1.0 is described as the old v1.1 behaviour, 4.0 as the recommended preset, and very high values are said to over-copy the reference until removals and replacements stop working. The single-image graph ships at 4; the mask and two-image graphs ship at 3.

The instruction itself goes through Krea2EditGroundedEncode — a plain-English sentence, not a tag soup. The saved examples are written as direct actions ("add a pair of thin black cat-eye sunglasses", "change the outfit of the woman in Image A to the complete sporty outfit from Image B") and then explicitly list what should not change: face, hairstyle, clothing, background, lighting.

01 — Single-image editor, with a second reference hiding in plain sight

The first graph looks like a single-image editor and behaves like one by default, but it ships with a whole group named "2 · SECOND REFERENCE (optional — bypassed)". Inside that group is a second LoadImage with an example person, saved bypassed.

That matters because it means you do not need a different workflow to do "put this person into that scene": you enable the group. The order rule from the notes is that the scene goes in slot 1 and the person goes in slot 2. If you enable the group and put the person first, you get the opposite of what you wanted and no error message.

The active instruction in this file edits a product area into a clean-beauty tube while asking to preserve the model's exact face, smile, pose and lighting. Output prefix is krea2_identity_edit.

02 — Mask editor: the edit is constrained, the mask is grown

The masked graph is where the workflow gets practical for product and detail work. It adds a mask input and a GrowMask with expand 6 and taper on, and the saved example is a real object edit: add thin black cat-eye sunglasses, keep the face, hairstyle, flower accessory, makeup, clothing, background and lighting untouched.

Two things to notice. First, the mask is grown before it gates the edit — if you skip that, seams are the usual result. Second, the instruction still names everything that must stay; the mask tells the model where it may write, but the sentence tells it what may not move.

Encode size in this saved state is 768, ref_boost is 3, and output goes to ComfyUI/Kera2Edit.

03 — Two-image editor: scene plus reference, one sentence

The third graph is the dedicated two-image path: two active LoadImage nodes, Krea2EditModelPatch in two-image mode at ref_boost 3, and an instruction that transfers a full sporty outfit from image B onto the woman in image A.

The saved example is a good template because it does three things at once: it identifies the subject by position ("the woman in Image A"), it enumerates the pieces being transferred (white sports bra, navy high-waisted bottoms, white thigh-high socks, white sneakers), and it lists the invariants (facial identity, hairstyle, original scene).

The graph's notes recommend inputs around 1 MP, and at or below about 1.5 MP for two-person images. That is a practical constraint, not a hard technical limit, but it lines up with where this family of editors tends to fall apart.

How I would approach a first run

Start with the single-image editor, keep ref_boost at 4, and write a simple physical change with a clear "preserve" clause. Then move to the mask editor for anything that must stay local. Only reach for the two-image graph when the reference genuinely is a second source of identity or wardrobe — otherwise you are handing the model two competing references for no reason.

Keep the turbo tuning unless you are deliberately switching to the raw model; mixing a turbo LoRA with raw-model sampling settings is a reliable way to get mush.

What this is not

It is not a benchmark or a quality review. The saved prompts, images and preview nodes are the author's demo content; none of them proves an output. The node notes describe intended behaviour, not measured behaviour.

It is also not an installation guide. The graphs reference specific weights and custom nodes (Krea2EditGroundedEncode, Krea2EditModelPatch) that must exist in your install at compatible versions.

Getting the workflows

Published by this channel. The graphs are free; weights, LoRA and example media are not bundled and carry their own terms.

• Hugging Face: AI-KSK/krea2-identity-edit-three-modes-282 — three JSONs, source covers, SHA256 manifest. Start here.
• ModelScope (魔搭): Sunjian520/krea2-identity-edit-three-modes-282
• Civitai: 01 single-image · 02 mask editor · 03 two-image
• Run without a local GPU: 01 · 02 · 03 — these carry the author's invite code rh-v1111.
• Tutorial: YouTube · Bilibili

Verify workflow, node, model, LoRA, source-image and cover rights separately. No blanket commercial or redistribution license is asserted for the graphs or the example media.