Krea 2 Identity Edit:
Krea2IdentityEdit creates conditioning for Krea 2 image edits that use visual references to guide identity and appearance. It combines an edit instruction with image-aware text encoding and, when references are present, encoded reference images that the compatible editing model can use during generation. You can guide a scene, a subject, or both; control how strongly and when references influence the result; and limit changes to a selected region. It also supports combining its conditioning with compatible scene or style conditioning. With no reference image, it behaves like a plain Krea 2 text-conditioning node. The node prepares guidance for a downstream workflow; it does not render an image by itself.
Krea 2 Identity Edit Input Parameters:
clip
Required CLIP input. Connect the Krea 2-compatible text encoder used to turn your instruction and any reference-image descriptions into conditioning. The node expects a qwen3vl_4b vision text encoder loaded with CLIPLoader type krea2; an incompatible encoder cannot process the image prompt correctly. No default or numeric range is specified.
prompt
Required multiline STRING input containing your edit instruction; dynamic prompts are supported. Describe the change you want, such as changing clothing while keeping the person recognizable. For the recommended negative conditioning node, leave this instruction empty and connect the same reference images. No default or character-length range is specified. A preset name wired here is treated as literal prompt text, not as a preset command.
vae
Optional VAE input. When you supply reference images, this encodes them into the reference data used for identity preservation; the node raises an error if references are present but no VAE is connected. Use the VAE appropriate to your Krea 2 workflow. No default or numeric range is specified.
image
Optional IMAGE input for the first reference, typically the scene or the sole image being edited. When used alongside image2, this image comes first in reference order. Only the first frame is used if you connect an image batch. No default or numeric range is specified.
image2
Optional IMAGE input for the designated subject reference. It is placed after image and any chained sources, so the last-reference strength and mask controls apply to it in a two-reference workflow. Only the first frame is used if you connect an image batch. No default or numeric range is specified.
grounding_px
Optional INT input that caps the longest side of the first image sent to the vision text encoder; it affects how much image detail that encoder receives, separately from VAE reference encoding. Default: 768; minimum: 0; maximum: 4096; step: 32. Set 0 to avoid this resizing. Higher values can retain more detail but increase the image size processed for visual understanding.
fuse_with
Optional CONDITIONING input for conditioning you want placed before this node’s edit conditioning, such as compatible moodboard guidance. The node prepends its token rows while retaining this node’s reference-latent settings. For a predictable combination, use conditioning produced by a compatible Krea 2 node in the same pack. If the input contains multiple scheduled segments, only its first segment is fused. No default or numeric range is specified.
sources
Optional KREA2_SOURCES input for references from a Krea2 Edit Source Chain. These references are inserted between image and image2, keeping image2 last when it is connected. This helps you add sources without transferring the last-reference controls away from the designated subject. The source warns that its edit LoRA was trained with one or two references and that more than two may blend identities. No default or numeric range is specified.
ref_boost
Optional FLOAT input controlling how strongly the output attends to the last reference, usually the subject in a two-reference setup. Default: 1.0; minimum: 0.0; maximum: 10.0; step: 0.05. 1.0 adds no boost, values above it pull more strongly toward the reference’s appearance, and values below it loosen that pull. The node tooltip reports an edit-LoRA author suggestion of 2–6; treat that as a range to test, not a guaranteed best setting. Set this on positive conditioning and normally leave negative conditioning at 1.0.
ref_boost_a
Optional FLOAT input applying the corresponding boost to references before the last one, such as the scene in a two-reference setup. Default: 1.0; minimum: 0.0; maximum: 10.0; step: 0.05. 1.0 adds no boost. It has no effect with only one reference.
target_latent
Optional LATENT input representing the intended sampling grid, typically your empty sampling latent. With a VAE and references connected, the node fits reference images to that output resolution in pixel space before encoding them, instead of relying on later latent resizing. This is the path used by the fit_mode control and can help with resolution or aspect-ratio mismatches. It has no effect without a source image. No default or numeric range is specified.
fit_mode
Optional choice of fit or crop (legacy); default: fit. This setting applies when target_latent and vae are connected. fit generally preserves a mismatched source image’s aspect ratio within the target grid, using a small center crop when aspect ratios nearly match; crop (legacy) center-crops to the target aspect ratio and then resizes, matching the older geometry described in the source. Choose based on the edit weights and whether preserving the full source frame or using legacy crop behavior matters to your workflow.
ref_boost_mask
Optional MASK input selecting a region on the last reference where ref_boost should concentrate, such as a face. Draw the mask in the reference image’s frame, not the output frame. It lets you strengthen identity cues without boosting the entire reference equally. If ref_boost remains 1.0, the node notes that there is no boost for the mask to scope. No default or numeric range is specified.
ref_start
Optional FLOAT input specifying when reference guidance begins as a fraction of sampling progress. Default: 0.0; minimum: 0.0; maximum: 1.0; step: 0.05. For example, 0.2 starts reference guidance later, giving the composition some time to form first. It must be less than ref_end to create an active window.
ref_end
Optional FLOAT input specifying when reference guidance ends as a fraction of sampling progress. Default: 1.0; minimum: 0.0; maximum: 1.0; step: 0.05. Ending earlier removes the references from subsequent sampling steps. If ref_start is greater than or equal to ref_end, the node ignores the window and keeps references active throughout instead.
edit_mask
Optional MASK input for restricting where the edit may change the image: white marks the editable region, while the first source is re-imposed outside it at each denoising step. Draw it in the first source image’s frame. The source notes that this requires the source to align with the target grid, for example through a matching output aspect ratio or crop (legacy) geometry. With CFG above 1, connect the same mask to the negative edit node. No default or numeric range is specified.
edit_mask_feather
Optional INT input softening the boundary of edit_mask. Default: 2; minimum: 0; maximum: 32. The unit is latent pixels, where one latent pixel is approximately eight image pixels; 0 gives a hard edge. A softer edge can reduce an abrupt transition between edited and preserved areas.
isolate_refs
Optional BOOLEAN input; default: False. With multiple references, enabling it prevents references from attending to one another, which can reduce identity or style bleed between them. It has no effect with a single reference. Options: True or False.
boost_blocks
Optional choice of all, early, mid, or late; default: all. This experimental control limits where the reference-boost settings apply within the model: all is the normal setting, while early is described as leaning toward composition and pose and late toward texture and detail. It can be combined with the boost mask and timing window. If both boost values are 1.0, there is no boost to restrict.
grounding_px_subject
Optional INT input setting a separate vision-encoder image-size cap for references after the first, often the subject. Default: 0; minimum: 0; maximum: 4096; step: 32. At 0, those references use grounding_px; a nonzero value lets you allocate a different cap to later references, such as more detail for a face.
picture_labels
Optional BOOLEAN input; default: False. When enabled, this experimental training-parity setting prefixes each reference’s grounded prompt with a Picture N: label. The source says validated community results used the default-off behavior, so compare results before adopting it and use the same setting on positive and negative nodes. Options: True or False.
ref_t0_modulation
Optional BOOLEAN input; default: False. This experimental training-parity setting requests timestep-zero modulation of clean reference tokens. The source notes that the cited training setup uses it but the LoRA author’s inference node does not; compare outputs before adopting it and keep the setting the same on positive and negative nodes. Options: True or False.
system_prompt
Optional multiline STRING input overriding the vision system prompt used to interpret references. Default: an empty string, which retains the stock Krea 2 template. A custom instruction can change what the encoder emphasizes, such as a person’s facial features rather than an illustration style, but the source warns that an off-template prompt may be mildly outside the training distribution. Use identical text on positive and negative nodes and compare results. No character-length range is specified.
attention
Optional choice controlling the attention backend for the reference pass only; default: auto. The source describes auto as switching to PyTorch when xformers cannot run on the GPU, pytorch as always using PyTorch SDPA for this pass, and default as leaving the backend unchanged. The full set of choices comes from ATTENTION_MODES, whose definition is not included in the provided context, so additional choices cannot be confirmed here. Plain text-to-image processing is unaffected.
Krea 2 Identity Edit Output Parameters:
conditioning
The single CONDITIONING output contains the encoded instruction and, when images are supplied, the accompanying reference information and active settings for downstream Krea 2 generation. It is guidance for the model, not a finished image or a preview of the edit. Its contents can reflect reference order, boosts, timing, masks, target-grid fitting, and compatible fused conditioning, depending on what you connected.
Krea 2 Identity Edit Usage Tips:
- For an identity-preserving edit, connect the compatible Krea 2 vision
clip, avae, the source image, and a clear positive edit instruction. Create a corresponding negative node with an empty prompt and the same images. - For a scene plus a separate person, put the scene in
imageand the person inimage2. Adjustref_boostfor the person andref_boost_afor the earlier scene reference. - Connect your sampling latent to
target_latentwhen reference and output sizes or aspect ratios differ. Start withfitfor the documented v1.2 geometry; testcrop (legacy)when using older weights or when your mask workflow needs a source aligned to the output frame. - To retain a face without copying everything else from a portrait, draw
ref_boost_maskover the face on the subject reference and raiseref_boostabove1.0. Start with moderate changes and compare outputs. - Use
edit_maskwhen you need areas outside an edit to remain close to the first source. Check frame alignment first, and use the same mask on positive and negative nodes when CFG is above1. - If reference guidance constrains composition too early, try a later
ref_start, such as0.2. Keepref_startbelowref_end; otherwise the timing window is ignored. - Keep reference counts to one or two when identity separation matters. For multiple references, test
isolate_refsif their appearances bleed together; treat three or more references as an experiment beyond the stated training range. - When combining moodboard and edit conditioning through
fuse_with, prefer compatible conditioning from this pack and keeppicture_labelsconsistent across the fused inputs.
Krea 2 Identity Edit Common Errors and Solutions:
Krea 2 Identity Edit: the startup patch preflight failed (ComfyUI ABI change) - identity references are disabled this session.
- Explanation: The node detected that the required startup patch did not pass its compatibility check, so it refuses to process image references.
- Solution: Check the console startup message for the underlying compatibility issue and update RedNode Studio as the error directs. Retry after the patch preflight succeeds.
Krea 2 Identity Edit: connect the 'vae' input (your Krea 2 / qwen_image VAE).
- Explanation: At least one reference image is connected, but the node cannot encode its reference latents without a VAE.
- Solution: Connect the VAE used by your Krea 2 workflow to
vae.
Krea 2 Identity Edit: this CLIP cannot encode Krea 2 image prompts.
- Explanation: The connected
clipfailed while tokenizing the instruction and references. The full error includes an inner error with more detail. - Solution: Load the
qwen3vl_4bvision text encoder with CLIPLoader typekrea2and connect it toclip. If the error persists, inspect the inner error rather than assuming a different cause.
Krea 2 Identity Edit: wrong text-encoder type - load the qwen3vl_4b text encoder with CLIPLoader type 'krea2'.
- Explanation: Tokenization returned an encoder format that does not contain the expected
qwen3vl_4bcomponent. - Solution: Replace the connected text encoder with
qwen3vl_4bloaded as typekrea2.
[Krea2 Identity Edit] WARNING: the prompt is 'custom (use settings)', which is a PRESET NAME rather than an edit instruction.
- Explanation: A preset label, or a similar recognized label such as
custom (live),node's own, ornone, reachedprompt. The node warns you but still encodes those words literally. - Solution: Check the connection to
promptand supply an actual edit instruction. Leave the prompt empty only where you intentionally need empty negative conditioning.
[Krea2 Identity Edit] WARNING: ref_start (...) >= ref_end (...) - the refs would NEVER be active.
- Explanation: The reference timing window has no duration. The node ignores that window and keeps references on for every step.
- Solution: Set
ref_startlower thanref_end, such as0.2and1.0.
[Krea2 Identity Edit] 3 references - the edit LoRA trained on 1-2; expect identity blending beyond that
- Explanation: This is a warning, not a stopped run. With more than two references, separate identities may become mixed.
- Solution: Reduce the workflow to one or two essential references when distinct identity matters; compare outputs carefully if you retain more.
[Krea2 Identity Edit] input batch of ... - using frame 1
- Explanation: Identity references are single images in this node. When an
IMAGEinput contains a batch, only its first frame is used. - Solution: Connect the specific image you want as the reference instead of expecting the batch to provide multiple references.
[Krea2 Identity Edit] note: ref_boost_mask set but ref_boost is 1.0 - the mask has nothing to scope
- Explanation: The mask limits a boost, but the last reference currently has no added boost.
- Solution: Increase
ref_boostabove1.0if you intend to strengthen the masked reference region, or disconnect the mask if no boost is needed.
[Krea2 Identity Edit] note: isolate_refs has no effect with a single reference
- Explanation: There is no second reference from which to isolate the first.
- Solution: Leave
isolate_refsoff for single-reference edits, or add a second reference only when the edit actually calls for one.
[Krea2 Identity Edit] note: boost_blocks set but the ref_boost dials are 1.0 - nothing to restrict
- Explanation: A block range cannot limit a boost that has not been applied.
- Solution: Use
allwhile the boosts are1.0, or set the applicable reference boost above or below1.0before testing a restricted block range.
[Krea2 Identity Edit] fuse_with carries scheduled segments - only the first is fused
- Explanation: The fused input contains multiple scheduled conditioning segments, but this node uses only the first.
- Solution: Supply conditioning whose first segment is the one you intend to combine; do not expect later segments to be included.
[Krea2 Identity Edit] fuse_with input used a different Picture-label policy than this node
- Explanation: The fused conditioning and this node used different
picture_labelssettings, so their reference-description formats do not match. - Solution: Set
picture_labelsconsistently on both conditioning sources.
[Krea2 Identity Edit] fuse_with input is not from this pack - assuming it ends with the standard 5-row Krea 2 template tail.
- Explanation: The fused conditioning lacks the pack’s template-tail metadata. The node makes an assumption when removing its ending before fusion, which may not match that input’s actual format.
- Solution: Feed compatible conditioning from a Krea 2 Moodboard Encode node or another node in this pack for the documented fusion behavior.
[Krea2 Identity Edit] note: target_latent has no effect without a source image
- Explanation: There is no reference image to fit to the target grid.
- Solution: Connect a source image if you intend to use reference fitting; otherwise you can leave
target_latentdisconnected.
[Krea2 Identity Edit] note: edit_mask has no effect without a source image
- Explanation: The mask’s preservation behavior depends on a first source image.
- Solution: Connect the source image you want to preserve outside the masked region, or remove the unused mask.
