ComfyUI Node
Eric Krea2 Edit
Instruction / identity edit (Krea 2 Identity Edit recipe) for Multi-Stage Ultra's `edit` socket.
Eric Krea2 Edit
- source_image
- source_image_b
- ref_boost_mask
- edit
- grounding_preview
◄instruction►
◄fit_modefit►
◄ref_boost1.00►
◄ref_boost_a1.00►
◄grounding_px768►
◄size_fromsource►
◄ground_negativetrue►
◄system_prompt►
◄vision_processor_sourceH:\Testing\Qwen3-VL-4B-Instruct-heretic-7refusal►
CategoryEric/Krea2
Inputs (12)
| Name | Type | Default | Description |
|---|---|---|---|
| source_image | IMAGE | Image to edit (the scene / main reference, RoPE frame 1). | |
| instruction | STRING | Edit instruction, e.g. 'recolor the car to matte black'. NegPiP (phrase:-1) works here when the Negative Guidance node has NegPiP on. | |
| source_image_bopt | IMAGE | Second reference (the subject, RoPE frame 2) for person-into-scene edits. Place two people in ONE pass rather than chaining. | |
| fit_modeopt | COMBO | fit | fit = v1.2 training geometry (resample to the target grid at a centered offset; any aspect ratio). crop (legacy) = center-crop to the output aspect - only for v1/v1.1 weights. |
| ref_boostopt | FLOAT | 1.000–100 | Reference-fidelity dial on the LAST reference (the subject in two-ref edits, else the source): multiplies target->reference attention. 1 = off, ~4 strong likeness (model card), >10 breaks removals, <1 loosens. |
| ref_boost_aopt | FLOAT | 1.000–100 | Same dial for the FIRST reference (the scene) in two-ref edits. No effect with one reference. |
| ref_boost_maskopt | MASK | Optional region of the LAST reference to boost (e.g. the face), in that image's pixels. | |
| grounding_pxopt | INT | 7680–2048 | Longest side of the image(s) Qwen3-VL sees. Trained 384-768. Lower = stronger instruction adherence, higher = stronger likeness. Lower it if subjects duplicate. 0 = native size. |
| size_fromopt | COMBO | source | source: Stage 1 takes the source aspect at Ultra's s1_megapixels (width/height on Ultra still win if set). ultra settings: Ultra's aspect_ratio decides; the fit geometry handles the mismatch. |
| ground_negativeopt | BOOLEAN | true | With guidance (CFG) on: the negative is the EMPTY instruction grounded on the same image(s) - the trained unconditional. Off = Ultra's negative_prompt as plain text. |
| system_promptopt | STRING | Advanced: override the grounding system prompt (empty = Krea 2 / training default). | |
| vision_processor_sourceopt | STRING | H:\Testing\Qwen3-VL-4B-Instruct-heretic-7refusal | Folder with the Qwen3-VL image-processor config (same as the Vision Prompt node). |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| edit | KREA2_EDIT | — |
| grounding_preview | IMAGE | — |