RedNode Studio Settings (Advanced)

Every Krea 2 RedNode dial, named in plain English

By RedNodeAI·Created 3 months ago·Updated 3 days ago· 14

settings

◄transferstyle►

◄reference_processingfull image►

◄style_directivetrue►

◄hide_style_refstrue►

◄style_detail_px384►

◄likeness_vs_obedience768►

◄reference_fidelity2.50►

◄scene_fidelity1.00►

◄fit_modefit►

◄identity_start0.00►

◄identity_end1.00►

◄isolate_refsfalse►

◄boost_blocksall►

◄subject_likeness_px0►

◄edit_mask_feather2►

◄picture_labelsfalse►

◄ref_t0_modulationfalse►

◄vision_system_prompt►

◄attentionauto►

The advanced control surface for Krea2RedNode. One node, no image inputs, no conditioning - it just emits a KREA2_SETTINGS object that you plug into the RedNode's settings input, and the moment you connect it, the preset is replaced entirely. Every dial on Krea2RedNodeSettings wins.

What makes it worth a page of its own is that it names the fiddly technical parameters in plain language. The underlying nodes call things extract, indirect, budget_px, grounding_px, ref_boost. This one calls them transfer, hide_style_refs, style_detail_px, likeness_vs_obedience, reference_fidelity - and the tooltips tell you both the plain meaning and the technical parameter behind it.

The dials

transfer - what the style references contribute: style (palette/lighting/texture/mood) or subject (composition/content, with your prompt controlling the look). This is the extract mode.
reference_processing - full image, quadrant crops (2x2), or fine tiles (4x4). Crops and tiles scramble composition so only style survives from the references.
style_directive - adds the "style from the refs, subjects from the text" sentence next to the references.
hide_style_refs - ON by default, and this is the one worth understanding. Style references are deleted after encoding, so the image model never actually sees them; style transfers through prompt re-contextualization, and poses or people in the refs structurally cannot leak. That's the indirect mode from the moodboard nodes. Leave it on unless you specifically want the refs' content to transfer.
style_detail_px - resolution budget per style reference fed to the vision encoder (budget_px). 384 default; lower keeps token count down with lots of refs.
likeness_vs_obedience - the face-vs-instruction tradeoff (grounding_px). Lower = follows the instruction more; higher = preserves likeness more. 768 is balanced, 1024+ when the subject is a face.
reference_fidelity - pull toward the identity reference's appearance (ref_boost on the subject ref). 1.0 = off; 2–6 recommended with the v1.2 edit LoRA. The default here is 2.5, which tells you the author's opinion on how much boost the face usually wants.
scene_fidelity - the same dial for the scene reference in two-source setups (ref_boost_a). 1.0 = off.
fit_mode - fit is the v1.2 pixel-space geometry (blur-proof, any aspect ratio); crop (legacy) is the v1/v1.1 behavior for older LoRA weights.

How to use it

The workflow is: build your Krea2RedNode as normal, then add this node and connect its settings output to the RedNode's settings input. Everything you set here overrides the preset. Pair it with the RedNode's output_latent connected to your empty sampling latent so the fit geometry actually engages, and you've got the full manual identity-edit experience without touching the legacy building blocks.

It's also a genuinely good teaching tool: because each dial names its technical counterpart in the tooltip, you can use this node to learn what ref_boost or grounding_px actually do, then go back to the raw nodes with a clue.

Installing

Same pack, same one-line install - no third-party deps, GPL-3.0:

git clone https://github.com/RedNodeAI/ComfyUI-Krea2Moodboard ComfyUI/custom_nodes/ComfyUI-Krea2Moodboard

or via ComfyUI Manager ("Krea2Moodboard") and a restart. It needs the standard Krea 2 stack (ComfyUI 0.25.0+, qwen3vl_4b via CLIPLoader type krea2, qwen_image_vae) and a krea2_edit LoRA at 1.0 for anything identity-related. If the presets on the RedNode are 90% of what you want but one thing bugs you, this node is the smallest possible way to fix just that thing.

CategoryRedNode/Studio

Inputs (19)

Name	Type	Default	Description
transfer	COMBO	style	what the style references contribute: style = palette/lighting/texture/mood; subject = composition/content, the prompt controls the look. (technical: extract mode)
reference_processing	COMBO	full image	crops/tiles scramble composition so only style survives from the references
style_directive	BOOLEAN	true	adds a 'style from the refs, subjects from the text' sentence next to the references
hide_style_refs	BOOLEAN	true	style references are deleted after encoding, so the image model never sees them — style transfers via the prompt; poses/people in the refs cannot be copied directly (a small residual influence via the re-contextualized prompt can remain). (technical: indirect mode)
style_detail_px	INT	384128–1536	resolution budget per style reference fed to the vision encoder. (technical: budget_px)
likeness_vs_obedience	INT	7680–2048	cap on the identity source fed to the vision encoder: lower = follows the instruction more, higher = preserves likeness more (768 balanced, 1024+ for faces). (technical: grounding_px)
reference_fidelity	FLOAT	2.500–10	pull toward the identity reference's appearance: 1.0 = off, 2-6 recommended with the v1.2 edit LoRA. (technical: ref_boost on the subject ref)
scene_fidelity	FLOAT	1.000–10	same dial for the scene reference in two-source setups. (technical: ref_boost_a)
fit_mode	COMBO	fit	fit = v1.2 pixel-space geometry (blur-proof, any aspect ratio); crop = v1/v1.1 legacy behavior
identity_start	FLOAT	0.000–1	when the identity references become active during sampling (0 = first step). Start later (e.g. 0.2) to let the composition form freely before the identity locks in. (technical: ref timestep window)
identity_end	FLOAT	1.000–1	when the identity references switch off (1 = last step). ~0.3 = pose-only transfer: the ref drives composition, identity never locks.
isolate_refs	BOOLEAN	false	two-subject setups: stop the references reading each other — fixes identity/style bleed between two people. No effect with a single reference. (technical: ref-to-ref attention block)
boost_blocks	COMBO	all	experimental: where in the network the fidelity dials act — early leans composition/pose, late leans texture/detail. (technical: DiT depth range for ref_boost)
subject_likeness_px	INT	00–4096	separate likeness cap for the SUBJECT when a scene is also connected (the subject often deserves 1024+ while the scene doesn't). 0 = same as likeness_vs_obedience. (technical: grounding_px for the 2nd+ refs)
edit_mask_feather	INT	20–32	soft edge for the edit_mask input on the main node, in latent pixels (1 latent px ≈ 8 image px); 0 = hard edge
picture_labels	BOOLEAN	false	EXPERIMENTAL training-parity: 'Picture N: ' prefixes in the grounded prompt (ai-toolkit layout). Applied to both outputs. A/B before adopting.
ref_t0_modulation	BOOLEAN	false	EXPERIMENTAL training-parity: modulate reference tokens at timestep 0 (as ai-toolkit trains). Applied to both outputs. A/B before adopting.
vision_system_prompt	STRING		ADVANCED: override the vision system prompt for how the encoder READS the references (empty = training default). Example for anime→real: 'Describe the person's identity, face, hair and clothing as a real human being; ignore the illustration style.' Applied to both outputs; A/B.
attention	COMBO	auto	attention backend for the reference pass only. auto = fix the RTX 50xx/Blackwell xformers crash when it applies; pytorch = always plain SDPA here (avoids sage's quantized-attention smearing on reference detail); default = never change it.

Outputs (1)

Name	Type	Description
settings	KREA2_SETTINGS	—