Krea 2 Settings

Krea 2 is the model the community actually moved to: the 12B open-weights release from Krea, licensed free under $1M revenue, that landed three weeks after Ideogram 4 and took over as the local general-purpose default. AIOKrea2Settings is the pack's control panel for it - one AIO_MODEL_SETTINGS output that goes into AIO Image Generate's model_settings socket with model_type set to krea2. The profile default is the provided workflow's path: 8-step er_sde sampler with the simple scheduler, CFG 1.0, on a 1344x2048 canvas.

The inputs that matter

The two you'll actually turn are the enhancer and the reference denoise:

• enhancer_enabled / enhancer_strength - the Krea2T prompt-adherence enhancer (adapted from capitan01R's ComfyUI-Krea2T-Enhancer). On by default at strength 1.0. This is the model patch that makes Krea 2 actually listen to the prompt instead of drifting into "pretty but unrelated." Set enhancer_strength to 0.0 to disable it entirely.
• reference_denoise - how much noise is applied when image 1 is connected as the Krea 2 reference (default 0.65). Lower preserves more of the reference; higher lets the prompt rewrite more. It's ignored entirely when no reference is connected, which makes it safe to leave set.
• inpaint_positive_prompt - an optional positive prompt used only for Krea 2 inpaint runs, so your text-to-image prompt and your inpaint prompt can diverge.
• max_length - Krea 2's text-conditioning token chunk length, default 4096 and capped at 4096. You can lower it; you can't exceed it.
• fp16_accumulation_enabled - matches the provided workflow's torch matmul setting, on by default, via ComfyUI model callbacks when supported. It's the kind of knob you leave alone unless your output quality changes with it.

Then the shared performance block - precision_policy, attention_mode, torch_compile_mode/torch_compile_backend, performance_apply_timing - same semantics as every other settings node in this pack.

The optional prompt_builder socket accepts the AIO_IDEOGRAM4_PROMPT from Ideogram 4 Prompt Builder. Krea 2 shares the Qwen/Ideogram structured-prompt DNA, so plugging the builder in gives you its JSON prompt and resolved dimensions for normal generation. And flip the builder's bbox_order to xy (Qwen-style) for Krea 2 - the default yx is Ideogram's convention.

Wiring and install

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/helto4real/comfyui-all-on-one-image-generation-node
cd comfyui-all-on-one-image-generation-node
python -m pip install -r requirements.txt

Or ComfyUI Manager → search "AIO Image Generate" → install → restart. Nodes live under AIO/Image. On the main node: model_type=krea2, diffusion model + Qwen3-VL text encoder + Qwen Image VAE in the dropdowns, and connect this settings node's AIO_MODEL_SETTINGS to model_settings. For reference img2img, connect your image to image 1 and play with reference_denoise; for inpaint, attach AIO Inpaint instead.

Gotchas

• image 1 is a reference or an inpaint source, never both. When an active AIO Inpaint config is connected it takes priority and image 1 becomes the inpaint source; the reference path only applies when inpaint is absent. The README notes a retained workflow connection alone doesn't select inpaint - the config has to actually be active.
• Krea only takes the first reference. image 2–image 4 are a FLUX.2 Klein thing; Krea ignores everything past image 1.
• Dimensions must be multiples of 16. Krea 2 won't happily sample a canvas that isn't.
• The enhancer is the point. If your Krea 2 output feels like it's ignoring the prompt, check enhancer_enabled before you start blaming the sampler - this pack ships the enhancer on by default for a reason.

Two dials, honestly: enhancer_strength for prompt adherence, reference_denoise for how much of your reference survives. Everything else is set-and-forget.