AIO Image Generate

The name is not a lie. AIO Image Generate is the flagship of the helto4real "AIO Image Generate" pack, and it genuinely does what the pack title promises: pick a model family, pick three model files, type a prompt, hit run. It replaces the usual wall of loaders, encoders, samplers and decoders with one node that knows how to drive four different open-weights families - FLUX.2 Klein 9B, Ideogram 4, Krea 2, and Z-Image Turbo. No API, no key, all local.

What it actually is

The node is a facade. Under the hood the pack splits everything into profiles (services/profiles.py), adapters per model family, loader backends and validation - each family brings its own default sampler, scheduler, step count and canvas size, and at execution time the node lazily composes plain ComfyUI primitives (UNET loader, CLIP encode, ksampler, VAE decode). That's why you can switch families by changing one dropdown instead of rebuilding the graph. It's also why the pack imports without loading torch or GGUF: nothing heavy happens until you actually run.

The catch is that the node resolves model files lazily at execution time - nothing is auto-downloaded. If the dropdowns are empty or a file is missing, you get a failure at run, not a friendly installer.

The inputs that matter

Most of the node is "set once and forget," but these are the ones you'll actually touch:

• model_type - flux2_klein_9b, ideogram4, krea2, or z_image_turbo. This controls every default, so set it first.
• diffusion_model, text_encoder, vae - filename dropdowns searched across models/diffusion_models, text_encoders, clip, vae, and GGUF folders when a backend is installed. Pick the files for the family you chose.
• positive_prompt / negative_prompt - and here's the trap: use_zero_negative_conditioning defaults to on, which means the negative prompt isn't used as real sampler negative conditioning on any model (most of these families don't consume negatives anyway - the flow-matching generation inverts the SDXL rules, see the KB's concepts.md). Turn that toggle off if you actually want a negative encoded.
• size mode, max side, aspect ratio, multiple value - output sizing; set multiple value to 16 for both Ideogram 4 and Krea 2, since both demand multiples of 16 per side.
• steps and cfg - both default to 0, which means "use this family's profile default" (4 steps for distilled Klein, 20 for Ideogram 4, 8 for Krea 2 / Z-Image Turbo). An explicit number overrides.
• seed and batch_count - one image per seed, batch just walks seeds.
• second_pass_enabled - turns on a built-in upscale-then-low-denoise-refine pass, like a hires fix baked in.

The optional sockets matter a lot for building a real workflow: model_settings (from the matching settings node), lora_config (from AIO LoRA Configuration), inpaint (from AIO Inpaint), image 1–image 4 (references for Klein / Krea 2), and model/clip (externally patched models from AIO Load Pipeline Models).

The outputs and where they go

image is the main one - wire it to Preview or Save Image. You also get latent, width, height, run_info (a JSON string with what actually ran), and three bundled outputs you can't use directly: model_info, pid_info, and inpaint_info. Those need their companion nodes (AIO Model Info, AIO PID Info, AIO Inpaint Info) to unpack. This replaced older direct model/VAE/PID outputs, so if you loaded an old workflow that wired them straight into other nodes, you'll need to reconnect through the info nodes.

Installing it

ComfyUI Manager: search "AIO Image Generate" (that's the pack title) and install. Or do it by hand:

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   # helto-privacy + cryptography

Then restart ComfyUI; all nodes appear under AIO/Image. The Python dependencies are light - the heavy part is the model files, which you fetch yourself: Ideogram 4 wants its conditional + unconditional models plus Qwen3-VL text encoder and the FLUX.2 VAE; Krea 2 wants its own Qwen3-VL encoder and Qwen Image VAE; Klein and Z-Image use their standard ComfyUI packs.

Where people get burned

• Nothing generates because the model dropdown is empty or wrong. The node doesn't download anything. Download the family's files into the right models/ folders first.
• The negative prompt does nothing. See above - use_zero_negative_conditioning is on by default. That's intentional, not a bug.
• GGUF files with no GGUF backend. Pick a .gguf file and the node raises "no compatible GGUF backend detected" - you need ComfyUI-GGUF installed, and the pack won't silently fall back to safetensors.
• Settings mismatch. Connect FLUX.2 Klein 9B Settings while model_type is z_image_turbo and the node rejects it rather than guessing. Match the settings node to the family.

It's a young pack (every node here shows zero impressions on comfy.icu), so treat it as 0.1.0-era software: the README is detailed, the tests exist, but the community hasn't stress-tested it yet. For a single family's one-node workflow it's remarkably tidy.