Eric Krea2 Generate

This is the "just make a picture" node in the pack. Wire a KREA2_PIPELINE from the Loader in, type a prompt, hit queue, and out comes an IMAGE plus a KREA2_LATENT you can feed into the fancier machinery. If you've used the stock KSampler workflow, this will feel familiar - it's the one node in the pack that doesn't make you think about stages or sigmas.

The best feature is auto_settings, which defaults on and exists to stop beginners from fighting Krea 2's two checkpoints. It reads the is_distilled flag from the loaded pipeline and applies the right steps and guidance for you: Turbo gets 8 steps at guidance 0, Raw gets 28 steps at 4.5. Those are exactly the conventions Krea ships - Turbo is TDM-distilled for 8 steps and no CFG, Raw is the undistilled base that wants real guidance. Turn auto_settings off and steps/guidance_scale go live.

One convention worth internalizing, because it comes up all over this pack: Krea's guidance_scale is the g convention where CFG = 1 + g. So guidance_scale = 0 is CFG 1.0 (no guidance), 0.2 is CFG 1.2, 4.5 is CFG 5.5. Turbo was distilled for g=0 and this node won't let you over-guide it in auto mode.

The rest of the required inputs are plain:

• prompt / negative_prompt - Krea 2 wants natural language, not tag soup. Long, specific, comma-separated sentences describing materials, light, and camera work beat keyword lists every time. Quote words to render them as literal text.
• width / height - 256–4096, snapped to /16, 1024 default. Turbo is calibrated up to ~2k; that's what the fast 2x decode path is for, not brute-forcing bigger canvas.
• seed - a plain integer; 0 means "random" in the ComfyUI sense.

There's one optional input, num_images (1–8), if you want a quick batch at the same settings.

Outputs: image is the standard IMAGE tensor for preview/save, and latent is the packed KREA2_LATENT. That latent output is the interesting one - Generate is effectively your "Stage 1" shortcut. You can run it to an image and stop, or take the latent and hand it to Multi-Stage Ultra's init_latent input for img2img refinement, or push it through Latent Resize before anything decodes it. Nothing forces you to do multi-stage sampling just to make a good image; for Turbo daily use, Generate → Upscale Decode (2x) is the whole fast 4k recipe.

How to install

Part of EricRollei/Krea2_ComfyUI_Advanced. ComfyUI Manager → search "Krea2_ComfyUI_Advanced", or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Krea2_ComfyUI_Advanced

Restart, then make sure you've done the two install steps the whole pack needs: a diffusers build from git that includes Krea2Pipeline (pip install --upgrade --force-reinstall --no-deps git+https://github.com/huggingface/diffusers.git) and transformers>=4.57.0. You also need Krea 2 weights in diffusers layout - the loader defaults point at a path that won't exist on your machine.

Common issues

• Prompt does nothing / refuses. That's Krea 2, not this node. The open checkpoint went through an alignment pass the hosted model didn't get, and the trained-in refusal craters around violence and body descriptions are a known complaint. Long natural-language prompts plus specificity ("Teen Titans (2003)", not "Teen Titans") are the workaround; a rebalance/uncensor LoRA is the stronger one.
• Mushy detail on Turbo. The ecosystem's fix is dropping the sampler, not the steps - this node uses the stock euler path. If you need sampler control, that's what Multi-Stage Ultra and the Sigmas node are for.
• Slow on Raw. 28 steps at 4.5 is the honest Raw setting. If it feels like a slog, that's the checkpoint working as intended; Turbo exists for speed.