Doom Krea2 Sampler

Doom Krea2 Sampler is the pack's sampler for Krea 2, and its entire selling point is that it's pre-set correctly. Krea 2 is a distilled turbo model: it wants 8 steps, guidance of 1, and a fixed shift of 1.15 - not the resolution-dependent shift Flux uses. This node ships with exactly those defaults, so you load the model, type a prompt, and it behaves. No step-count archaeology, no "why is this oversaturated" detour.

That's worth more than it sounds. Distilled models punish you for using the defaults you learned on SDXL - the KB's distillation doc is blunt that CFG 7, 20 steps, and a Karras-ish schedule produce oversaturated junk on turbo checkpoints. Krea 2's own community history is a masterclass in this: the model is great, and most of its early "it looks awful" reports were people running it like a full model. This node removes that entire failure mode by construction.

How it works

The three dials that matter are already where Krea 2 wants them:

• steps - default 8. That's the turbo recipe. The tooltip notes Krea 2 RAW wants 52 - if you're running the undistilled RAW checkpoint, that's the number. The node doesn't switch it for you, but at least it tells you.
• shift - default 1.15, and explicitly not resolution-dependent. Flux samplers in this pack scale shift with resolution; Krea 2 does not, and the code's comment says the official value is fixed at 1.15. Leave it.
• CFG hidden at 1.0. There's no guidance input and no negative prompt input at all. Krea 2's DiT is guidance-distilled and has no guidance embedding, so a single positive pass at CFG 1 is the whole story - and ComfyUI skips computing the negative pass at exactly 1, which is a real speed win.

The sampler defaults to res_multistep sampler + simple scheduler, which is the Krea2-compatible pairing. Mechanically it builds the empty latent (16-channel Wan21-style, auto-fixed by fix_empty_latent_channels), injects nothing extra into the conditioning, and samples. Outputs are output and denoised_output latents - decode with the qwen_image_vae (the tooltip names it explicitly: Krea 2 ships with the Qwen-Image VAE, a detail that trips people decoding with the wrong VAE).

Inputs you'll set: model (e.g. krea2_turbo_fp8_scaled), conditioning, noise_seed, steps (8 turbo / 52 RAW), width/height, batch_size. That's it.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/PeterMikhai/Doom_Flux_NodePack

Restart or ComfyUI Manager. No extra deps; current ComfyUI required (V3 API). README's DoomAI_nodes.git line is stale - repo is Doom_Flux_NodePack. Model side: Krea 2 checkpoint + its Qwen3-VL text encoder + qwen_image_vae.safetensors.

Common issues

• Decoded with the wrong VAE → colors look off / washed. Use qwen_image_vae, not a Flux VAE.
• Running RAW at 8 steps → mushy, undercooked. RAW wants 52; turbo wants 8. Know which checkpoint you loaded.
• "I want more prompt adherence / style." There's no CFG dial to crank. If Krea 2 under-delivers, the lever is a LoRA or a different checkpoint - not guidance, which the architecture doesn't even take.

It's the rare sampler where the defaults are the feature. If you run Krea 2, you'll reach for this node because it means the only thing left to argue with is the prompt.