# DnD (Dust and Dreams) — Krea 2 Workflow v1.5Text-to-image workflow for Krea 2 Turbo with a dual-sampler switch, optionalimage-reference editing, optional SDE noise injection, and a built-initerate-then-commit loop so you can tune post-processing against a locked seedwithout flooding your output folder. Runs on modest hardware — the VAE decode istiled specifically so large images work on Macs.---## Requirements### Custom node packs| Pack | Used for ||---|---|| [RES4LYF](<a target="_blank" rel="ugc" href="https://github.com/ClownsharkBatwing/RES4LYF">https://github.com/ClownsharkBatwing/RES4LYF</a>) | ClownsharKSampler_Beta, ClownOptions_SDE_Beta, the <code>beta57</code> scheduler || [rgthree-comfy](<a target="_blank" rel="ugc" href="https://github.com/rgthree/rgthree-comfy">https://github.com/rgthree/rgthree-comfy</a>) | Power Lora Loader, Image Comparer, Label || [ComfyUI-KJNodes](<a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">https://github.com/kijai/ComfyUI-KJNodes</a>) | LazySwitchKJ (the sampler switch) || [ComfyUI-Krea2-Ostris-Edit](<a target="_blank" rel="ugc" href="https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit">https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit</a>) | Krea 2 text encode + model patch || [comfyui-art-venture](<a target="_blank" rel="ugc" href="https://github.com/sipherxyz/comfyui-art-venture">https://github.com/sipherxyz/comfyui-art-venture</a>) | ColorCorrect |&gt; ⚠️ Don't trust "Install Missing Custom Nodes" blindly for this workflow —&gt; the embedded metadata misattributes several nodes to a different pack. The&gt; table above is the authoritative list. Everything not listed is comfy-core.### Models| File | Goes in | Note ||---|---|---|| <code>krea2_turbo_bf16.safetensors</code>](<a target="_blank" rel="ugc" href="https://civarchive.com/models/2726029/krea-2-turbo-official-comfy-org-checkpoints-krea2">https://civarchive.com/models/2726029/krea-2-turbo-official-comfy-org-checkpoints-krea2</a>) | <code>models/diffusion_models/</code> | the Krea 2 Turbo base (official Comfy-Org checkpoint) || <code>qwen3VLInstruct4bHeretic_v10.safetensors</code>](<a target="_blank" rel="ugc" href="https://huggingface.co/Kutches/Kr3a/blob/main/qwen3VLInstruct4bHeretic_v10.safetensors">https://huggingface.co/Kutches/Kr3a/blob/main/qwen3VLInstruct4bHeretic_v10.safetensors</a>) | <code>models/text_encoders/</code> | loaded with CLIP type krea2 — the workflow ships configured for this uncensored (heretic/abliterated) encoder || <code>qwen3vl_4b_bf16.safetensors</code>](<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Krea-2/blob/main/text_encoders/qwen3vl_4b_bf16.safetensors">https://huggingface.co/Comfy-Org/Krea-2/blob/main/text_encoders/qwen3vl_4b_bf16.safetensors</a>) | <code>models/text_encoders/</code> | the stock encoder, if you'd rather not use the heretic variant — just reselect it in the CLIPLoader || <code>wan_2.1_vae.safetensors</code>](<a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors">https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors</a>) | <code>models/vae/</code> | || <code>krea2_style_reference.safetensors</code>](<a target="_blank" rel="ugc" href="https://huggingface.co/ostris/krea2_turbo_style_reference">https://huggingface.co/ostris/krea2_turbo_style_reference</a>) | <code>models/loras/</code> | optional, ships disabled — grab it from the linked page if you want to use it |The LoRA loader references a <code>Krea 2/style/</code> subfolder — if your loras liveelsewhere, just reselect them in the Power Lora Loader.### ComfyUI versionRecent ComfyUI required: the <code>krea2</code> CLIP type and the selection-toolbar playbutton both need a current build. If clicking a node doesn't show a floatingtoolbar with a blue ▶ above it, update.---## Quick start1. Load the workflow. Fix any red model dropdowns by reselecting your local files.2. Type your prompt in the green node (positive). The red node is the negative — Turbo runs near-CFG-1, so it's empty by default and fine to leave that way.3. Set your resolution in Image Size (default 1920×816).4. Hit Run — everything executes, and you'll get about 3 images in your output folder per run (the comparer writes its two, plus the saved output). Fine for a final render. For iterating, use the loop below instead — that's the whole point of this workflow.## The iterate / commit loop (the intended way to use this)The SaveImage node is set to trigger-only, so you can render without committing:1. Click the Image Comparer node — a blue ▶ play button appears in the toolbar above it ("Execute to selected output nodes"; the branch gets an orange border). Hit it: the workflow renders and previews **without saving anything**.2. Like the composition? Lock your seed (both samplers ship <code>fixed</code>), then tweak ColorCorrect and play the comparer again. It shows base vs. color-corrected in a slide comparer — drag to compare.3. Happy? Click the SaveImage node and hit its play button. Exactly one image is committed to disk.## The knobs- Sampler switch <code>KSampler Switch</code>, true/false): choose between the ClownsharKSampler and the regular ComfyUI KSampler. - true → ClownsharKSampler (RES4LYF) — the recommended sampler. Ships <code>exponential/ddim</code> + <code>beta57</code>, 8 steps, CFG 1.0. - false → regular KSampler — included for comparisons, or to use samplers Clownshark doesn't offer, like <code>euler_ancestral</code> (ships <code>euler_ancestral/simple</code>, 10 steps, CFG 1.1).- Noise injection (OPTIONAL group): ClownOptions SDE settings feeding the ClownsharKSampler — gaussian/hard, eta 0.5. Adds texture/variance back into the distilled model's output. Leave as-is to start.- ColorCorrect: ships with temperature +10 (slightly warm), everything else neutral. This is the node you iterate on in the loop above.- Image references (bypassed by default): un-bypass the LoadImage and feed up to 3 reference images into the Krea 2 edit encoder for image-conditioned generation. Leave bypassed for pure txt2img.- LoRAs: the Power Lora Loader ships with only the optional style-reference lora (disabled). Stack your own Krea 2 loras here; mind that filter-removal loras add their own guidance — if things look overcooked, drop CFG toward 1.0.## Notes- Mac users: the tiled VAE decode is there for you — large resolutions decode without OOM. Both samplers here are MPS-safe.- Turbo wants low CFG (~1). If you raise steps/CFG expecting more quality, you'll get less — that's distillation, not a bug.

Files with SHA256: 6004d2540fb9253f13563b786e85d4948062401858fd6d7d00bc746d2c87c69d