This is the workflow I use to explore and compare combinations of samplers, schedulers, LoRAs, and checkpoints. It also has support for Ostris' Krea 2 Turbo style reference.# 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 youroutput 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'sthe 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 thetoolbar above it ("Execute to selected output nodes"; the branch gets anorange border). Hit it: the workflow renders and previews **without savinganything**.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 oneimage is committed to disk.## The knobs- Sampler switch <code>KSampler Switch</code>, true/false): choose between theClownsharKSampler 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 samplersClownshark 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 theClownsharKSampler — gaussian/hard, eta 0.5. Adds texture/variance back intothe distilled model's output. Leave as-is to start.- ColorCorrect: ships with temperature +10 (slightly warm), everything elseneutral. This is the node you iterate on in the loop above.- Image references (bypassed by default): un-bypass the LoadImage and feedup to 3 reference images into the Krea 2 edit encoder for image-conditionedgeneration. Leave bypassed for pure txt2img.- LoRAs: the Power Lora Loader ships with only the optional style-referencelora (disabled). Stack your own Krea 2 loras here; mind that filter-removalloras 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 resolutionsdecode 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: dd611a0551c5246041398fd152e0197a97731c73d3980e26ab25b28493a9730d