Two Krea 2 workflows from my character/portrait pipeline, published free.

What you get
<ul>
<li>Version 1 — Two-Pass Recipe. Simple, readable graph. About 110 s per image at 1536x2048.</li>
<li>Version 2 — Pro-Grade 3-Pass + 4K. Adds a low-denoise detail pass and a 4K finish. 3072x4096, about 135 s on an RTX 5090.</li>
</ul>

The one setting that matters

Both versions work the same way: every sampling pass runs at a LOWER denoise than the one before it, so pass 1's composition survives and the later passes spend their steps on texture instead of re-inventing the image.

The common mistake is running every pass at denoise 1.0. Each pass then re-rolls the image, and you end up with a 4th-rate second idea instead of a refined first one.

Pick version 1 for speed and a graph you can read at a glance. Pick version 2 if you want the full ladder (1.00 -&gt; 0.45 -&gt; 0.28) and the 4K finish.

Running them

Both files use STOCK ComfyUI nodes only — no custom node packs, no missing-node boxes on load. The character-LoRA node ships BYPASSED with a placeholder, so you can drop in your own LoRA. Krea 2 LoRAs are model-only, so LoraLoaderModelOnly is the correct loader for them.

Ships pointing at Dark Beast Krea 2 FP8 (the public release). Any Krea 2 checkpoint works — just re-pick it on the UNETLoader node.

About the cover image

It is a LoRA-enabled production render, not a re-run of these files. It used version 1's two-pass structure (heun/ddim_uniform 12 steps, then dpmpp_2m_sde/beta 14 steps at denoise 0.40) on my private merged base, with two character LoRAs enabled: chloe_face_v5_v30 at 0.6 and chloe_body_v6 at 0.5.

Both are free downloads, from the "2.0" release of the Chloe model: <a href="https://civarchive.com/models/2846216?modelVersionId=3357040" rel="ugc">Chloe 2.0 — the face and body LoRAs used on this cover</a>

The LoRA node ships bypassed, so a LoRA-free graph gives the base model's own subject instead. That is the honest result of the file you downloaded, and it is why the base is named on the UNETLoader node. Press Ctrl+B to enable the node, then drop in a Krea 2 character LoRA — mine or your own — to get this look.

Character LoRAs I trained on this base are on my profile, if you want the faces these recipes were built around.

Krea 2 Recipes — Two-Pass & Pro-Grade 4K

Three-pass refinement on a falling denoise ladder, then a 4x ESRGAN finish to 3072x4096.
20 nodes in 6 labelled groups.

STAGE 1  composition    960x1280   heun / ddim_uniform    12 st  denoise 1.00
STAGE 2  refine        1536x2048   dpmpp_2m_sde / beta    12 st  denoise 0.45
STAGE 3  micro-detail  1536x2048   dpmpp_2m_sde / beta    14 st  denoise 0.28
STAGE 4  4K finish     upscale 4x -&gt; 3072x4096 -&gt; sharpen

Stage 4 is post-process, not sampling: 4x ESRGAN to 6144x8192, then ImageScale back to
3072x4096 with lanczos. Downscaling an oversampled image is what makes it read as sharp
rather than merely big.

Verified by real execution before upload — twice. 3072x4096 output, all three samplers and
all three post-process stages confirmed present in the output file's own embedded metadata,
with denoise = [1.0, 0.45, 0.28] and steps = [12, 12, 14]. 2.26% clipped highlights.
Render time 135 s on an RTX 5090 (a colder, contended run measured 642 s).

Needs one extra file: a 4x ESRGAN-family upscale model (4x-UltraSharp used here). Swap it
for whichever you have, or bypass stage 4 to stop at 1536x2048.

Stock ComfyUI nodes only. LoRA node ships bypassed — Ctrl+B to enable.

Files with SHA256: 044480eb810fc2e67f151d512f7e5762d8173849c75d4395415c6a000a6cce10