Comfyui Krea2

Krea2 on ComfyUI

Verified 2026-08-27 against ComfyUI 0.30.0. Engine mechanics (endpoints, polling, error reading) are in the comfyui skill — this is the model adapter only.

Krea2 is NOT a FLUX model

That assumption costs an hour. Krea2 uses a 12-layer Qwen3-VL text-encoder stack (12 × 2560 = 30720 conditioning features). Feeding it FLUX's T5+CLIP (4096) fails with:

Krea2 expects conditioning with 12x2560=30720 features (a 12-layer Qwen3-VL stack) but got 4096.

And it needs the Qwen image VAE, not FLUX's ae.safetensors. The wrong VAE does NOT error — it decodes to flat mauve fabric-like mush that looks like a bad seed. If output is degenerate but the job succeeded, check the VAE first.

Verified graph (API format)

UNET = "krea2_turbo_fp8_scaled.safetensors"   # or krea2_raw.safetensors (slower, higher fidelity)
CLIP = "qwen3vl_4b_fp8_scaled.safetensors"    # type MUST be "krea2"
VAE  = "qwen_image_vae.safetensors"           # NOT ae.safetensors

{
 "1": {"class_type":"UNETLoader",     "inputs":{"unet_name":UNET,"weight_dtype":"default"}},
 "2": {"class_type":"CLIPLoader",     "inputs":{"clip_name":CLIP,"type":"krea2"}},
 "3": {"class_type":"VAELoader",      "inputs":{"vae_name":VAE}},
 "4": {"class_type":"CLIPTextEncode", "inputs":{"text":PROMPT,"clip":["2",0]}},
 "5": {"class_type":"ConditioningZeroOut","inputs":{"conditioning":["4",0]}},   # negative
 "6": {"class_type":"EmptyLatentImage","inputs":{"width":1024,"height":1024,"batch_size":1}},
 "7": {"class_type":"KSampler","inputs":{"seed":SEED,"steps":8,"cfg":1.0,
        "sampler_name":"euler","scheduler":"simple","denoise":1.0,
        "model":["1",0],"positive":["4",0],"negative":["5",0],"latent_image":["6",0]}},
 "8": {"class_type":"VAEDecode","inputs":{"samples":["7",0],"vae":["3",0]}},
 "9": {"class_type":"SaveImage","inputs":{"filename_prefix":"out","images":["8",0]}}
}

Node choices that are NOT free to vary:

• EmptyLatentImage — not EmptySD3LatentImage.
• ConditioningZeroOut off the positive is the template's negative; an empty CLIPTextEncode also works but zero-out is what ships.
• Turbo wants cfg 1.0 and ~8 steps with euler/simple. Raising cfg on a turbo model degrades it; raising steps mostly wastes time.

Performance

6 s for 1024² at 8 steps on the reference box (4 s at 768²). Cheap enough to iterate on prompts and regenerate whole asset sets rather than fight one image.

Optional style LoRA

krea2_darkbrush.safetensors at ~0.8 via LoraLoaderModelOnly between the UNET and the sampler ({"class_type":"LoraLoaderModelOnly","inputs":{"lora_name":...,"strength_model":0.8,"model":["1",0]}}), then point the sampler's model at it. Ships in the stock template, off by default.

Prompting for game textures

Five rules, each learned by getting it wrong on a full 26-texture set and A/B-ing the fix. They matter more than the wording of any individual prompt.

1. Steps, not adjectives, produce detail

At 8 steps this model returns soft uniform micro-weave — limestone that reads as knitting, cobbles that read as mesh. At 20 steps the same prompt returns real blocks with mortar joints. If a texture looks characterless, raise steps before rewriting the prompt. 1024² at 20 steps is ~15 s; still cheap enough to regenerate a whole set.

2. NEVER phrase a prompt as a negation

Diffusion conditions on what you name. Observed, repeatedly:

Say what you want instead. This applies to the old advice "no shadows, no objects" too — write "flat even ambient lighting, uniform brightness edge to edge" and "continuous surface".

3. Describe the MATERIAL, not the object or the marks

The single highest-value rule; it fixed four textures that had each resisted three attempts.

Naming a countable object gets you a tiling grid of that object.

4. Do NOT anchor physical scale

"A 3 metre wide area", "individual stones 20 cm across", "shot from 3 metres away" — tried across eight textures, made every one more abstract, not less. Dropped. Rule 1 is the real lever.

5. An albedo is what a surface IS, never how it was lit

Ask for flat even lighting explicitly. A baked vignette is invisible at 1:1 and quilts the moment it tiles — one generation returned plaster with a 4×4 grid of vignettes inside a single 1024² tile, which rendered as a padded cell across every wall. Score it by blurring the luminance at several radii and measuring p95−p05; a single large radius cannot see a repeating vignette grid.

Judge textures TILED, never at 1:1

Every defect above is invisible in the raw PNG and obvious the moment it repeats. Always render a 3×3 tiled contact sheet before accepting a set. This is not optional — it is the only honest view.

Seams: cross-fade the edges, do not blur them

Do not offset by half and paint a blurred copy over the seam. It removes the seam and bakes a smeared grey cross into the tile, which then repeats as a visible grid of lines across every wall — worse than the seam it fixes, and it shipped in all 26 of our textures before anyone noticed.

Cross-fade the opposing edge bands instead, so the outer rows match exactly — seamless by construction, no blur anywhere:

def fade(a, axis, feather=0.18):     # a = HxWx3 float array
    n = a.shape[axis]; b = max(4, int(n * feather))
    t = np.linspace(0.5, 1.0, b, dtype=np.float32)       # 0.5 at the edge -> 1.0 inside
    t = t.reshape((b,1,1) if axis == 0 else (1,b,1))
    out = a.copy()
    if axis == 0:
        head, tail = a[:b], a[n-b:]
        out[:b] = head*t + tail*(1-t); out[n-b:] = tail*t[::-1] + head*(1-t[::-1])
    else:
        head, tail = a[:, :b], a[:, n-b:]
        out[:, :b] = head*t + tail*(1-t); out[:, n-b:] = tail*t[:, ::-1] + head*(1-t[:, ::-1])
    return out
img = fade(fade(img, 1), 0)

Prompting for fine even detail (rule 3) keeps the cross-fade ghosting invisible.

Alpha cut-outs without a matting model

For foliage, wire or netting cards when the install has no rembg/BiRefNet/SAM: generate the subject on flat black, then key alpha from luminance. Unpremultiply the colour (rgb / alpha) or every edge pixel keeps a black fringe when composited.

Record the prompt with the texture

Prompts that live only in shell history cannot be reproduced or improved — we had to re-author 26 of them. Keep a prompt table in version control and mirror it into the asset manifest.

Source of truth for all of the above: the stock image_krea2_turbo_t2i template, readable at GET /templates/image_krea2_turbo_t2i.json (the real nodes are inside definitions.subgraphs).