Krea 2 text to image that does not run out of memory at the decode. Drag the JSON in, load three models, write a paragraph. About 17 seconds a picture on our card.
Required models
<ul>
<li>Diffusion model: <a rel="ugc" href="https://civarchive.com/models/958009?modelVersionId=3139241">RedCraft 3.0 Krea 2</a>, the INT8 convrot build, <code>redcraft23_30Krea2_int8_convrot.safetensors</code>, in <code>models/diffusion_models</code>. Plain Krea 2 and Krea 2 Turbo load too; RedCraft is what we tested.</li>
<li>Text encoder: the Qwen3-VL 4B text encoder from the Krea 2 downloads (see <a rel="ugc" href="https://comfyui-wiki.com/en/tutorial/advanced/image/krea/krea-2">the Krea 2 guide on comfyui-wiki.com</a>); we run a community int8 ConvRot build of it (<code>qwen3vl_4b_int8_convrot</code>), and the bf16 or fp8 file works the same. It goes in <code>models/text_encoders</code>, loaded with CLIPLoader type <code>krea2</code>.</li>
<li>VAE: the Qwen-Image VAE, <code>qwen_image_vae.safetensors</code>, in <code>models/vae</code>, from <a rel="ugc" href="https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae">Comfy-Org on Hugging Face</a>.</li>
</ul>
Settings
<ul>
<li>Steps <code>8</code>. It is a distilled model, more steps do not help.</li>
<li>CFG <code>1.0</code>, sampler <code>er_sde</code>, scheduler <code>simple</code>, denoise <code>1.0</code></li>
<li>Size: 1 megapixel buckets, <code>896x1152</code>, <code>1152x896</code> or <code>1024x1024</code></li>
<li>Negative: a <code>ConditioningZeroOut</code>. At CFG 1.0 a negative prompt does nothing, so the graph does not pretend.</li>
<li>Decode: <code>VAEDecodeTiled</code>, tile 512, overlap 64</li>
<li>LoRA slot: one <code>LoraLoaderModelOnly</code>, bypassed and empty, behind a switch set to false. To use a Krea 2 LoRA, un-bypass it, pick the file, set the switch to true.</li>
</ul>
Why the tiled decode. On our 24 GB card the plain decode began to fail on some prompts, asking for 2.95 GB more than was left with all three models loaded. The tiled decode fixed it. Against the plain decode it differs by 0.2 out of 255 on average, which nobody can see.
Tips we learned the hard way
<ul>
<li>Density beats brevity. Write a full paragraph: who or what, where, the light, the camera. A short prompt gets you a short picture.</li>
<li>Say what you want, never what you do not. No negative prompt, and no SDXL habits: weight brackets and the old separator keyword do nothing here. Write "alone in the frame, no other people" instead of a negative list.</li>
<li>Name the medium. "Drawn in clean cel-shaded anime line art, drawn rather than photographed" holds a drawing; "semi-realistic" alone drifts to photo.</li>
<li>Stay at 1 MP. The buckets above are what the model was trained around.</li>
</ul>
Tested on ROCm (a 7900 XTX). Core ComfyUI nodes only, so CUDA builds should run it as is.
If this worked for you, a thumbs-up here helps the next person find it.

Krea 2 Tiled txt2img: No-OOM 8-Step Workflow

Krea 2 text to image that does not run out of memory at the decode. Drag the JSON in, load three models, write a paragraph. About 17 seconds a picture on our card.
Required models
<ul>
<li>Diffusion model: <a href="https://civitai.red/models/958009?modelVersionId=3139241" rel="ugc">RedCraft 3.0 Krea 2</a>, the INT8 convrot build, <code>redcraft23_30Krea2_int8_convrot.safetensors</code>, in <code>models/diffusion_models</code>. Plain Krea 2 and Krea 2 Turbo load too; RedCraft is what we tested.</li>
<li>Text encoder: the Qwen3-VL 4B text encoder from the Krea 2 downloads (see <a href="https://comfyui-wiki.com/en/tutorial/advanced/image/krea/krea-2" rel="ugc">the Krea 2 guide on comfyui-wiki.com</a>); we run a community int8 ConvRot build of it (<code>qwen3vl_4b_int8_convrot</code>), and the bf16 or fp8 file works the same. It goes in <code>models/text_encoders</code>, loaded with CLIPLoader type <code>krea2</code>.</li>
<li>VAE: the Qwen-Image VAE, <code>qwen_image_vae.safetensors</code>, in <code>models/vae</code>, from <a href="https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae" rel="ugc">Comfy-Org on Hugging Face</a>.</li>
</ul>
Settings
<ul>
<li>Steps <code>8</code>. It is a distilled model, more steps do not help.</li>
<li>CFG <code>1.0</code>, sampler <code>er_sde</code>, scheduler <code>simple</code>, denoise <code>1.0</code></li>
<li>Size: 1 megapixel buckets, <code>896x1152</code>, <code>1152x896</code> or <code>1024x1024</code></li>
<li>Negative: a <code>ConditioningZeroOut</code>. At CFG 1.0 a negative prompt does nothing, so the graph does not pretend.</li>
<li>Decode: <code>VAEDecodeTiled</code>, tile 512, overlap 64</li>
<li>LoRA slot: one <code>LoraLoaderModelOnly</code>, bypassed and empty, behind a switch set to false. To use a Krea 2 LoRA, un-bypass it, pick the file, set the switch to true.</li>
</ul>
Why the tiled decode. On our 24 GB card the plain decode began to fail on some prompts, asking for 2.95 GB more than was left with all three models loaded. The tiled decode fixed it. Against the plain decode it differs by 0.2 out of 255 on average, which nobody can see.
Tips we learned the hard way
<ul>
<li>Density beats brevity. Write a full paragraph: who or what, where, the light, the camera. A short prompt gets you a short picture.</li>
<li>Say what you want, never what you do not. No negative prompt, and no SDXL habits: weight brackets and the old separator keyword do nothing here. Write "alone in the frame, no other people" instead of a negative list.</li>
<li>Name the medium. "Drawn in clean cel-shaded anime line art, drawn rather than photographed" holds a drawing; "semi-realistic" alone drifts to photo.</li>
<li>Stay at 1 MP. The buckets above are what the model was trained around.</li>
</ul>
Tested on ROCm (a 7900 XTX). Core ComfyUI nodes only, so CUDA builds should run it as is.

Files with SHA256: 367cfa6d5fc1feb513069435da4ed3e01170b2ec206452cab49991e64898c668