<h3 id="krea-2-(bf16)-in-comfyui:-20482048-with-chunked-execution">Krea 2 (bf16) in ComfyUI: 2048×2048 with chunked execution</h3><p>A text-to-image workflow for Krea 2 that renders at 2048×2048 in 8 steps. It uses the LayerStream: Krea2 Chunking node and a tiled VAE decode so large images don't hit the memory wall.</p><h3 id="what-problem-it-solves">What problem it solves</h3><p>Big image models have two memory costs: the weights, and the activations. The activations grow with image size. For a full-size model like the 24.5 GB bf16 Krea 2, ComfyUI streams the weights, but the activations at large sizes can still spill into slow shared memory.</p><p>The node runs the transformer in slices, so the activation peak stays small. Decoding with tiles does the same for the VAE step.</p><h3 id="real-world-result">Real-world result</h3><p>On a 24 GB card with default launch flags, one production run of the bf16 model took about 30 seconds with the node. Without the node, it did not finish in a reasonable time. That is one run on one machine, so treat it as a report, not a guarantee.</p><h3 id="does-it-change-the-picture-please-read-this">Does it change the picture? Please read this</h3><p>Not bit-identical. I won't claim that for Krea 2, because it isn't true:</p><ul><li><p>With a single chunk the output was bit-identical.</p></li><li><p>With several chunks, the bf16 math rounds slightly differently depending on how many rows each slice has. The output is numerically equivalent but not pixel-identical. I measured about 40.7 dB PSNR against the stock render at 1 step, about 29–30 dB at 8 steps (1024²) and 38.5 dB at 4 MP. Compare a few renders side by side yourself, since I can't judge your taste for you.</p></li><li><p>I did not test whether the int8 or fp8 versions are exact. The tests above are for bf16.</p></li></ul><h3 id="speed-and-limits">Speed and limits</h3><ul><li><p>Slower: chunking costs time because weights are re-fetched for each slice. In my hard-cap tests it was about 2.7 to 4 times slower than stock where the stock run fit at all.</p></li><li><p>At 1 MP, the launch flags alone were enough. The node isn't needed there.</p></li><li><p>At 4 MP the node lowered PyTorch's memory floor by about 1 GB (stock worked at 3 GB, the node at 2 GB, both with the tiled VAE). That is PyTorch's own allocations, not total GPU usage, and not a promise for small cards.</p></li><li><p>Tested only on Windows with an RTX 4090 (24 GB), ComfyUI 0.39.0.</p></li><li><p>The node patches ComfyUI's Krea 2 code, so a future ComfyUI update could break it.</p></li></ul><h3 id="requirements">Requirements</h3><ul><li><p>ComfyUI 0.39.0 or newer</p></li><li><p>The osiworx-pack custom node pack, from ComfyUI-Manager or <a target="_blank" rel="ugc" href="https://github.com/osi1880vr/osiworx-pack">https://github.com/osi1880vr/osiworx-pack</a></p></li><li><p>Krea 2 model, the Qwen3-VL 4B text encoder (fp8) and the Qwen Image VAE. The file names in the loaders are from my machine, so pick your own. The workflow's default name is <code>krea2_turbo_int8_convrot.safetensors</code>, but my measurements were on the bf16 file.</p></li><li><p>Model files are not included.</p></li></ul><p>Measurements and untested areas are in the README: <a target="_blank" rel="ugc" href="https://github.com/osi1880vr/osiworx-pack">https://github.com/osi1880vr/osiworx-pack</a></p>

Files with SHA256: a43ab0890a31038870a81f6b983510ae8bb93c083e9278b8bd2348d9f5797e9c