This is a workflow for Krea 2 T2I and Qwen Image 2.1 TI2I, with an effective prompt enhancer/expander/director based on the abliterated/heretic versions of Qwen Image 2.1 Prompt Enhancer T2I/I2I and/or Qwen 3.8 27B LLM that run in Ollama.Important Pros:<ul><li>Remarkable text rendering for Krea 2 and Qwen Image 2.</li><li>Remarkable enrichment of prompts, tailored to each model (Krea 2, and Qwen 2.1) and T2I or I2I task, suitable to niche artistic generations that need NSFW freedom.</li><li>Very high quality generations up to 4MP, with an optional upscaler up to 8-16MP</li></ul>Important Cons:<ul><li>The workflow is complicated, relies on several custom nodes and may need an up to date ComfyUI.</li><li>While the low resolution generation without prompt processing needs resources like other workflows, high resolution generations with BF16 models and LLMs enabled need 32-96MB VRAM GPUs. RTX PRO 4500/5000/6000 Blackwell recommended.</li></ul>The details are available in the notes inside the workflow. A few points are worth highlighting:<ul><li>The base of this workflow were the ComfyUI workflows for Krea 2 and Qwen Image 2.1, and of course getting inspiration from community workflows.</li><li>The main additional components in this workflow are the prompt enhancers (PE) that were introduced by open image models such as Qwen Image 2.1 and Flux 3 Klein. These checkpoints that are finetuned LLMs based on Qwen 3.5 9B; they are not just some LLMs, but act as directors and prompt engineers or what a huge image model would do implicitly. Unfortunately many workflows or generations totally ignore them. This workflow maintains the Qwen Image 2.1 PE LLMs, but also extends their use to Krea 2 Turbo/Raw that were published without a dedicated PE.</li><li>The credit for heretic version of the Qwen Image PE's are pottokao's: <a target="_blank" rel="ugc" href="https://huggingface.co/pottokao/Qwen-Image-2.1-PE-T2I-Heretic">https://huggingface.co/pottokao/Qwen-Image-2.1-PE-T2I-Heretic</a> </li><li>The workflow also uses an abliterated version of the general purpose LLM, Qwen 3.8 27B for prompt expansion. The abliteration credit is for HuiHui AI. <a target="_blank" rel="ugc" href="https://ollama.com/huihui_ai/Qwen3.8-abliterated">https://ollama.com/huihui_ai/Qwen3.8-abliterated</a></li><li>LLMs need to get installed and run on Ollama outside of ComfyUI, but are controlled and used through connecting nodes inside ComfyUI.</li><li>The main contribution of the workflow are assembling these together, and importantly rewriting the system prompts for all LLMs separately for Krea 2 and Qwen Image 2.1 and each T2I or I2I task. LLMs helped to merge my prompting nodes, existing prompting skills by the community, like <a target="_blank" rel="ugc" href="https://github.com/rakshazi/krea2-prompting/tree/main">https://github.com/rakshazi/krea2-prompting/tree/main</a></li></ul>As with any other resource, the workflows provide substantial freedom for niche artistic concepts, including SFW and NSFW. The responsible use of resources are on each individual user.

Files with SHA256: b877a1c5b80ba5ceef03e704a22a0486e3dde827dfba7ec483035cc5d338c0bb