Writes a prompt that recreates one image, to reuse it or to check what a captioner sees. The models are huihui-ai abliterations (they don't refuse): Qwen3.8 27B as GGUF quants (default Q4_K_M) or Qwen3-VL 8B / Qwen3.5 9B / 4B as safetensors. Enable one <code>Model:</code> group. The default instruction asks for one dense recreate-this-image paragraph; edit it for tags or shorter prompts. Qwen3.8 27B needs flash attention in the QwenVL GGUF node on a 24 GB card (see the note; ComfyUI-QwenVL 2.3.2 leaves it off). Can't apply the patch, or no CUDA build of llama-cpp-python (common on Windows)? Enable the Qwen3-VL 8B / Qwen3.5 group instead: it runs on plain PyTorch, no llama.cpp needed.Needs ComfyUI-QwenVL 2.3.2 or newer. Its downloader node gained a setting in 2.3.2; on older releases the Download models nodes' values land in the wrong boxes.v7.1: the Image node opens ComfyUI's own <code>example.png</code> instead of a file you don't have, and the note states the QwenVL version.Files: <code>kiko-qwenvl3-abliterated-v7.1.json</code><h2 id="custom-nodes">Custom Nodes</h2><ul><li><a target="_blank" rel="ugc" href="https://github.com/ComfyAssets/ComfyUI-KikoTools">ComfyUI-KikoTools</a></li><li><a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">ComfyUI-KJNodes</a></li><li><a target="_blank" rel="ugc" href="https://github.com/1038lab/ComfyUI-QwenVL">ComfyUI-QwenVL</a> 2.3.2 or newer</li><li><a target="_blank" rel="ugc" href="https://github.com/rgthree/rgthree-comfy">rgthree-comfy</a></li></ul><h2 id="models">Models</h2><ul><li><code>Huihui-Qwen3.8-27B-abliterated-UD-DW-Q4_K_M.gguf</code> plus its vision (mmproj) file → <code>models/LLM/</code> (QwenVL picks the subfolder): <a target="_blank" rel="ugc" href="https://huggingface.co/huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF">https://huggingface.co/huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF</a> (other quants download by themselves when picked)</li><li><code>Huihui-Qwen3-VL-8B-Instruct-abliterated</code> (folder) → <code>models/LLM/huihui-ai/</code>: <a target="_blank" rel="ugc" href="https://huggingface.co/huihui-ai/Huihui-Qwen3-VL-8B-Instruct-abliterated">https://huggingface.co/huihui-ai/Huihui-Qwen3-VL-8B-Instruct-abliterated</a></li><li>Optional: <code>Huihui-Qwen3.5-9B-abliterated</code>, <code>Huihui-Qwen3.5-4B-abliterated</code> (huggingface.co/huihui-ai)</li></ul>The Download models group fetches and registers them through QwenVL's downloader.<h2 id="first-run">First run</h2>Each workflow has a bypassed Download models group. Turn it on in Toggle stages, queue once, then turn it off again. Files you already have are skipped. Civitai models and gated Hugging Face models need a token, set where ComfyUI starts (never in the workflow):<pre><code># Linux / macOS (~/.bashrc, ~/.zshrc or your start script)
export CIVITAI_API_KEY="your-civitai-key"
export HF_TOKEN="hf_your-token"

# Windows (Command Prompt once, then a new window)
setx CIVITAI_API_KEY "your-civitai-key"
setx HF_TOKEN "hf_your-token"</code></pre>Restart ComfyUI after setting them.<h2 id="workflow-image">Workflow image</h2>No example render: this workflow writes text, not images.<code>kiko-qwenvl3-image-to-prompt-workflow.png</code> is a picture of the whole graph (v7; v7.1 only changes the image file and the note). Load the <code>.json</code> to use it.Happy rendering! Kiko

Kiko image to prompt (Qwen3.8 27B GGUF / Qwen3-VL, abliterated)

⚠️ Superseded by <a target="_blank" rel="ugc" href="https://civitai.com/models/3042678?modelVersionId=3460041">v7.1</a>, please use that instead. v7.1 fixes:<ul><li>Missing image error on load: v7's Image node pointed at a file from my machine. v7.1 opens ComfyUI's own <code>example.png</code>.</li><li>Scrambled settings: the Download models nodes need ComfyUI-QwenVL 2.3.2 or newer; on older releases their values land in the wrong boxes. v7.1 states this in the note.</li><li>Windows / no llama.cpp: if you can't run the Qwen3.8 27B GGUF (no CUDA build of llama-cpp-python, or no flash-attention patch on a 24 GB card), enable the Qwen3-VL 8B / Qwen3.5 group instead; it runs on plain PyTorch.</li></ul>

Files with SHA256: 3ed8fec5a3d234d4f8f2db4e88ab7dc3d892be3800680d6244f7e54a99ed2558