<h1 id="promptwaffle-wf-qwen-2.1-turbo-character-lora-pipeline-v2">Promptwaffle WF — Qwen 2.1 Turbo Character LoRA Pipeline v2</h1>Turn character references into a varied image dataset for your next character LoRA.This ComfyUI workflow uses Qwen-Image 2.1 Turbo to generate portraits, expressions, body views, outfits, scenes, and lighting variations from your reference images. Four built-in prompt packs, automatic canvas selection, and a central control panel make it easier to run a complete dataset without rebuilding the graph.This workflow generates dataset images. LoRA training and captioning are handled separately.<h2 id="features">Features</h2><ul><li>Three reference slots: face, full-body front, and optional full-body back, with individual switches.</li><li>Four 70-prompt packs: Men, Women, Keep Identity, and Custom.</li><li>Automatic canvas selection: square, portrait, or landscape based on the prompt row.</li><li>Dual LoRA stacks: separate style/character and skin/detail stacks, with trigger-word controls.</li><li>Turbo sampling: 8 steps, CFG 1.0, Euler, and the included manual sigma schedule.</li><li>Optional PW Color Tools: palette extraction, grading, curves, optics, grain, and look/LUT export.</li><li>Organized output: character-based naming, PNG metadata, and raw-versus-graded comparison.</li><li>Memory controls: optional text-encoder unloading after encoding.</li></ul>The Keep Identity pack asks the model to preserve clothing and hairstyle. The Men and Women packs introduce more variation. Check and curate the generated images before training.<h2 id="required-model-files">Required model files</h2>Paths below are relative to your ComfyUI installation.ModelExact filename selected in this workflowFolderDiffusion model<code>qwen_image_2.1_turbo_bf16.safetensorsmodels/diffusion_models/Qwen/</code>Text encoder<code>qwen3vl_8b_nvfp4_heretic.safetensorsmodels/text_encoders/</code>VAE<code>qwen_image_2.1_vae_bf16.safetensorsmodels/vae/</code>Downloads:<ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Qwen-Image-2.1">Comfy-Org Qwen-Image 2.1 model files</a></li><li><a target="_blank" rel="ugc" href="https://huggingface.co/chfm/Qwen-Image-2.1-Text-Encoder-Heretic-NVFP4">Heretic NVFP4 text encoder</a></li></ul>The Turbo diffusion model and VAE are available from Comfy-Org. The selected Heretic NVFP4 encoder is a separate download. If you use a supported alternative encoder, select its filename in M2 after loading the workflow.If you place the diffusion model directly in <code>diffusion_models/</code> instead of the <code>Qwen/</code> subfolder, reselect it in M1.No additional LoRA is required to start. Both stacks are empty. Only load LoRAs explicitly compatible with Qwen-Image 2.1.<h2 id="required-custom-nodes">Required custom nodes</h2>Node packUsed for<a target="_blank" rel="ugc" href="https://github.com/willmiao/ComfyUI-Lora-Manager">ComfyUI-Lora-Manager</a>LoRA stacks, combining stacks, loading, trigger words, information, and metadata saving<a target="_blank" rel="ugc" href="https://github.com/Oratorian/ComfyUI-Switchboard">ComfyUI-Switchboard</a>Section control panels and Boolean switches<a target="_blank" rel="ugc" href="https://github.com/rgthree/rgthree-comfy">rgthree-comfy</a>Any Switch routing<a target="_blank" rel="ugc" href="https://github.com/ltdrdata/ComfyUI-Impact-Pack">ComfyUI-Impact-Pack</a>Comparisons and conditional canvas selection<a target="_blank" rel="ugc" href="https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes">Comfyroll Custom Nodes</a>CR Prompt List<a target="_blank" rel="ugc" href="https://github.com/Azornes/Comfyui-Resolution-Master">Comfyui-Resolution-Master</a>Output resolution controls<a target="_blank" rel="ugc" href="https://github.com/Fablestarexpanse/PW_ColorTools">PW Color Tools</a>Palette, Look, Curves, Optics, Grain, and Look I/O<a target="_blank" rel="ugc" href="https://github.com/Deno2026/comfyui-deno-custom-nodes">Deno custom nodes</a><code>DenoTextEncoderUnload</code>Install the dependencies, restart ComfyUI, and reload the workflow. Install PW Color Tools manually into <code>custom_nodes/</code> if it is unavailable through Manager.Use a current ComfyUI build containing TextEncodeQwenImage21, QwenImage21Cache, and the standard utility nodes used by this graph. The image comparison node is part of ComfyUI.Optional sections are switched off by default, but their node packs should still be installed to load the supplied graph completely.<h2 id="quick-start">Quick start</h2><ol><li>Load <code>PW_Qwen2.1_Turbo_CharacterLora_WF_V2.json</code>.</li><li>Replace the reference images with your own character. Enable the reference slots you want to use.</li><li>Set OUTPUT FOLDER and CHARACTER / FILE NAME.</li><li>Enable exactly one prompt pack in the Control Panel.</li><li>Reset PROMPT INDEX to 0.</li><li>Generate a test image, then queue 70 runs for the complete pack.</li></ol>Each queued run uses one prompt row. The index increments automatically and must be reset before starting another batch.<h2 id="canvas-sizes">Canvas sizes</h2>Prompt rowsOrientationSaved size1–18Square2048 × 204819–64Portrait1792 × 224065–70Landscape2240 × 1792For custom lists, keep this row order or adjust SQUARE ROWS END and VERTICAL ROWS END.<h2 id="notes">Notes</h2><ul><li>Negative prompts do not influence sampling at CFG 1.0, although the lists remain in the graph.</li><li>Leave color grading off for clean dataset generation, or configure your own grade first.</li><li>If you encounter out-of-memory errors, try the Unload Text Encoder switch, smaller reference resolution, or a smaller output canvas.</li><li>The Men and Women packs include nude adult body-reference prompts in rows 42–49. Remove those rows if unwanted; the Keep Identity pack excludes them.</li><li>Reference images and model weights must be supplied separately.</li></ul>Built by Promptwaffle

Qwen 2.1 Turbo Character Lora Set Maker

V3 changelog<ul><li>Added a single Turbo / Base switch that changes the diffusion model, CFG, and sampling schedule together.<ul><li>Turbo ON: CFG 1 with the original 8-step sigmas.</li><li>Base OFF: CFG 3 with BasicScheduler, simple, 25 steps.</li></ul></li><li>Only the selected diffusion-model branch loads. No additional custom-node packs required.</li><li>Negative prompts remain connected and take effect in Base mode at CFG above 1.</li><li>Fixed a duplicate group ID.</li><li>Updated the notes for sampling, reference switches, resized reference dimensions, model files, and LoRA compatibility.</li><li>Added a warning about the current male nudity limitation: outputs may have missing or censored genital anatomy. An unlocked model version may be needed to address it.</li><li>Preserved the 70-image batch setup, prompt packs, reference defaults, LoRA rig, colour grading, and layout.</li></ul>Validation: Turbo and Base sampling paths both passed separate 512 × 512 test renders. The full reference-and-LoRA dataset workflow has not been tested end to end.

Files with SHA256: 2967ea483befd33619d449a772f968b46ce8be1c5310bf896ddb7598c207bc59